Using synthetic data for pretraining partial discharge detection in overhead transmission lines Authors: Lukáš Klein, Jan Fulneček, Ondřej Kabot, Jiří Dvorský, Lukáš Prokop Published: 2025 ABSTRACT Accurate detection of partial discharges (PDs) in medium-voltage overhead transmission lines is critical for preemptive maintenance and avoiding costly outages, yet it is challenged by scarce labeled data and pervasive electromagnetic interference. This paper investigates a hybrid simulation-and-data-driven framework in which synthetically generated PD signals are used to pretrain deep neural networks and are subsequently fine-tuned on a limited set of real overhead-line measurements. The synthetic pipeline systematically varies PD repetition rates, amplitude distributions, vegetation-contact scenarios, and noise conditions, producing diverse time-series and spectrogram-like representations that approximate real operating environments. We conduct a comprehensive ablation study across multiple architectures—Convolutional Neural Networks (CNNs), a Vision Transformer (ViT), and a Long Short-Term Memory (LSTM) network—and analyze their sensitivity to granular sweeps of synthetic-data parameters. CNN-based models decisively outperform ViT and LSTM counterparts on the spectrogram-based classification task, while ViT and LSTM fail to learn meaningful representation. For the successful CNNs, pretraining on carefully parameterized synthetic datasets—particularly those reflecting higher PD activity, such as our Datasets 3 and 4—consistently improves downstream performance on real data, boosting the Matthews Correlation Coefficient (MCC) on imbalanced, cost-sensitive test sets by roughly 10–20% compared with training from scratch. At the same time, we show that poorly aligned synthetic data can degrade generalization, underscoring the need for accurate noise calibration and domain-aligned simulation. Overall, the results confirm that (i) architectural choice is pivotal for PD detection in overhead lines and (ii) well-designed synthetic data is a powerful, practical lever for achieving reliable and cost-effective PD monitoring when real labeled data are limited. Introduction Partial discharges are localized dielectric breakdowns that manifest in high-voltage insulation whenever the local electric field surpasses a critical threshold1,2. Though each individual discharge involves only modest energy release, their collective electrical, thermal and chemical effect can erode insulation over time and precipitate catastrophic system failures3,4. To mitigate vegetation-related outages and enhance safety, medium-voltage (MV) overhead distribution systems often use cross-linked polyethylene (XLPE)-covered conductors5. While this covering prevents direct short circuits from casual contact, it does not fully eliminate the possibility of PD under adverse conditions, such as contamination or sustained contact with branches3. Furthermore, the conductor covering can partially shield the radiated electromagnetic signals from PD events and may influence the discharge characteristics themselves, posing unique challenges for detection compared to bare conductors. Challenges in PD detection Early and reliable detection of PD is crucial for the timely identification and mitigation of high-impedance faults, reducing the risk of unplanned outages and potential fire hazards6. Conventional galvanic techniques (i.e., sensors that clamp directly onto energized lines) can yield accurate measurements yet typically require service interruptions for installation, incurring significant operational costs7. Non-contact or “radiometric” sensing via antennas has thus been proposed as a more practical alternative, as it monitors radiated electromagnetic signals from PD events without necessitating direct conductor access6,8. However, robust detection in the field remains difficult; signals can be masked by noise, corona discharges, and other transient electromagnetic phenomena, particularly in overhead configurations with variable environmental conditions9,10 and also distance of detection range (power lines span across many kilometers). Recent deep-learning pipelines have nevertheless achieved F-scores above 0.95 on laboratory benchmarks, e.g., Yang et al.’s parallel recognition ensemble on hybrid HV lines11 and Rauscher et al.’s data-augmented CNNs for electrical machines12. Synthetic data for PD classification One key obstacle hindering the development of accurate, machine-learning-based PD detectors is a general scarcity of labeled data. Two open-access antenna-based datasets now provide multi-year XLPE covered-conductor measurements13,14, yet each contains only a few hundred positively-labeled PD snapshots—far below what most deep models require. Field measurements tend to be costly, time-consuming, and limited in coverage—especially for sporadic events like partial discharges in overhead distribution systems13. As a solution, synthetic data generation has garnered increasing interest for augmenting or substituting real measurements in fault diagnosis pipelines15,16. They are also widely used in other fields of deep learning17. Traditionally, researchers have adopted two broad approaches to synthetic data generation. On one hand, physics-based simulation employs first-principles models or digital twins to reproduce the underlying discharge processes and create waveforms consistent with known electromagnetic phenomena18,19. Such methods can systematically produce large datasets covering rare or extreme conditions. Comparable physics-based strategies have demonstrated value for mechanical fault detection20,21, allowing models to learn from a range of scenarios that might be prohibitively difficult to collect experimentally. However, ensuring high-fidelity simulations of PD is challenging, given the complex stochastic behavior of discharges and the partial shielding effects introduced by conductor coverings. On the other hand, data-driven generative models offer a more empirical route. By learning statistical patterns from available real data and then sampling from the inferred distribution, these models can produce new, realistic training examples. This family of techniques includes Variational Autoencoders (VAEs)22 and, more prominently, Generative Adversarial Networks (GANs)23. In partial discharge research, GAN-based augmentation has been shown to improve classification accuracy, particularly when existing labeled datasets are limited24. Indeed, cGANs–GAN variants incorporating auxiliary information like operating conditions—have proven effective in synthesizing fault data with diverse amplitudes and noise conditions. However, purely GAN-driven approaches generally demand sufficient baseline datasets (i.e., some real PD examples to learn from), which may not always be feasible for overhead transmission power lines in early-stage monitoring25. Conditional generative adversarial networks Among data-driven approaches, cGANs stand out for their capacity to yield high-fidelity, labeled outputs tailored to specific environmental or operational contexts26,27. This strength is especially relevant for partial discharge classification, where signal amplitudes and spectral content vary widely depending on insulation health, voltage level, contamination, and vegetation interactions. Yet cGANs typically call for extensive tuning, extensive computation, and a sufficiently robust real-data foundation to ensure stable training. In practice, overhead distribution networks often lack the large annotated datasets necessary for purely data-driven pipelines25. The result can be synthetic outputs of questionable realism or coverage gaps that undermine classifier performance in real-world conditions. Physics-based simulation, conversely, offers a path to generate diverse data even without an extensive collection of initial real examples, making it potentially more suitable when real labeled data is severely limited. Gap and motivation Although both physics-based simulations and cGAN-based augmentation demonstrate promise, their application to medium-voltage overhead covered conductors remains comparatively underexplored. Generating synthetic PD signals under realistic wideband noise and partial shielding conditions—while also capturing sporadic events like vegetation contact—poses significant modeling and computational challenges. Moreover, it is unclear whether specific distributions of synthetic data (e.g., emphasizing high-rate PD events or wider amplitude variability) systematically yield more robust detectors, or how best to fuse synthetic pretraining with a small pool of real measurements. Objectives and contributions Motivated by these gaps, the present work proposes a hybrid simulation-and-data-driven approach to generate synthetic PD datasets that reflect the unique operational realities of XLPE-covered overhead transmission power lines. Our key contributions are: Flexible simulation framework A software environment capable of producing antenna-based PD time-series at 40 MS/s within a 20 ms capture window, controllable by parameters such as PD rates, contact scenarios, and noise levels.Diverse synthetic datasets Multiple synthetic distributions encompassing different PD amplitudes, repetition rates, vegetation contact conditions, and wideband interference factors, including variations simulating moderate to very high PD activity levels.Systematic parameter analysis We quantify the impact of varying synthetic data parameters (e.g., noise, PD intensity) on the final performance of the models after fine-tuning on real data, identifying parameter configurations that optimize for key metrics like ROC AUC and MCC.Two-phase training pipeline A workflow that uses synthetic data for pretraining deep neural networks, followed by fine-tuning on real overhead PD recordings, thus addressing label scarcity while retaining domain relevance.Comprehensive performance analysis We systematically quantify the impact of various synthetic data configurations on classification metrics—focusing on the Matthews Correlation Coefficient (MCC)—and highlight the scenarios in which synthetic pretraining most effectively enhances real-world detection, identifying specific dataset characteristics (like those in Datasets 3 and 4) that lead to the best outcomes.Comprehensive architectural evaluation We systematically compare the performance of CNN-based models against a Vision Transformer (ViT) and a Long Short-Term Memory (LSTM) network to determine architectural suitability for spectrogram-based PD detection.The rest of this article is structured as follows: Section Results details the detection and classification outcomes, comparing models trained with and without synthetic pretraining on real test sets. Section Discussion interprets the observed behaviors, examines the practical implications, and suggests guidelines for effective synthetic-data generation. Section Methods describes the synthetic data generation approach, the real-data acquisition setup, and the end-to-end model training pipeline. Finally, Section Conclusion summarizes the main findings and indicates future directions for deploying hybrid simulation-data-driven PD detectors in field applications. Flexible simulation framework A software environment capable of producing antenna-based PD time-series at 40 MS/s within a 20 ms capture window, controllable by parameters such as PD rates, contact scenarios, and noise levels. Flexible simulation framework A software environment capable of producing antenna-based PD time-series at 40 MS/s within a 20 ms capture window, controllable by parameters such as PD rates, contact scenarios, and noise levels. Diverse synthetic datasets Multiple synthetic distributions encompassing different PD amplitudes, repetition rates, vegetation contact conditions, and wideband interference factors, including variations simulating moderate to very high PD activity levels. Diverse synthetic datasets Multiple synthetic distributions encompassing different PD amplitudes, repetition rates, vegetation contact conditions, and wideband interference factors, including variations simulating moderate to very high PD activity levels. Systematic parameter analysis We quantify the impact of varying synthetic data parameters (e.g., noise, PD intensity) on the final performance of the models after fine-tuning on real data, identifying parameter configurations that optimize for key metrics like ROC AUC and MCC. Systematic parameter analysis We quantify the impact of varying synthetic data parameters (e.g., noise, PD intensity) on the final performance of the models after fine-tuning on real data, identifying parameter configurations that optimize for key metrics like ROC AUC and MCC. Two-phase training pipeline A workflow that uses synthetic data for pretraining deep neural networks, followed by fine-tuning on real overhead PD recordings, thus addressing label scarcity while retaining domain relevance. Two-phase training pipeline A workflow that uses synthetic data for pretraining deep neural networks, followed by fine-tuning on real overhead PD recordings, thus addressing label scarcity while retaining domain relevance. Comprehensive performance analysis We systematically quantify the impact of various synthetic data configurations on classification metrics—focusing on the Matthews Correlation Coefficient (MCC)—and highlight the scenarios in which synthetic pretraining most effectively enhances real-world detection, identifying specific dataset characteristics (like those in Datasets 3 and 4) that lead to the best outcomes. Comprehensive performance analysis We systematically quantify the impact of various synthetic data configurations on classification metrics—focusing on the Matthews Correlation Coefficient (MCC)—and highlight the scenarios in which synthetic pretraining most effectively enhances real-world detection, identifying specific dataset characteristics (like those in Datasets 3 and 4) that lead to the best outcomes. Comprehensive architectural evaluation We systematically compare the performance of CNN-based models against a Vision Transformer (ViT) and a Long Short-Term Memory (LSTM) network to determine architectural suitability for spectrogram-based PD detection. Comprehensive architectural evaluation We systematically compare the performance of CNN-based models against a Vision Transformer (ViT) and a Long Short-Term Memory (LSTM) network to determine architectural suitability for spectrogram-based PD detection. Results Overall performance across model types and synthetic data We evaluated three model architectures—Max-Pooling, Traditional, and SOTA-Inspired—across six synthetic-data conditions (Dataset 1, Dataset 2, Dataset 3, Dataset 4,All Datasets Combined, and No Pretraining for no pretraining). Performance was measured using the MCC, F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} score, and accuracy. Figures 1, 2, and 3 presents box plots of these three metrics for each combination of model type and synthetic-data pretraining.Fig. 1Box plots of MCC by model type and synthetic data. Dataset 3 and Dataset 4 achieved the highest median MCC across the three model architectures. Pretraining with Dataset 1, Dataset 2, or the All Datasets Combined resulted in mixed performance, while skipping pretraining (No Pretraining) did not necessarily lead to the worst outcome for all models.Fig. 2Box plots of F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} scores by model type and synthetic data. The SOTA-Inspired model with Dataset 3 and Dataset 4 achieved the highest F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document}, whereas other synthetic datasets (Dataset 1, Dataset 2, and All Datasets Combined) sometimes led to lower F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} than models trained without pretraining (No Pretraining).Fig. 3Box plots of accuracy by model type and synthetic data. Dataset 3 and Dataset 4 consistently improved accuracy across models. However, models pretrained on Dataset 1, Dataset 2, or All Datasets Combined exhibited varied performance, sometimes lower than models without pretraining (No Pretraining). Box plots of MCC by model type and synthetic data. Dataset 3 and Dataset 4 achieved the highest median MCC across the three model architectures. Pretraining with Dataset 1, Dataset 2, or the All Datasets Combined resulted in mixed performance, while skipping pretraining (No Pretraining) did not necessarily lead to the worst outcome for all models. Box plots of F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} scores by model type and synthetic data. The SOTA-Inspired model with Dataset 3 and Dataset 4 achieved the highest F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document}, whereas other synthetic datasets (Dataset 1, Dataset 2, and All Datasets Combined) sometimes led to lower F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} than models trained without pretraining (No Pretraining). Box plots of accuracy by model type and synthetic data. Dataset 3 and Dataset 4 consistently improved accuracy across models. However, models pretrained on Dataset 1, Dataset 2, or All Datasets Combined exhibited varied performance, sometimes lower than models without pretraining (No Pretraining). Overall, pretraining with Dataset 3 or Dataset 4 yielded the highest median MCC (as can be seen in Table 1), F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} score, and accuracy across all model types. In contrast, skipping pretraining (No Pretraining) generally resulted in lower performance compared to Dataset 3 or Dataset 4, though its impact varied across models. The SOTA-Inspired model experienced the largest performance drop (over 0.20 in MCC) when pretraining was omitted. Among the synthetic alternatives, Dataset 3 or Dataset 4 proved to be the most beneficial for partial discharge detection, whereas Dataset 1, Dataset 2, and the All Datasets Combined provided mixed results, sometimes offering only marginal improvements or even underperforming compared to no pretraining.Table 1Comparison of MCC grouped by dataset.DatasetGroupMean MCCMinMaxStd. Dev.Dataset 1Max-Pooling0.500− 0.0130.8290.189Traditional0.460− 0.0320.7270.145SOTA-Inspired0.279− 0.0380.6080.194Dataset 2Max-Pooling0.5910.0000.8240.162Traditional0.4940.0530.7770.126SOTA-Inspired0.3800.0990.6380.157Dataset 3Max-Pooling0.7370.6270.8360.049Traditional0.7100.4320.8140.070SOTA-Inspired0.6770.4930.8290.075Dataset 4Max-Pooling0.7270.1040.8240.094Traditional0.6990.4810.8450.073SOTA-Inspired0.6810.4540.8760.082All Datasets CombinedMax-Pooling0.4970.0000.7880.173Traditional0.4900.2890.7530.129SOTA-Inspired0.236− 0.0660.6860.214N/AMax-Pooling0.521− 0.0460.8710.274Traditional0.6650.4290.8270.090SOTA-Inspired0.445− 0.0260.7710.252Significant values are in bold. Table 1Comparison of MCC grouped by dataset.DatasetGroupMean MCCMinMaxStd. Dev.Dataset 1Max-Pooling0.500− 0.0130.8290.189Traditional0.460− 0.0320.7270.145SOTA-Inspired0.279− 0.0380.6080.194Dataset 2Max-Pooling0.5910.0000.8240.162Traditional0.4940.0530.7770.126SOTA-Inspired0.3800.0990.6380.157Dataset 3Max-Pooling0.7370.6270.8360.049Traditional0.7100.4320.8140.070SOTA-Inspired0.6770.4930.8290.075Dataset 4Max-Pooling0.7270.1040.8240.094Traditional0.6990.4810.8450.073SOTA-Inspired0.6810.4540.8760.082All Datasets CombinedMax-Pooling0.4970.0000.7880.173Traditional0.4900.2890.7530.129SOTA-Inspired0.236− 0.0660.6860.214N/AMax-Pooling0.521− 0.0460.8710.274Traditional0.6650.4290.8270.090SOTA-Inspired0.445− 0.0260.7710.252Significant values are in bold. Comparison of MCC grouped by dataset. Significant values are in bold. Statistical analysis of MCC Normality and homogeneity of variance We examined whether MCC values in each of the 18 groups (3 model architectures \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 6 synthetic-data conditions) satisfied normality and homoscedasticity. Shapiro–Wilk tests revealed that some groups deviated from normality (e.g., SOTA-Inspired | Dataset 2, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p=0.0386$$\end{document}), while others did not (e.g., Max-Pooling | Dataset 3, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p=0.9033$$\end{document}). Levene’s test further indicated a violation of homogeneity of variances (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p<0.0001$$\end{document}). Because of these violations (see Table 2), we employed the non-parametric Kruskal–Wallis test to compare MCC across all 18 groups. All statistical significance testing was conducted using an alpha level of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.05$$\end{document}.Table 2Normality and Homogeneity of Variance Test Results (MCC).MethodDatasetNormality Test (p-value)Homogeneity Test (p-value)SOTA-InspiredDataset 20.03860.0000SOTA-InspiredDataset 30.83510.0000SOTA-InspiredDataset 40.16590.0000SOTA-InspiredDataset 10.01530.0000SOTA-InspiredAll Datasets Combined0.02620.0000SOTA-InspiredN/A0.00000.0000Max-PoolingDataset 20.00010.0000Max-PoolingDataset 30.90330.0000Max-PoolingDataset 10.08940.0000Max-PoolingAll Datasets Combined0.23270.0000Max-PoolingDataset 40.00000.0000Max-PoolingN/A0.00000.0000TraditionalDataset 20.27930.0000TraditionalN/A0.11980.0000TraditionalDataset 10.10650.0000TraditionalAll Datasets Combined0.10790.0000TraditionalDataset 40.05270.0000TraditionalDataset 30.00030.0000 Table 2Normality and Homogeneity of Variance Test Results (MCC).MethodDatasetNormality Test (p-value)Homogeneity Test (p-value)SOTA-InspiredDataset 20.03860.0000SOTA-InspiredDataset 30.83510.0000SOTA-InspiredDataset 40.16590.0000SOTA-InspiredDataset 10.01530.0000SOTA-InspiredAll Datasets Combined0.02620.0000SOTA-InspiredN/A0.00000.0000Max-PoolingDataset 20.00010.0000Max-PoolingDataset 30.90330.0000Max-PoolingDataset 10.08940.0000Max-PoolingAll Datasets Combined0.23270.0000Max-PoolingDataset 40.00000.0000Max-PoolingN/A0.00000.0000TraditionalDataset 20.27930.0000TraditionalN/A0.11980.0000TraditionalDataset 10.10650.0000TraditionalAll Datasets Combined0.10790.0000TraditionalDataset 40.05270.0000TraditionalDataset 30.00030.0000 Normality and Homogeneity of Variance Test Results (MCC). Kruskal–Wallis omnibus test The Kruskal–Wallis test yielded \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H=479.486506$$\end{document} (df \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$=17$$\end{document}) with a corresponding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p=3.939812\times 10^{-91}$$\end{document}, indicating highly significant differences in MCC across the 18 model—dataset combinations. This confirms that both model architecture and the choice of synthetic pretraining data significantly affect classification performance. Post-hoc Dunn’s test with Bonferroni correction We conducted Dunn’s post-hoc test (with Bonferroni adjustment) to pinpoint which pairs of groups differed significantly. Results are shown in Table 3. The key observations from the post-hoc analysis include:Dataset 3 and Dataset 4 outperforms other synthetic conditions In nearly all comparisons, groups that used Dataset 3 or Dataset 4 achieved significantly higher MCC than those using Dataset 1, Dataset 2, or All Datasets Combined, with p-values majority of cases (over half) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$<10^{-10}$$\end{document}. Crucially, pairwise comparisons between groups using Dataset 3 and Dataset 4 for the same model architecture generally showed no statistically significant difference (p > 0.05 after Bonferroni correction), indicating their comparable high effectiveness.Skipping pretraining often reduces performance but can outperform suboptimal synthetic data While omitting pretraining (No Pretraining) generally leads to lower performance compared to Dataset 3, its impact varies by model type. In some cases, models without pretraining outperform those trained on Dataset 1, Dataset 2, or the All Datasets Combined. However, across all architectures, Dataset 3 and Dataset 4 consistently provides the most significant performance boost, with p-values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ll 0.01$$\end{document} in pairwise comparisons.Table 3Bonferroni-corrected p-values from Dunn’s post-hoc test for pairwise comparisons of MCC scores (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha =0.05$$\end{document}, Bonferroni).Max-PoolingSOTA-InspiredTraditionalDataset 1Dataset 2Dataset 3Dataset 4Virtual RawDataset 1Dataset 2Dataset 3Dataset 4Virtual RawDataset 1Dataset 2Dataset 3Dataset 4Virtual Rawx0.5454390.0000000.0000001.000000x1.0000000.0000000.0000001.000000x1.0000000.0000000.0000001.0000000.545439x0.0000550.0000020.2668121.000000x0.0000000.0000001.0000001.000000x0.0000000.0000001.0000000.0000000.000055x1.0000000.0000000.0000000.000000x1.0000000.0000000.0000000.000000x1.0000000.0000000.0000000.0000021.000000x0.0000000.0000000.0000001.000000x0.0000000.0000000.0000001.000000x0.0000001.0000000.2668120.0000000.000000x1.0000001.0000000.0000000.000000x1.0000001.0000000.0000000.000000x Dataset 3 and Dataset 4 outperforms other synthetic conditions In nearly all comparisons, groups that used Dataset 3 or Dataset 4 achieved significantly higher MCC than those using Dataset 1, Dataset 2, or All Datasets Combined, with p-values majority of cases (over half) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$<10^{-10}$$\end{document}. Crucially, pairwise comparisons between groups using Dataset 3 and Dataset 4 for the same model architecture generally showed no statistically significant difference (p > 0.05 after Bonferroni correction), indicating their comparable high effectiveness. Dataset 3 and Dataset 4 outperforms other synthetic conditions In nearly all comparisons, groups that used Dataset 3 or Dataset 4 achieved significantly higher MCC than those using Dataset 1, Dataset 2, or All Datasets Combined, with p-values majority of cases (over half) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$<10^{-10}$$\end{document}. Crucially, pairwise comparisons between groups using Dataset 3 and Dataset 4 for the same model architecture generally showed no statistically significant difference (p > 0.05 after Bonferroni correction), indicating their comparable high effectiveness. Skipping pretraining often reduces performance but can outperform suboptimal synthetic data While omitting pretraining (No Pretraining) generally leads to lower performance compared to Dataset 3, its impact varies by model type. In some cases, models without pretraining outperform those trained on Dataset 1, Dataset 2, or the All Datasets Combined. However, across all architectures, Dataset 3 and Dataset 4 consistently provides the most significant performance boost, with p-values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ll 0.01$$\end{document} in pairwise comparisons. Skipping pretraining often reduces performance but can outperform suboptimal synthetic data While omitting pretraining (No Pretraining) generally leads to lower performance compared to Dataset 3, its impact varies by model type. In some cases, models without pretraining outperform those trained on Dataset 1, Dataset 2, or the All Datasets Combined. However, across all architectures, Dataset 3 and Dataset 4 consistently provides the most significant performance boost, with p-values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ll 0.01$$\end{document} in pairwise comparisons. Table 3Bonferroni-corrected p-values from Dunn’s post-hoc test for pairwise comparisons of MCC scores (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha =0.05$$\end{document}, Bonferroni).Max-PoolingSOTA-InspiredTraditionalDataset 1Dataset 2Dataset 3Dataset 4Virtual RawDataset 1Dataset 2Dataset 3Dataset 4Virtual RawDataset 1Dataset 2Dataset 3Dataset 4Virtual Rawx0.5454390.0000000.0000001.000000x1.0000000.0000000.0000001.000000x1.0000000.0000000.0000001.0000000.545439x0.0000550.0000020.2668121.000000x0.0000000.0000001.0000001.000000x0.0000000.0000001.0000000.0000000.000055x1.0000000.0000000.0000000.000000x1.0000000.0000000.0000000.000000x1.0000000.0000000.0000000.0000021.000000x0.0000000.0000000.0000001.000000x0.0000000.0000000.0000001.000000x0.0000001.0000000.2668120.0000000.000000x1.0000001.0000000.0000000.000000x1.0000001.0000000.0000000.000000x Bonferroni-corrected p-values from Dunn’s post-hoc test for pairwise comparisons of MCC scores (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha =0.05$$\end{document}, Bonferroni). Pairwise comparisons between Dataset 3/Dataset 4 and other conditions Given their strong performance, we conducted focused pairwise tests comparing Dataset 3 and Dataset 4 to each other synthetic-data condition for every model type (Table 4). Depending on normality and variance checks, we used either the Mann–Whitney U test or Welch’s t-test at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.05$$\end{document}. The results consistently show statistically significant advantages for Dataset 3 and Dataset 4, often with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 10^{-8}$$\end{document} when compared to Dataset 1 or Dataset 2, and also significant improvements relative to All Datasets Combined and No Pretraining. As shown in Table 4, the differences between Dataset 3 and Dataset 4 themselves were typically not statistically significant (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} in the ‘Sig. Better?’ column), highlighting their similar efficacy.Table 4Pairwise Comparisons. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document} indicates Dataset 3 performed significantly better based on the chosen test (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.05$$\end{document}); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} indicates no statistically significant difference (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p \ge 0.05$$\end{document}).Model typeDataset 3 vsTest usedStatisticp-valueMean DiffMedian DiffSig. Better?Max-PoolingDataset 1T-test (Welch’s)9.3652.65e-14+0.237+0.229\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}Max-PoolingDataset 2Mann-Whitney U1389.5005.44e-08+0.146+0.122\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}Max-PoolingDataset 4Mann-Whitney U1165.0000.570+0.010-0.003\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}Max-PoolingN/AMann-Whitney U7149.0002.70e-10+0.216+0.112\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}Max-PoolingAll Datasets CombinedT-test (Welch’s)9.1626.66e-13+0.240+0.226\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SOTA-InspiredDataset 1Mann-Whitney U1664.0002.56e-14+0.398+0.429\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SOTA-InspiredDataset 2Mann-Whitney U1154.0009.02e-12+0.296+0.255\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SOTA-InspiredDataset 4T-test (Student’s)-0.2350.814-0.004+0.002\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}SOTA-InspiredN/AMann-Whitney U1762.0003.77e-09+0.231+0.136\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SOTA-InspiredAll Datasets CombinedMann-Whitney U1138.0003.31e-11+0.440+0.526\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}TraditionalDataset 1Mann-Whitney U3097.0003.52e-17+0.250+0.265\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}TraditionalDataset 2Mann-Whitney U3172.5009.55e-17+0.216+0.213\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}TraditionalDataset 4Mann-Whitney U2154.5000.229+0.011-0.005\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}TraditionalN/AMann-Whitney U1442.0000.002+0.045+0.033\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}TraditionalAll Datasets CombinedMann-Whitney U2158.5002.72e-13+0.220+0.209\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document} Table 4Pairwise Comparisons. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document} indicates Dataset 3 performed significantly better based on the chosen test (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.05$$\end{document}); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} indicates no statistically significant difference (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p \ge 0.05$$\end{document}).Model typeDataset 3 vsTest usedStatisticp-valueMean DiffMedian DiffSig. Better?Max-PoolingDataset 1T-test (Welch’s)9.3652.65e-14+0.237+0.229\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}Max-PoolingDataset 2Mann-Whitney U1389.5005.44e-08+0.146+0.122\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}Max-PoolingDataset 4Mann-Whitney U1165.0000.570+0.010-0.003\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}Max-PoolingN/AMann-Whitney U7149.0002.70e-10+0.216+0.112\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}Max-PoolingAll Datasets CombinedT-test (Welch’s)9.1626.66e-13+0.240+0.226\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SOTA-InspiredDataset 1Mann-Whitney U1664.0002.56e-14+0.398+0.429\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SOTA-InspiredDataset 2Mann-Whitney U1154.0009.02e-12+0.296+0.255\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SOTA-InspiredDataset 4T-test (Student’s)-0.2350.814-0.004+0.002\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}SOTA-InspiredN/AMann-Whitney U1762.0003.77e-09+0.231+0.136\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}SOTA-InspiredAll Datasets CombinedMann-Whitney U1138.0003.31e-11+0.440+0.526\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}TraditionalDataset 1Mann-Whitney U3097.0003.52e-17+0.250+0.265\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}TraditionalDataset 2Mann-Whitney U3172.5009.55e-17+0.216+0.213\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}TraditionalDataset 4Mann-Whitney U2154.5000.229+0.011-0.005\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}TraditionalN/AMann-Whitney U1442.0000.002+0.045+0.033\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}TraditionalAll Datasets CombinedMann-Whitney U2158.5002.72e-13+0.220+0.209\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document} Pairwise Comparisons. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document} indicates Dataset 3 performed significantly better based on the chosen test (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.05$$\end{document}); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} indicates no statistically significant difference (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p \ge 0.05$$\end{document}). Visual comparison of time-domain signals and spectrograms A qualitative visual comparison was conducted between both the time-domain waveforms and the corresponding spectrograms generated from the synthetic data pipeline and those derived from real-world overhead transmission power line measurements. In the time domain, the synthetic signals replicate the presence of sharp, transient pulses characteristic of partial discharge (PD) superimposed on a noise floor (compare Figs. 5 and 6). Similarly, the synthetic spectrograms exhibit structural resemblances to real data spectrograms, capturing the broadband energy bursts associated with PD events (compare Figs. 4 and 4 Real Signal - Spectrogram). Domain expert assessment, considering both representations, confirmed that the synthetic data captures relevant PD characteristics deemed suitable for the intended pretraining application.Fig. 4Spectrograms of synthetic datasets (d1–d4) and a real partial discharge signal.Fig. 5Real Signal - Time-domain representation showing amplitude variation over time.Fig. 6Time-domain representations of synthetic signals under different sample configurations. Spectrograms of synthetic datasets (d1–d4) and a real partial discharge signal. Real Signal - Time-domain representation showing amplitude variation over time. Time-domain representations of synthetic signals under different sample configurations. However, specific visual differences were noted across both domains. In the time domain, the baseline noise texture and amplitude in synthetic signals may differ from the real measurements, which sometimes exhibit specific periodic interference or distinct noise floor characteristics. Real signals occasionally show broader, low-amplitude oscillations (potentially related to power line harmonics or specific noise sources like switching power supplies) not fully replicated synthetically. In the frequency domain (spectrograms), these differences manifest as variations in background texture and the potential absence of certain continuous narrowband or broadband interference components present in the real data. Furthermore, synthetic datasets designed for severe conditions (Dataset 3, Dataset 4) intentionally feature more frequent and higher-amplitude PD pulses in both time-domain and spectrogram representations compared to average real samples. Despite these visual differences, the success of pretraining, especially with Dataset 3 and Dataset 4, suggests the models effectively learned the core discriminative features of PD. We hypothesize that the CNN, particularly when processing spectrograms, focuses on the characteristic broadband, transient energy bursts associated with PD events. These key features were successfully replicated in the synthetic data (especially their intensity and frequency variations in Datasets 3/4), making the model robust to variations in background noise texture or the absence of specific narrowband interference components during classification. F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document}-score and accuracy trends Although MCC offers a robust view under class imbalance, F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} score and accuracy also confirm the dominance of Dataset 3. As shown in Fig. 2 and 3, most Dataset 3-pretrained models achieve F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1>0.90$$\end{document} and accuracy exceeding 0.90. In contrast, skipping pretraining or using less-optimized synthetic data (e.g., Dataset 1) resulted in lower metrics and wider variance. Parallel Kruskal–Wallis and post-hoc tests (not shown) reveal the same pattern of statistically significant differences, underscoring the consistent benefit of Dataset 3. Performance comparison across architectures To ascertain the optimal model architecture for spectrogram-based partial-discharge (PD) detection, we conducted a comparative evaluation of five distinct models: a SOTA-inspired convolutional neural network (CNN), a max-pooling CNN, a traditional CNN, a Long Short-Term Memory (LSTM) network, and a Vision Transformer (ViT). The models were assessed across a comprehensive suite of synthetic pretraining configurations, with performance quantified primarily by the Receiver Operating Characteristic Area Under the Curve (ROC AUC) and further validated using the Matthews Correlation Coefficient (MCC). The results revealed a distinct performance hierarchy among the architectures, as illustrated by the ROC AUC heatmap in Fig. 7. The three convolutional architectures consistently demonstrated superior and robust performance, achieving mean ROC AUC values in excess of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.84$$\end{document} across the majority of synthetic data configurations. The SOTA-inspired CNN emerged as the premier model, maintaining a mean ROC AUC of approximately \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.88$$\end{document}. This finding is strongly corroborated by an analysis of the mean MCC effect for each model, presented in Table 5. The convolutional models all exhibited a positive effect relative to the overall mean performance, with the SOTA-inspired CNN showing the largest positive deviation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta = +0.038$$\end{document}). In stark contrast, the LSTM and ViT architectures demonstrated a profound inability to learn discriminative features from the spectrogram inputs, both yielding large negative effects (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta = -0.349$$\end{document}) and ROC AUC scores near the random-chance baseline of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.5$$\end{document}. This performance disparity underscores the critical role of architectural inductive bias; for this task, models designed for two-dimensional spatial feature extraction (CNNs) are substantially more effective than those oriented toward sequential data (LSTM) or global patch-based attention (ViT) under the given data regime.Table 5Mean MCC effect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document}) of the Model Head choice versus the overall mean.Model head choiceEffect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document} vs. Overall)LSTM-0.349ViT-0.349SOTA-inspired CNN+0.038Max-Pooling CNN+0.035Traditional CNN+0.019 Table 5Mean MCC effect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document}) of the Model Head choice versus the overall mean.Model head choiceEffect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document} vs. Overall)LSTM-0.349ViT-0.349SOTA-inspired CNN+0.038Max-Pooling CNN+0.035Traditional CNN+0.019 Mean MCC effect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document}) of the Model Head choice versus the overall mean. Fig. 7Heatmap of mean ROC AUC scores for all evaluated architectures across synthetic-pretraining configurations. CNN-based architectures (SOTA-inspired CNN, max-pooling CNN, traditional CNN) achieve consistently high performance (lighter/yellow regions), while the LSTM-based model and the Vision Transformer remain near random baselines (dark regions). Heatmap of mean ROC AUC scores for all evaluated architectures across synthetic-pretraining configurations. CNN-based architectures (SOTA-inspired CNN, max-pooling CNN, traditional CNN) achieve consistently high performance (lighter/yellow regions), while the LSTM-based model and the Vision Transformer remain near random baselines (dark regions). Sensitivity analysis of the SOTA-inspired CNN to synthetic-data parameters Given its superior performance, the SOTA-inspired CNN was selected for a detailed sensitivity analysis to determine the influence of individual synthetic-data generation parameters. While its ROC AUC remained high across most configurations, a quantitative analysis based on the MCC revealed clear dependencies on specific pretraining conditions, as detailed in Table 6.Table 6Mean MCC effect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document}) of hyperparameter choices, relative to the overall mean and the mean within each model.FactorChoice\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document}) vs.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document} vs. Model MeanOverallLSTMViTSOTA-CNNMax-Pool CNNTrad. CNNNoise Level200,000-0.123+0.000+0.000+0.011-0.009-0.021500,000-0.122-0.000-0.000-0.003-0.037-0.0061,000,000-0.110+0.000+0.000-0.007+0.042+0.029Vegetation Mean1.10-0.120+0.000-0.000-0.012-0.002+0.0042.20-0.116-0.000+0.000+0.011+0.002-0.005Vegetation Rate800-0.110-0.000+0.000+0.016+0.006-0.0043,200-0.118+0.001-0.000+0.004+0.000-0.0065,000-0.127-0.001+0.000-0.023-0.006+0.010Vegetation Sigma0.45-0.113-0.000+0.000+0.006+0.004+0.0021.85-0.116+0.000-0.000-0.001+0.004+0.0032.00-0.125-0.000+0.000-0.005-0.008-0.005 Table 6Mean MCC effect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document}) of hyperparameter choices, relative to the overall mean and the mean within each model.FactorChoice\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document}) vs.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document} vs. Model MeanOverallLSTMViTSOTA-CNNMax-Pool CNNTrad. CNNNoise Level200,000-0.123+0.000+0.000+0.011-0.009-0.021500,000-0.122-0.000-0.000-0.003-0.037-0.0061,000,000-0.110+0.000+0.000-0.007+0.042+0.029Vegetation Mean1.10-0.120+0.000-0.000-0.012-0.002+0.0042.20-0.116-0.000+0.000+0.011+0.002-0.005Vegetation Rate800-0.110-0.000+0.000+0.016+0.006-0.0043,200-0.118+0.001-0.000+0.004+0.000-0.0065,000-0.127-0.001+0.000-0.023-0.006+0.010Vegetation Sigma0.45-0.113-0.000+0.000+0.006+0.004+0.0021.85-0.116+0.000-0.000-0.001+0.004+0.0032.00-0.125-0.000+0.000-0.005-0.008-0.005 Mean MCC effect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta$$\end{document}) of hyperparameter choices, relative to the overall mean and the mean within each model. Effect of PD Amplitude (vegetation mean) Pretraining with more salient PD events was found to enhance downstream performance. The quantitative data in Table 6 shows that increasing the mean amplitude of the vegetation-induced PD signal from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\texttt {vegetation mean} = 1.1$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2$$\end{document} shifted the MCC effect for the SOTA-CNN from negative (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta = -0.012$$\end{document}) to positive (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta = +0.011$$\end{document}). This numerical finding is visually confirmed by the upward shift in the performance distribution depicted in Fig. 8. These results support the hypothesis that exposure to stronger, more clearly separable targets during pretraining facilitates the learning of robust time-frequency representations that transfer more effectively to real-world data.Fig. 8Performance distribution for the SOTA-inspired CNN as a function of the mean PD amplitude (vegetation mean) in the synthetic pretraining data. Higher-amplitude settings (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2$$\end{document}) shift the distribution upward, indicating a consistent relative gain. Performance distribution for the SOTA-inspired CNN as a function of the mean PD amplitude (vegetation mean) in the synthetic pretraining data. Higher-amplitude settings (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2$$\end{document}) shift the distribution upward, indicating a consistent relative gain. Effect of repetition rate (vegetation rate) Contrary to the intuition that a greater number of training events is invariably beneficial, our analysis reveals an optimal range for event density. Table 6 indicates that a moderate repetition rate of 800 pulses/window yielded a significant positive effect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta = +0.016$$\end{document}), whereas a very high rate of 5,000 pulses/window resulted in a detrimental effect (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta = -0.023$$\end{document}). This trend is consistent across different noise levels, as illustrated in the heatmap in Fig. 9. We posit that excessively dense synthetic PD activity generates training examples that diverge substantially from the sparser, event-driven characteristics of authentic monitoring data, thereby diminishing the transferability of the learned features.Fig. 9Interaction between synthetic noise level and PD repetition rate for the SOTA-inspired CNN. Moderate repetition rates (800 pulses/window) consistently lead to better relative performance than highly dense scenarios (5 000 pulses/window), even when the background noise level is increased. Interaction between synthetic noise level and PD repetition rate for the SOTA-inspired CNN. Moderate repetition rates (800 pulses/window) consistently lead to better relative performance than highly dense scenarios (5 000 pulses/window), even when the background noise level is increased. Parameter interactions Optimal model performance was achieved not by optimizing parameters in isolation but through their synergistic combination. While Table 6 quantifies the main effects of each parameter, interactive effects are visualized in Fig. 10. This figure highlights that the most substantial performance gains were realized by combining a moderate repetition rate (800 pulses/window) with a higher amplitude variance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\texttt {vegetation sigma} = 1.85$$\end{document}). This observation suggests that an optimal synthetic-pretraining strategy should simultaneously enhance the saliency of the target phenomenon (via higher mean and variance of PD amplitudes) and maintain a realistic event density analogous to field conditions. By systematically identifying and selecting these favorable parameter ranges—specifically, moderate PD repetition rates coupled with higher signal intensities and variances—we achieved a peak mean ROC AUC of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.910$$\end{document}. This result demonstrates that highly effective PD detectors can be developed with a primary reliance on a carefully optimized synthetic pretraining regime.Fig. 10Interaction between amplitude variance (vegetation sigma) and repetition rate for the SOTA-inspired CNN. The best relative performance is obtained for the region combining moderate repetition rates (800 pulses/window) with higher variance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\texttt {vegetation sigma} = 1.85$$\end{document}). Interaction between amplitude variance (vegetation sigma) and repetition rate for the SOTA-inspired CNN. The best relative performance is obtained for the region combining moderate repetition rates (800 pulses/window) with higher variance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\texttt {vegetation sigma} = 1.85$$\end{document}). Summary of key findings The statistical analyses highlighted three main conclusions: Pretraining effectiveness depended strongly on the dataset characteristics. While pretraining on synthetic data could enhance performance, the benefits varied significantly. Datasets simulating higher PD activity (Dataset 3, Dataset 4) led to substantial improvements, whereas other synthetic datasets (Dataset 1, Dataset 2, and All Datasets Combined) provided only marginal gains or, in some cases, degraded performance relative to no pretraining.Dataset 3 and Dataset 4 consistently yielded the highest performance Across all model types, pretraining with Dataset 3 or Dataset 4 resulted in significantly higher MCC, F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} score, and accuracy compared to other synthetic datasets and the no-pretraining baseline. The performance difference between Dataset 3 and Dataset 4 was generally not statistically significant, suggesting that multiple configurations simulating high PD activity can be effective.Skipping pretraining could be preferable to using suboptimal synthetic data While Dataset 3 and Dataset 4 clearly improved model performance, models pretrained on Dataset 1, Dataset 2, or the All Datasets Combined often performed no better, or sometimes worse, than those without any pretraining (No Pretraining). This indicated that ineffective pretraining could introduce biases or misleading feature representations, ultimately harming model performance. Pretraining effectiveness depended strongly on the dataset characteristics. While pretraining on synthetic data could enhance performance, the benefits varied significantly. Datasets simulating higher PD activity (Dataset 3, Dataset 4) led to substantial improvements, whereas other synthetic datasets (Dataset 1, Dataset 2, and All Datasets Combined) provided only marginal gains or, in some cases, degraded performance relative to no pretraining. Pretraining effectiveness depended strongly on the dataset characteristics. While pretraining on synthetic data could enhance performance, the benefits varied significantly. Datasets simulating higher PD activity (Dataset 3, Dataset 4) led to substantial improvements, whereas other synthetic datasets (Dataset 1, Dataset 2, and All Datasets Combined) provided only marginal gains or, in some cases, degraded performance relative to no pretraining. Dataset 3 and Dataset 4 consistently yielded the highest performance Across all model types, pretraining with Dataset 3 or Dataset 4 resulted in significantly higher MCC, F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} score, and accuracy compared to other synthetic datasets and the no-pretraining baseline. The performance difference between Dataset 3 and Dataset 4 was generally not statistically significant, suggesting that multiple configurations simulating high PD activity can be effective. Dataset 3 and Dataset 4 consistently yielded the highest performance Across all model types, pretraining with Dataset 3 or Dataset 4 resulted in significantly higher MCC, F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} score, and accuracy compared to other synthetic datasets and the no-pretraining baseline. The performance difference between Dataset 3 and Dataset 4 was generally not statistically significant, suggesting that multiple configurations simulating high PD activity can be effective. Skipping pretraining could be preferable to using suboptimal synthetic data While Dataset 3 and Dataset 4 clearly improved model performance, models pretrained on Dataset 1, Dataset 2, or the All Datasets Combined often performed no better, or sometimes worse, than those without any pretraining (No Pretraining). This indicated that ineffective pretraining could introduce biases or misleading feature representations, ultimately harming model performance. Skipping pretraining could be preferable to using suboptimal synthetic data While Dataset 3 and Dataset 4 clearly improved model performance, models pretrained on Dataset 1, Dataset 2, or the All Datasets Combined often performed no better, or sometimes worse, than those without any pretraining (No Pretraining). This indicated that ineffective pretraining could introduce biases or misleading feature representations, ultimately harming model performance. Discussion The present study demonstrates that synthetic-data pretraining can substantially enhance partial discharge (PD) detection in overhead transmission lines, albeit with strong dependence on the quality and characteristics of the synthetic datasets. The results presented in Fig. 1–3 and the accompanying statistical analyses underscore two main findings: (1) not all synthetic datasets confer the same degree of performance improvement, and (2) datasets simulating higher levels of PD activity, specifically Dataset 3 and Dataset 4 in our experiments, stand out as the most effective resources for pretraining. The superior performance associated with Dataset 3 and Dataset 4 can likely be attributed to their simulation parameters, which were designed to reflect higher PD activity (see Table 7). Specifically, the increased PD repetition rates (1600 and 3200 pulses/window, respectively, compared to 800 in Datasets 1/2) and higher amplitude variability (Veg. PD Amp. Sigma of 1.45 and 1.85, and higher mean in Dataset 4) likely exposed the pretrained models to a richer, more challenging set of discriminative features characteristic of significant or evolving fault conditions. This better prepared the models to identify such patterns within the noise and variability of the real test data. Interpretation of visual fidelity and pretraining effectiveness The visual discrepancies observed between synthetic and real signals, evident in both time-domain waveforms and spectrogram representations, require careful consideration regarding their impact on pretraining. While perfect visual replication across both domains was not achieved—particularly with respect to specific ambient noise signatures (e.g., noise floor texture, periodic interference) and the deliberate exaggeration of PD activity in datasets such as Dataset 3 and Dataset 4—the significant performance improvements obtained through pretraining strongly suggest that capturing the core functional characteristics of PD signals is more important than reproducing all visual details. Our hypothesis is that the synthetic data pipeline successfully embedded key discriminative features relevant to PD physics, recognizable in both time and frequency domains. This includes the sharp transient nature and amplitude variations of pulses (visible in time-domain plots) and their corresponding broadband spectral signatures (visible in spectrograms), together with realistic variations in repetition rates and clustering. The neural network, particularly when using a spectrogram input representation derived from these time-domain signals, appears to leverage these fundamental learned features effectively during pretraining. This, in turn, facilitates successful transfer to the real-data domain during fine-tuning, despite the apparent visual domain gaps in both representations. The notable success of Dataset 3, characterized by more intense PD activity, further supports the idea that exposing the model to a diverse range of functionally relevant signal variations—even if visually imperfect or slightly exaggerated—is highly beneficial for robust learning. This simulation-centric approach, which focuses on embedding physically meaningful characteristics, contrasts with generative approaches such as conditional GANs (cGANs) that often prioritize closer visual resemblance in the target domain25. In our own experiment with a cGAN trained on only 1,020 labeled real samples, the model was unstable and failed to generate high-fidelity samples; we attribute this behaviour to the acute scarcity of real data, which increases the risk of mode collapse and the appearance of unrealistic artifacts. By comparison, several studies reporting strong cGAN-based gains (e.g., MCC improvements from 0.701 to 0.795) rely on substantially larger and more diverse real datasets (often on the order of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$>10{,}000$$\end{document} samples) to learn a reliable data distribution12. Our setting is closer to the typical industrial scenario, where only a limited number of labeled real measurements is available (in our case, approximately 1,700 samples for fine-tuning), and under such conditions the physics-informed simulation route proves to be more stable and more controllable than purely data-driven generation. Taken together, these findings highlight an important implication for simulation-based pretraining: achieving perfect visual fidelity across all signal characteristics (including precise background noise replication) may be less critical than accurately capturing the time–frequency signature of PD pulses—their location, bandwidth, intensity, and repetition patterns. A simulation framework grounded in first principles and domain expertise is not primarily constrained by the size of an existing real dataset and can systematically generate a wide spectrum of plausible PD scenarios. This makes it a practical and robust pathway for pretraining in data-scarce environments, such as overhead transmission power line PD monitoring, while still remaining compatible with GAN-based augmentation in regimes where ample real data is available. Impact of suboptimal pretraining Although synthetic pretraining is conceptually appealing for mitigating data scarcity, the results caution that it can be detrimental if the synthetic data fail to mirror key real-world characteristics. Our experiments show that models pretrained on certain datasets (Dataset 1, Dataset 2, All Datasets Combined) do not always outperform models trained de novo on limited real data (No Pretraining). In fact, Dunn’s post-hoc comparisons indicate that these synthetic datasets sometimes lead to statistically lower MCC than no pretraining at all. This observation implies that ineffective synthetic data may inadvertently induce biased or misleading feature representations, contrasting sharply with the benefits seen from the more intense PD simulations in Datasets 3 and 4. From an operational standpoint, these findings underscore the importance of investing resources into the design or curation of synthetic datasets before field deployment. Rather than relying on a generic or off-the-shelf approach to synthetic PD waveforms, stakeholders should tailor the data-generation methodology to the specific conductor types, expected contamination levels, and vegetation interactions pertinent to the target environment. Such efforts can substantially reduce the risk of negative transfer, potentially prioritizing simulations that capture challenging or high-activity conditions. Architectural considerations for spectrogram-based PD detection The results of the architectural comparison indicate a clear preference for convolutional models in the context of spectrogram-based partial-discharge (PD) classification. The three CNN-based architectures consistently outperformed the recurrent and transformer-based variants. This outcome suggests that the most discriminative information for PD detection is encoded in the two-dimensional time—frequency structure of the spectrograms. Convolutional networks are inherently well aligned with this representation: their localized receptive fields and hierarchical feature extraction enable them to identify characteristic PD signatures, such as short broadband energy bursts, their orientation, and their relative positioning within the spectrogram. In practical terms, convolutional filters act as learnable detectors of these time–frequency patterns and can robustly separate PD-related activity from background interference. By contrast, the LSTM model processes the spectrogram as a temporal sequence of frequency vectors. Such a one-dimensional treatment is likely to discard or underutilize spatial correlations across the frequency axis, which appear to be essential for this task. The Vision Transformer model, while competitive in large-scale vision tasks, typically requires substantial pretraining (e.g. on ImageNet-scale corpora) to learn effective patch-level representations. Under the comparatively limited data conditions of this work (combining synthetic and real signals), the CNN backbone—notably the ResNet50V2 variant employed in the SOTA-Inspired configuration—benefits from strong inductive biases (locality, translation invariance) that make it more data-efficient and better matched to image-like inputs. Refining the strategy for synthetic data generation The granular ablation of synthetic-data parameters further refines the data-generation strategy. Slight but consistent improvements in MCC for higher vegetation-induced PD amplitudes (higher Veg Mean, cf. Fig. 8) indicate that pretraining on scenarios with more pronounced PD signatures is beneficial. This is in agreement with earlier experiments using synthetic datasets that emphasized higher PD activity: making the target phenomenon more salient during pretraining improves the model’s ability to recognize it under real, potentially noisier, conditions. An equally important observation is the robustness of the top CNN architectures to variations in synthetic noise level (Fig. 8). Once a CNN has learned the core PD signature, moderate changes in the background noise do not substantially degrade performance. This suggests the following design guideline for synthetic-data pipelines: while a realistic noise model remains desirable, the fidelity, variability, and frequency of the target PD events in the synthetic data are more critical than an exact reproduction of every noise component present in the field. Practical implications and FP/FN trade-offs From an operational standpoint, the ROC AUC results are particularly relevant. High AUC values (exceeding 0.9 for the SOTA-Inspired architecture) imply that the classifier maintains good separation between PD and non-PD classes over a broad range of decision thresholds. This flexibility allows system operators to select a threshold that prioritizes either high sensitivity (low FN rate) or high specificity (low FP rate), depending on the maintenance policy and cost structure. In power-distribution applications, false negatives are typically more costly, as undetected PD activity may lead to accelerated insulation degradation, unplanned outages, or safety risks. False positives, while undesirable, usually result only in additional inspection costs. A model with high ROC AUC thus enables a more favorable FP/FN operating point and supports economically viable predictive-maintenance workflows. Conversely, the inability of the LSTM and ViT models to achieve meaningful AUC scores indicates that these architectures, in the present setting, do not provide the level of discriminative control required for deployment in the field. Limitations and future research directions Residual synthetic–real domain gap Despite the promising results, a persistent limitation is the domain gap between synthetic and real-world PD signatures. Even though Dataset 3 and Dataset 4 reduce this discrepancy, current synthetic pipelines still simplify several non-ideal, site-specific, and time-varying phenomena. Our generator already models multiple interference classes—colored and impulsive noise, power-line harmonics, and narrowband RF carriers—to approximate the electromagnetic environment of medium-voltage overhead lines, but real deployments often exhibit structured, nonstationary components from power-electronic equipment, periodic or stochastic switching events, harmonic distortion, and local electromagnetic coupling. Environmental effects (severe weather, multi-source corona, vegetative canopies) are likewise only partially represented. Richer noise and interference modeling Future versions of the synthetic generator should therefore incorporate more expressive disturbance models. Concretely, this includes: (i) stochastic switching events with variable rise times, durations, and repetition patterns to emulate power-electronic and switching surges; (ii) clustered harmonics and amplitude- or frequency-modulated carriers to mimic coupled RF and power-electronic activity from nearby infrastructure; and (iii) adaptive, site-derived noise profiles learned from measured background spectra so that diurnal load changes, weather-induced coupling, or maintenance operations are reflected in the synthetic signals. These additions would allow the synthetic dataset to better approximate the spectral texture and temporal variability of real antenna- or SDR-based PD measurements, further narrowing the synthetic–real gap. Line- and asset-specific simulation and advanced domain adaptation Another direction is to make the simulation step more line- and asset-aware. Even within overhead distribution systems, conductor geometry, phase spacing, grounding layout, and proximity to other infrastructure affect propagation, attenuation, and interference patterns. Encoding these features into line-specific models would yield more faithful PD waveforms and reduce the burden on downstream adaptation. Furthermore, while our two-phase pretraining-and-fine-tuning workflow effectively transfers knowledge, performance could be further enhanced by incorporating more explicit domain adaptation techniques. A promising extension would be to introduce a lightweight domain-adversarial fine-tuning step following the initial pretraining. In such a setup, a domain classifier is trained to distinguish between features extracted from synthetic and real data, while the main feature extractor is trained to “fool” this classifier. This encourages the model to learn domain-invariant representations that are robust to the subtle differences between the two data sources28,29. Similarly, contrastive domain-mapping methods could be employed to pull feature representations of the same class (e.g., “PD present”) from both synthetic and real domains closer together in the embedding space30,31. Integrating these advanced adaptation strategies represents a compelling avenue for future research, potentially offering further performance gains by more explicitly and systematically closing the synthetic-real gap. Topology-agnostic extension of the framework Although the present study focuses on covered conductors, the proposed hybrid synthetic-data pretraining strategy is largely topology-agnostic. PD monitoring is routinely applied to power transformers, XLPE cables, rotating machines32, and GIS installations33. The same pretraining–fine-tuning workflow can be transferred to these assets by adjusting a limited set of parameters—discharge amplitude distributions, propagation and reflection coefficients, expected interference profiles, and background noise models—to match the target environment (including bare conductors or underground cables). Benchmarking, deployment, and model drift Progress in the community will benefit from structured benchmarks that combine synthetic and real measurements from diverse overhead-line configurations, enabling reproducible and comparable evaluations across modeling approaches. Our results already indicate that the achieved inference latency (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}12.3 ms) is compatible with real-time stream processing, but long-term field deployment introduces model drift due to seasonal changes, conductor aging, or emerging interference sources. This motivates adaptive or continuous learning schemes—periodic fine-tuning on newly collected field data, semi-supervised exploitation of unlabeled streams, or online domain adaptation—to preserve accuracy over time. A practical path is to train a centrally maintained foundation model on a large, diversified synthetic corpus and to fine-tune it with small, site-specific datasets for rapid rollout. To increase operator trust and ease integration into SCADA or asset-management systems, future work should also integrate explainability (e.g., Grad-CAM on spectrogram inputs34) so that alarms can be traced back to the responsible time–frequency structures. Computational cost and practical applicability To assess the real-world feasibility of our proposed pipeline, we evaluated the computational costs associated with both synthetic data generation and model training. The synthetic data generation process was performed on a multi-core CPU environment using 32 threads of an Intel(R) Xeon(R) CPU E5-2630 v4 processor. The generation of 1,000 synthetic samples took approximately 96 minutes. It is important to note that this CPU-based generation python script was not performance-optimized; significant acceleration could be achieved through code vectorization or by porting the process to a GPU. As data generation is a one-time, offline task, this cost is considered manageable for creating large-scale pretraining datasets. The model training and inference phases were conducted on a single, consumer-grade NVIDIA GeForce GTX 1070 GPU, demonstrating that the approach does not require high-end computational resources. The training process was highly efficient. On average, the synthetic pretraining phase for 1000 samples completed in approximately 117 seconds (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\approx 2$$\end{document} minutes). The subsequent two-phase fine-tuning on real data required an average of 144 seconds for head-only training and 353 seconds for full-model fine-tuning, culminating in a total fine-tuning time of under 9 minutes per run. The total training time for a single model, including pretraining and fine-tuning, was therefore just over 10 minutes. For deployment, the inference speed is critical. Our trained models exhibited a mean latency of approximately 12.3 ms per spectrogram. This high throughput confirms the suitability of the models for near-real-time monitoring applications. Overall, the low training overhead and fast inference speeds underscore the practical applicability and cost-effectiveness of our synthetic pretraining framework for real-world PD detection systems. The practical relevance of these performance improvements goes beyond achieving better statistical indicators. In routine grid maintenance, the consequences of misclassification are strongly asymmetric. A false negative (FN)—i.e., failure to detect an actual partial discharge—may precipitate severe equipment damage, unplanned outages, and elevated safety risks, all of which translate into substantial financial and reputational losses. By contrast, a false positive (FP)—i.e., incorrectly flagging a healthy asset—typically leads only to an unnecessary inspection, which is comparatively less costly, though still operationally undesirable. In this setting, the higher Matthews Correlation Coefficient (MCC) obtained by models pretrained on Dataset 3 and Dataset 4 is particularly meaningful, because MCC offers a balanced assessment of classifier performance under class imbalance and unequal error costs. A model that simultaneously suppresses FNs and limits FPs improves the reliability of discharge detection and, at the same time, lowers the expected operational cost. Consequently, such models support more efficient and economically sustainable predictive-maintenance policies for overhead distribution networks. Concluding remarks In summary, our results confirm that the careful design and use of synthetic data can be a powerful tool for overcoming data limitations in PD detection for overhead transmission lines. While not all synthetic datasets yield equally positive outcomes, the marked success of Dataset 3 and Dataset 4 underscores the pivotal role of aligning simulated conditions with real operational complexities. By strategically integrating synthetic pretraining and selective fine-tuning, utilities and researchers can achieve heightened classification accuracy, faster model convergence, and reduced costs for data acquisition and annotation. These findings pave the way toward broader adoption of data-driven PD monitoring systems, improving insulation health assessment, reducing the risk of service interruptions, and supporting more robust grid operation in the long term. Methods Data acquisition and real dataset We employed a real-world, hourly sampled dataset introduced in13, collected using contactless antenna-based data acquisition systems installed on medium-voltage overhead transmission power lines. Each measurement was a 20 ms snapshot (one full mains cycle) sampled at 40 MS/s, yielding 800000 samples per snapshot6,7. The snapshot acquisition was repeated at hourly intervals. Expert-curated ground-truth labels (fault vs. no fault) were provided by an auxiliary galvanic contact sensor13, ensuring reliable annotations for model training and evaluation. The dataset is publicly available under:13. To develop and assess our models, we partitioned the real dataset into three subsets: 60% for training, 20% for validation, and the remaining 20% for testing. This split preserved ample data for both model development and unbiased performance estimation on held-out samples. Synthetic data generation To complement and augment the real data, we generated synthetic wideband signals resembling on-site measurements. These signals incorporated partial discharge (PD) events, corona activity, vegetation contact phenomena, and various noise and interference sources. The goal was to produce realistic time-domain waveforms that captured the complex electromagnetic environment observed in real overhead power lines. We also publish our code on https://github.com/Lukykl1/synth-pd-datahttps://github.com/Lukykl1/synth-pd-data as open source. The whole pipeline is captured in Fig. 11. To augment the real dataset and facilitate pretraining, we synthesized wideband signals that closely resembled on-site overhead-line measurements. These synthetic signals incorporated a range of phenomena:Vegetation contact pulses (PDs),Corona bursts,Vegetation-induced contact pulses,Reflections due to line discontinuities,Multiple noise sources (colored noise, impulsive noise, flicker, RF interference),Power-line interference (fundamental and harmonics),Antenna and amplifier responses.The overarching objective was to generate time-domain waveforms that captured the complex electromagnetic environment typically observed around medium-voltage overhead transmission power lines.Fig. 11Flowchart of the Synthetic Signal Generation Pipeline. The process begins with generating individual source signals (1), which are summed and then passed through a channel and sensor simulation (2). The final time-domain signal is converted to a spectrogram (3). The entire process is guided by a calibration loop (4) where synthetic outputs are compared against real-world data to refine simulation parameters. Vegetation contact pulses (PDs), Vegetation contact pulses (PDs), Corona bursts, Corona bursts, Vegetation-induced contact pulses, Vegetation-induced contact pulses, Reflections due to line discontinuities, Reflections due to line discontinuities, Multiple noise sources (colored noise, impulsive noise, flicker, RF interference), Multiple noise sources (colored noise, impulsive noise, flicker, RF interference), Power-line interference (fundamental and harmonics), Power-line interference (fundamental and harmonics), Antenna and amplifier responses. Antenna and amplifier responses. Flowchart of the Synthetic Signal Generation Pipeline. The process begins with generating individual source signals (1), which are summed and then passed through a channel and sensor simulation (2). The final time-domain signal is converted to a spectrogram (3). The entire process is guided by a calibration loop (4) where synthetic outputs are compared against real-world data to refine simulation parameters. Signal construction overview We synthesized a time-domain signal of duration \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document}, sampled at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{s}$$\end{document}. Thus, each waveform had\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} N \;=\;\lfloor T \cdot F_{s}\rfloor \end{aligned}$$\end{document}samples. In the time domain, we summed all pulses, noise, and interference components:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x(t) \;=\; x_{\textrm{PD}}(t) + x_{\textrm{corona}}(t) + x_{\textrm{veg}}(t) + x_{\textrm{noise}}(t) + x_{\textrm{PL}}(t) + x_{\textrm{ref}}(t), \end{aligned}$$\end{document}after which we applied frequency-dependent overhead-line attenuation, antenna response, and amplifier nonlinearity. Detailed descriptions of each component followed. Overhead transmission power line attenuation and phase shift We assumed that the signal propagated along an overhead transmission power line of length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d$$\end{document}. The line was characterized by a frequency-dependent attenuation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha (f)$$\end{document} and a phase constant \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta (f)$$\end{document}. For a signal at frequency \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha (f)$$\end{document} (in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {Np/m}$$\end{document}) described amplitude decay, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta (f)$$\end{document} governed phase shift. A commonly used model was:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \alpha (f) \;=\; a \sqrt{\frac{f}{f_{\textrm{ref}}}} \;+\; b \;+\; \sigma \left( \frac{f^{\,0.4}}{f_{\textrm{ref}}^{\,0.4}}\right) \;+\;\eta \, c \, f^{\,0.3}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a, b, c$$\end{document} were empirical constants, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document} reflected ground conductivity, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta$$\end{document} captured humidity, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\textrm{ref}}$$\end{document} was a reference frequency (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\,\textrm{MHz}$$\end{document}). The phase constant \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta (f)$$\end{document} was2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \beta (f) \;=\; \frac{2\pi \,f}{v_{p}}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{p}=\gamma \,c$$\end{document} was the phase velocity along the line (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma <1$$\end{document} to model subluminal propagation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document} is the speed of light). The line transfer function was then\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} H_{\textrm{line}}(f) \;=\; \exp \!\Bigl [-(\alpha (f) + j\,\beta (f))\,d\Bigr ]. \end{aligned}$$\end{document} Antenna response To simulate the receive antenna behavior, we included a frequency-dependent antenna gain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_{\textrm{ant}}(f)$$\end{document}. A logistic fade-in was used to transition from near-zero gain at lower frequencies, and a fade-out was applied above some upper frequency. Conceptually,3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} H_{\textrm{ant}}(f) \;=\; G \; \Phi _{\textrm{in}}(f) \; \Phi _{\textrm{out}}(f), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G$$\end{document} was the peak linear gain, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi _{\textrm{in}}(f)$$\end{document} modeled the lower-frequency gain ramp-up, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi _{\textrm{out}}(f)$$\end{document} modeled the high-frequency roll-off. Amplifier nonlinearity The received and filtered signal was passed through a mild amplifier saturation stage to replicate real-world amplifier behavior. We modeled the output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{\textrm{amp}}(t)$$\end{document} by4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{\textrm{amp}}(t) \;=\; \frac{x_{\textrm{in}}(t)}{\sqrt{\,1 + \Bigl (\frac{x_{\textrm{in}}(t)}{S_{\textrm{sat}}}\Bigr )^{2}}\,}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{sat}}$$\end{document} was the saturation threshold, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{\textrm{in}}(t)$$\end{document} was the input signal to the amplifier. Partial discharges and corona bursts We modeled both partial discharge (PD) pulses and corona discharges as damped sinusoids. A single pulse that began at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{0}$$\end{document} was described by5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} s(t) \;=\; {\left\{ \begin{array}{ll} A \,\exp \!\Bigl [-\alpha _{d}\,(t - t_{0})\Bigr ] \;\sin \!\Bigl [2\pi f_{\textrm{c}}(t - t_{0})\,Q\Bigr ], & t \ge t_{0}, \\ 0, & t < t_{0}, \end{array}\right. } \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A$$\end{document} was the amplitude, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _{d}$$\end{document} was the damping factor, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\textrm{c}}$$\end{document} was the center frequency, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q$$\end{document} controlled the ring-down rate. Corona discharge appeared as continuous, weaker pulses whose amplitude typically followed the line voltage cycle (50 Hz or 60 Hz). In the simulation, many low-amplitude corona pulses were generated over time, with statistical variability in repetition rate, amplitude, and damping parameters that depended on environmental factors such as humidity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta$$\end{document}) and altitude. Partial discharges in high-voltage networks appeared as broadband, short-lived transients. We used the same damped sinusoid form in (5) but with distinct parameter ranges (e.g., higher damping rates and broader frequency content). The amplitude and time of occurrence were drawn from empirically motivated or log-normal distributions, reflecting the variability found in real-world PD signals. Vegetation contact pulses Vegetation contact events were modeled as bursts of PD-like pulses over short windows (from sub-milliseconds to a few milliseconds). In such intervals:The repetition rate (hundreds to thousands of pulses per second) was higher than typical corona,Pulse amplitudes could exceed typical corona amplitudes,Crosstalk among phases could occur if vegetation simultaneously contacted multiple conductors.This burst-like behavior emulated the rapid succession of PD events when foliage physically contacted or brushed against an energized conductor. The repetition rate (hundreds to thousands of pulses per second) was higher than typical corona, The repetition rate (hundreds to thousands of pulses per second) was higher than typical corona, Pulse amplitudes could exceed typical corona amplitudes, Pulse amplitudes could exceed typical corona amplitudes, Crosstalk among phases could occur if vegetation simultaneously contacted multiple conductors. Crosstalk among phases could occur if vegetation simultaneously contacted multiple conductors. Reflections Reflections resulting from discontinuities in the line or nearby structures were introduced as scaled, time-delayed copies of the original signal. For a reflection path with distance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{\textrm{ref}}$$\end{document} and reflection coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Gamma$$\end{document}, the reflected component was6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{\textrm{ref}}(t) \;=\; \Gamma \,x\bigl (t-\tau \bigr ), \quad \tau \;=\; \frac{2 \,d_{\textrm{ref}}}{\gamma \,c}. \end{aligned}$$\end{document}Multiple reflections, each with unique \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(d_{\textrm{ref}}, \Gamma )$$\end{document}, were summed to emulate complex propagation paths. Noise and interference To replicate the electromagnetic interference environment around overhead transmission power lines, various noise sources were incorporated. One major source was power-line interference, which consisted of sinusoidal components at the fundamental line frequency (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_1 \approx 50\,\textrm{Hz}$$\end{document} or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$60\,\textrm{Hz}$$\end{document}) and its harmonics. The resulting signal could be expressed as7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{\textrm{PL}}(t) \;=\; \sum _{k=1}^{H} \frac{A_{m}(t)}{k}\; \sin \Bigl [2\pi \!\int _{0}^{t} f_{k}(\tau )\,d\tau + \phi _{k}\Bigr ], \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H$$\end{document} represented the number of harmonics, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_{m}(t)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{k}(t)$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi _{k}$$\end{document} captured time-varying amplitude, frequency, and phase drifts, respectively. Another important component was colored noise, characterized by a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1/f^\alpha$$\end{document} spectral profile to account for the stronger low-frequency components commonly observed in practical measurements. To reflect changing background conditions, the noise amplitude was varied in different segments. In addition to these continuous disturbances, impulsive noise was introduced in the form of high-amplitude, short-duration impulses that occurred at random instances. These impulses served to simulate transient events such as switching actions, arcing, or other sudden electrical phenomena. Further, broadcast carriers at radio frequencies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{b}$$\end{document} were included as narrowband sinusoidal tones, modeled as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{\textrm{BC}}(t) \;=\; \sum _{b} P_{b}\,\sin \bigl (2 \pi f_{b}\,t + \phi _{b}\bigr ), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{b}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi _{b}$$\end{document} denoted the carrier power and phase, respectively. A sparsity model was introduced to emulate intermittent transmissions, adding another layer of complexity to the simulated interference environment. Composite channel and data generation The total transfer function for the line and antenna was given by\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} H_{\textrm{total}}(f) \;=\; H_{\textrm{line}}(f)\,H_{\textrm{ant}}(f). \end{aligned}$$\end{document}After summing all pulses, noise, and interference in the time domain, we transformed the signal to the frequency domain via an FFT, multiplied by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_{\textrm{total}}(f)$$\end{document}, and then applied the inverse FFT to simulate channel effects. Finally, we applied the amplifier nonlinearity in (4). Synthetic signals were labeled according to the presence or absence of vegetation-induced partial discharge bursts within the simulated window. By sampling random parameters (e.g., pulse counts, amplitudes, reflection coefficients, noise intensities) across many iterations, we constructed a large library of synthetic data suitable for model training and testing. Synthetic datasets Each synthetic example was labeled \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1}$$\end{document} if it contained vegetation contact pulses and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0}$$\end{document} otherwise. By randomizing parameters (e.g., PD amplitudes, reflection distances, noise intensities) within realistic ranges, we generated diverse training samples. Different parameters for each dataset could be seen in Table 7. The specific parameter variations across Datasets 1-4 (Table 7) were chosen to systematically investigate the impact of increasing noise levels (Dataset 2 vs 1) and, more significantly, increasing PD activity intensity via repetition rate and amplitude statistics (Dataset 3 vs 1/2, Dataset 4 vs 3). This range was intended to reflect conditions from moderate background noise and PD activity up to potentially severe fault scenarios. Class balance To prevent the classifier from learning a trivial decision rule, all synthetic datasets were generated with an explicitly balanced class distribution, i.e. a 1:1 ratio between PD-containing ((y = 1)) and non-PD ((y = 0)) samples. No additional oversampling, undersampling, or class-weighting was applied on top of this synthetic balance. In contrast, the real-world evaluation datasets preserve their natural (typically more non-PD–heavy) class proportions, because our objective was to test how models trained on balanced synthetic data transfer to operational conditions.Table 7Comparison of key parameters across datasets.DatasetNoise base levelVeg. PD RateVeg. PD Amp. SigmaVeg. PD Amp. MeanDataset 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0 \times 10^5$$\end{document}8000.452.0Dataset 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0 \times 10^6$$\end{document}8001.452.0Dataset 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0 \times 10^5$$\end{document}16001.452.0Dataset 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0 \times 10^5$$\end{document}32001.852.2 Table 7Comparison of key parameters across datasets.DatasetNoise base levelVeg. PD RateVeg. PD Amp. SigmaVeg. PD Amp. MeanDataset 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0 \times 10^5$$\end{document}8000.452.0Dataset 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0 \times 10^6$$\end{document}8001.452.0Dataset 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0 \times 10^5$$\end{document}16001.452.0Dataset 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0 \times 10^5$$\end{document}32001.852.2 Comparison of key parameters across datasets. To explore different levels of complexity and noise, we produced four synthetic collections:Dataset 1 Moderate PD rates, moderate noise,Dataset 2 Higher noise levels and more pronounced PD,Dataset 3 Very frequent PD events with strong interference,Dataset 4 Extremely frequent and intense PD with high vegetation activity,All Datasets Combined A combined pool of the above four.All synthetic waveforms were then converted into spectrograms using a Short-Time Fourier Transform (STFT) with a 1024-sample window, 512-sample overlap, and a final resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$224 \times 224$$\end{document}. Figure 12 illustrated this spectrogram-generation pipeline. Dataset counts could be seen in Table 8.Table 8Sample counts for synthetic and real datasets used in training, validation, and testing.Dataset typeSplit/totalTotal samplesPositive samples (+)Negative samples (−)SyntheticDataset 1 (Pretrain)23,82811,91411,914Dataset 2 (Pretrain)15,5007,7507,750Dataset 3 (Pretrain)1,550775775Dataset 4 (Pretrain)1,550775775All Combined (Pretrain)42,42821,98921,989RealTrain (60%)1,020108912Validation (20%)34036304Test (20%)34036304Total1,7001801,520 Dataset 1 Moderate PD rates, moderate noise, Dataset 1 Moderate PD rates, moderate noise, Dataset 2 Higher noise levels and more pronounced PD, Dataset 2 Higher noise levels and more pronounced PD, Dataset 3 Very frequent PD events with strong interference, Dataset 3 Very frequent PD events with strong interference, Dataset 4 Extremely frequent and intense PD with high vegetation activity, Dataset 4 Extremely frequent and intense PD with high vegetation activity, All Datasets Combined A combined pool of the above four. All Datasets Combined A combined pool of the above four. Table 8Sample counts for synthetic and real datasets used in training, validation, and testing.Dataset typeSplit/totalTotal samplesPositive samples (+)Negative samples (−)SyntheticDataset 1 (Pretrain)23,82811,91411,914Dataset 2 (Pretrain)15,5007,7507,750Dataset 3 (Pretrain)1,550775775Dataset 4 (Pretrain)1,550775775All Combined (Pretrain)42,42821,98921,989RealTrain (60%)1,020108912Validation (20%)34036304Test (20%)34036304Total1,7001801,520 Sample counts for synthetic and real datasets used in training, validation, and testing. Evaluation metric For binary detection tasks such as identifying vegetation contact, we adopted the MCC to provide a balanced performance measure:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textrm{MCC} \;=\; \frac{(TP \times TN) - (FP \times FN)}{\sqrt{(TP+FP)\,(TP+FN)\,(TN+FP)\,(TN+FN)}}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TP$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TN$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$FP$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$FN$$\end{document} denoted true positives, true negatives, false positives, and false negatives, respectively. The MCC was particularly useful for imbalanced datasets, offering a single-value metric that captured all four confusion-matrix components. Implementation and usage notes All simulation steps were implemented in a numeric computing language and efficiently parallelized due to the independence of individual waveforms. Key parameters such as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha (f)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\textrm{c}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta$$\end{document}, reflection distances, and noise intensities were sampled from user-defined distributions, ensuring a diverse and representative synthetic dataset. As a final step, spectrograms or time-domain signals were stored for subsequent machine learning classification or signal-processing experiments. This approach enabled rapid prototyping and systematic testing of algorithms aimed at detecting partial discharge events, vegetation intrusions, and other transient phenomena in overhead transmission power lines.Fig. 12Synthetic spectrogram generation pipeline. Synthetic spectrogram generation pipeline. Model architecture and training Following our previous work35, we employed a well-established and reliable methodology for the classification of partial discharges. Specifically, we utilized a spectrogram-based classification approach leveraging a two-dimensional convolutional neural network (2D-CNN) with a ResNet50V2 backbone36. To systematically explore the impact of different classification heads, we experimented with four distinct top-layer configurations: Top-layer architectures We adopted a ResNet50V2 backbone and evaluated four distinct classification heads. Their architectural details and characteristics were summarized in Table 9.Table 9Comparison of classification heads used in the model.Head typeKey featuresMax-PoolingStandard max-pooling classification head.Traditional ssBased on prior work, employed batch normalization for generalization.SOTA-InspiredDeeper architecture with L2 regularization and dropout. Table 9Comparison of classification heads used in the model.Head typeKey featuresMax-PoolingStandard max-pooling classification head.Traditional ssBased on prior work, employed batch normalization for generalization.SOTA-InspiredDeeper architecture with L2 regularization and dropout. Comparison of classification heads used in the model. The classification heads differed primarily in how they processed extracted features. The Max-Pooling head applied a global max-pooling operation followed by a dense layer with 512 units, dropout regularization, and a final dense layer for classification. The Traditional head, commonly used in prior work, utilized global average pooling with a dropout rate of 0.7 before passing features to a single dense output layer. Inspired by state-of-the-art designs, the SOTA-Inspired head employed a deeper architecture, beginning with global average pooling, followed by batch normalization and two fully connected layers (1024 and 512 units) with L2 regularization, interleaved with dropout layers for improved stability. The input to the network consisted of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$224 \times 224 \times 3$$\end{document} spectrograms (where the three channels were copies of the single-channel spectrogram magnitude, formatted for compatibility with the pretrained ResNet). The models were trained to minimize the Binary Cross-Entropy loss function, suitable for this binary classification task. Extended ablation study To obtain a broader architectural perspective and a more fine-grained view of the synthetic-data space, we conducted an extended ablation study composed of two main components. Additional architectures To assess whether the trends observed for convolutional networks also hold for non-convolutional models, two alternative architectures were added to the benchmark: LSTM-based classifier This model employs a recurrent head based on Long Short-Term Memory (LSTM) units and processes the spectrogram as a sequence of temporal slices. In this setup, the spectrogram’s time axis is treated as an ordered sequence, allowing the model to capture long-range temporal dependencies and clustered PD activity.Vision Transformer A standard Vision Transformer (ViT) was included as a strong, non-convolutional baseline. The spectrogram is first partitioned into patches, which are then encoded as tokens and processed using multi-head self-attention. Since ViT architectures represent the state of the art in many computer-vision tasks, this model provides a meaningful point of comparison beyond CNNs.Granular parameter sweep Instead of relying on a small set of fixed synthetic datasets, we systematically varied the main generation parameters of the synthetic PD-signal pipeline in order to analyze their individual and combined effects on classification performance. The following parameters were swept:Noise base level three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$200\,\textrm{k}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$500\,\textrm{k}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\,\textrm{M}$$\end{document};Vegetation PD repetition rate (Veg Rate) three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$800$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\,200$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\,000$$\end{document} pulses/window;Vegetation PD amplitude sigma (Veg Sigma) three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.45$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.85$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.0$$\end{document};Vegetation PD amplitude mean (Veg Mean) two levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2$$\end{document}.This design makes it possible to study how noise conditions, PD occurrence statistics, and amplitude distributions influence model robustness, and whether certain architectural choices (convolutional, recurrent, attention-based) are more sensitive to particular types of synthetic variation. LSTM-based classifier This model employs a recurrent head based on Long Short-Term Memory (LSTM) units and processes the spectrogram as a sequence of temporal slices. In this setup, the spectrogram’s time axis is treated as an ordered sequence, allowing the model to capture long-range temporal dependencies and clustered PD activity. LSTM-based classifier This model employs a recurrent head based on Long Short-Term Memory (LSTM) units and processes the spectrogram as a sequence of temporal slices. In this setup, the spectrogram’s time axis is treated as an ordered sequence, allowing the model to capture long-range temporal dependencies and clustered PD activity. Vision Transformer A standard Vision Transformer (ViT) was included as a strong, non-convolutional baseline. The spectrogram is first partitioned into patches, which are then encoded as tokens and processed using multi-head self-attention. Since ViT architectures represent the state of the art in many computer-vision tasks, this model provides a meaningful point of comparison beyond CNNs. Vision Transformer A standard Vision Transformer (ViT) was included as a strong, non-convolutional baseline. The spectrogram is first partitioned into patches, which are then encoded as tokens and processed using multi-head self-attention. Since ViT architectures represent the state of the art in many computer-vision tasks, this model provides a meaningful point of comparison beyond CNNs. Noise base level three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$200\,\textrm{k}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$500\,\textrm{k}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\,\textrm{M}$$\end{document}; Noise base level three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$200\,\textrm{k}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$500\,\textrm{k}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\,\textrm{M}$$\end{document}; Vegetation PD repetition rate (Veg Rate) three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$800$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\,200$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\,000$$\end{document} pulses/window; Vegetation PD repetition rate (Veg Rate) three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$800$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\,200$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\,000$$\end{document} pulses/window; Vegetation PD amplitude sigma (Veg Sigma) three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.45$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.85$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.0$$\end{document}; Vegetation PD amplitude sigma (Veg Sigma) three levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.45$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.85$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.0$$\end{document}; Vegetation PD amplitude mean (Veg Mean) two levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2$$\end{document}. Vegetation PD amplitude mean (Veg Mean) two levels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2$$\end{document}. Training protocol For all experiments, the training procedure described in Section 4.4 was applied uniformly across all architectures and all synthetic-data configurations. All networks were trained on spectrogram inputs of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$224 \times 224 \times 3$$\end{document}, i.e. three-channel image-like representations derived from the original time-domain PD signals. Training procedure The models were trained using the Adam optimizer37 with an initial learning rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-4}$$\end{document}, minimizing the binary cross-entropy loss function, which is appropriate for this binary classification task. To dynamically adjust the learning rate during training, a ReduceLROnPlateau learning rate scheduler was employed. This scheduler monitored the validation loss and reduced the learning rate by a factor of 0.2 if the monitored metric did not improve for 5 epochs. Given that synthetic pretraining was a key objective, we implemented the following workflow:Optional pretraining step When synthetic pretraining was performed, the ResNet50V2 model (with its classification head) was first trained on the designated synthetic dataset (Dataset 1, 2, 3, 4, or All Combined) for a maximum of 12 epochs. Early stopping based on synthetic validation loss with a patience of 15 epochs. This stage aimed to initialize the ResNet backbone with potentially useful features learned from the synthetic domain before fine-tuning on the limited real data. For comparison, ablation experiments omitting this pretraining step were also conducted (No Pretraining condition).Training on real data (two phases) Following the optional pretraining, or starting directly for the No Pretraining condition, training proceeded on the real-world dataset split into two distinct phases: Phase 1 (head training) The weights of the ResNet50V2 backbone were initially frozen. Only the randomly initialized or pretrained classification head was trained on the real training data. This phase ran for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_1 = 12$$\end{document} epochs. Early stopping was utilized, monitoring the validation loss on the real validation set with a patience of 15 epochs. Model checkpointing saved the model weights that achieved the best performance on the monitored validation metric during this phase.Phase 2 (full fine-tuning) Subsequently, all layers of the network, including the ResNet backbone, were unfrozen. The entire model was then fine-tuned end-to-end on the real training data, typically starting with a lower learning rate (e.g., automatically reduced by the scheduler or reset to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-4}$$\end{document}). This fine-tuning phase continued for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_2 = 40$$\end{document} epochs. Similar to Phase 1, early stopping monitored the validation loss on the real validation set with a patience of 15 epochs. The final model selected and reported for evaluation was the one corresponding to the weights that achieved the best validation performance during this critical fine-tuning phase, as ensured by model checkpointing.Throughout the training and validation process, key evaluation metrics including accuracy, precision, recall, and particularly the Matthews Correlation Coefficient (MCC)38 were tracked to assess model performance. Our empirical observations consistently indicated that omitting Phase 1 (training the head first before unfreezing the backbone) generally resulted in suboptimal final model performance, underscoring the necessity of this staged training approach, especially when fine-tuning deep networks on limited target datasets. Optional pretraining step When synthetic pretraining was performed, the ResNet50V2 model (with its classification head) was first trained on the designated synthetic dataset (Dataset 1, 2, 3, 4, or All Combined) for a maximum of 12 epochs. Early stopping based on synthetic validation loss with a patience of 15 epochs. This stage aimed to initialize the ResNet backbone with potentially useful features learned from the synthetic domain before fine-tuning on the limited real data. For comparison, ablation experiments omitting this pretraining step were also conducted (No Pretraining condition). Optional pretraining step When synthetic pretraining was performed, the ResNet50V2 model (with its classification head) was first trained on the designated synthetic dataset (Dataset 1, 2, 3, 4, or All Combined) for a maximum of 12 epochs. Early stopping based on synthetic validation loss with a patience of 15 epochs. This stage aimed to initialize the ResNet backbone with potentially useful features learned from the synthetic domain before fine-tuning on the limited real data. For comparison, ablation experiments omitting this pretraining step were also conducted (No Pretraining condition). Training on real data (two phases) Following the optional pretraining, or starting directly for the No Pretraining condition, training proceeded on the real-world dataset split into two distinct phases: Phase 1 (head training) The weights of the ResNet50V2 backbone were initially frozen. Only the randomly initialized or pretrained classification head was trained on the real training data. This phase ran for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_1 = 12$$\end{document} epochs. Early stopping was utilized, monitoring the validation loss on the real validation set with a patience of 15 epochs. Model checkpointing saved the model weights that achieved the best performance on the monitored validation metric during this phase.Phase 2 (full fine-tuning) Subsequently, all layers of the network, including the ResNet backbone, were unfrozen. The entire model was then fine-tuned end-to-end on the real training data, typically starting with a lower learning rate (e.g., automatically reduced by the scheduler or reset to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-4}$$\end{document}). This fine-tuning phase continued for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_2 = 40$$\end{document} epochs. Similar to Phase 1, early stopping monitored the validation loss on the real validation set with a patience of 15 epochs. The final model selected and reported for evaluation was the one corresponding to the weights that achieved the best validation performance during this critical fine-tuning phase, as ensured by model checkpointing. Training on real data (two phases) Following the optional pretraining, or starting directly for the No Pretraining condition, training proceeded on the real-world dataset split into two distinct phases: Phase 1 (head training) The weights of the ResNet50V2 backbone were initially frozen. Only the randomly initialized or pretrained classification head was trained on the real training data. This phase ran for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_1 = 12$$\end{document} epochs. Early stopping was utilized, monitoring the validation loss on the real validation set with a patience of 15 epochs. Model checkpointing saved the model weights that achieved the best performance on the monitored validation metric during this phase.Phase 2 (full fine-tuning) Subsequently, all layers of the network, including the ResNet backbone, were unfrozen. The entire model was then fine-tuned end-to-end on the real training data, typically starting with a lower learning rate (e.g., automatically reduced by the scheduler or reset to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-4}$$\end{document}). This fine-tuning phase continued for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_2 = 40$$\end{document} epochs. Similar to Phase 1, early stopping monitored the validation loss on the real validation set with a patience of 15 epochs. The final model selected and reported for evaluation was the one corresponding to the weights that achieved the best validation performance during this critical fine-tuning phase, as ensured by model checkpointing. Phase 1 (head training) The weights of the ResNet50V2 backbone were initially frozen. Only the randomly initialized or pretrained classification head was trained on the real training data. This phase ran for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_1 = 12$$\end{document} epochs. Early stopping was utilized, monitoring the validation loss on the real validation set with a patience of 15 epochs. Model checkpointing saved the model weights that achieved the best performance on the monitored validation metric during this phase. Phase 1 (head training) The weights of the ResNet50V2 backbone were initially frozen. Only the randomly initialized or pretrained classification head was trained on the real training data. This phase ran for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_1 = 12$$\end{document} epochs. Early stopping was utilized, monitoring the validation loss on the real validation set with a patience of 15 epochs. Model checkpointing saved the model weights that achieved the best performance on the monitored validation metric during this phase. Phase 2 (full fine-tuning) Subsequently, all layers of the network, including the ResNet backbone, were unfrozen. The entire model was then fine-tuned end-to-end on the real training data, typically starting with a lower learning rate (e.g., automatically reduced by the scheduler or reset to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-4}$$\end{document}). This fine-tuning phase continued for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_2 = 40$$\end{document} epochs. Similar to Phase 1, early stopping monitored the validation loss on the real validation set with a patience of 15 epochs. The final model selected and reported for evaluation was the one corresponding to the weights that achieved the best validation performance during this critical fine-tuning phase, as ensured by model checkpointing. Phase 2 (full fine-tuning) Subsequently, all layers of the network, including the ResNet backbone, were unfrozen. The entire model was then fine-tuned end-to-end on the real training data, typically starting with a lower learning rate (e.g., automatically reduced by the scheduler or reset to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-4}$$\end{document}). This fine-tuning phase continued for a maximum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_2 = 40$$\end{document} epochs. Similar to Phase 1, early stopping monitored the validation loss on the real validation set with a patience of 15 epochs. The final model selected and reported for evaluation was the one corresponding to the weights that achieved the best validation performance during this critical fine-tuning phase, as ensured by model checkpointing. Evaluation protocol To ensure a rigorous assessment, model performance was evaluated using a dedicated real-world test set. The evaluation followed a structured approach that included several key strategies. First, we analyzed the confusion matrix, recording the counts of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) for each model. This allowed us to gain insights into the types of errors made and their impact on overall performance. Additionally, we computed the MCC, a robust metric particularly suited for imbalanced binary classification tasks, providing a more reliable measure of predictive quality than simple accuracy. Beyond these core metrics, we conducted a thorough sensitivity analysis to examine how various factors influenced model performance. Specifically, we investigated the impact of different synthetic dataset configurations, variations in classification head architectures, and the inclusion or exclusion of synthetic pretraining. To ensure the validity of our measurements, we performed at least 31 random cross-validations with different seeds. This helped mitigate the impact of randomness in data splits and provided a more reliable estimate of model performance. By integrating these evaluation strategies, we ensured a comprehensive understanding of the proposed models’ robustness and generalization capabilities across diverse experimental conditions. Conclusion This study demonstrates the significant potential of synthetic-data pretraining to enhance the accuracy and robustness of deep neural networks for partial discharge (PD) detection in medium-voltage overhead transmission power lines. By carefully crafting simulation pipelines that incorporate realistic PD rates, amplitude variability, vegetation contacts, reflection artifacts, and wideband electromagnetic noise, we created multiple synthetic datasets with varying degrees of fidelity. Our empirical analysis reveals that pretraining on a well-balanced and representative synthetic dataset yields statistically significant gains in MCC, F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document} score, and accuracy upon subsequent fine-tuning with real-world measurements. In particular, the dataset with higher PD repetition rates and broader amplitude distributions produced the most consistent improvements, underscoring the need to mirror actual overhead-line conditions as closely as possible. We also observed that pretraining on inadequately parameterized synthetic data can be detrimental, reinforcing the necessity for domain alignment and rigorous simulation design. From a practical standpoint, these findings highlight a path toward reducing field data requirements, lowering operational expenses, and strengthening PD detection pipelines for proactive grid maintenance. Our comparative analysis unequivocally shows that CNN-based architectures are significantly more effective for this task than LSTM or Vision Transformer models when using spectrogram inputs, highlighting the primacy of spatial feature extraction for identifying PD time-frequency signatures. Among the synthetic data configurations, those simulating more intense PD activity—characterized by higher repetition rates and amplitudes—consistently produced the best results, better preparing models for real-world conditions. We also found that top-performing models are robust to varying levels of synthetic noise, suggesting that fidelity in modeling the PD signal itself is paramount. Future work should focus on extending the proposed framework in two key directions. First, sophisticated domain adaptation techniques, such as domain-adversarial fine-tuning or contrastive mapping, should be investigated to more explicitly align the feature spaces of synthetic and real signals, especially under evolving environmental or loading conditions. Second, developing enhanced generative models capable of simulating extreme events, such as severe weather patterns, augmented vegetation encroachment, or heavy rime formation—would bolster real-world readiness. Addressing these avenues will ultimately facilitate more reliable and scalable PD monitoring solutions, helping utilities and operators safeguard overhead transmission networks against unexpected faults and service interruptions.