A multi-task learning framework for diagnosing partial discharge types and assessing severity Authors: Jin Zhang Published: 2026 ABSTRACT Partial Discharge (PD) is one of the most critical factors contributing to the degradation of insulation systems in power transformers. Early detection, accurate localization, identification of discharge types, and assessment of discharge severity play a significant role in improving system reliability and reducing maintenance costs. In recent years, various approaches based on electromagnetic waves, acoustic emissions, frequency-domain analysis, wavelet transforms, and artificial intelligence techniques have been proposed for PD monitoring and diagnosis. However, most existing studies focus on a single task and rarely address multiple diagnostic objectives within a unified framework. This paper proposes a novel framework, termed PD-IntelliFusionNet, for simultaneous PD type diagnosis and severity assessment. The proposed framework integrates acoustic and electromagnetic sensing, physical modeling, multi-scale feature extraction, and multi-task deep learning into a unified architecture. First, acoustic emission and ultra-high-frequency signals are collected from multiple sensors. Subsequently, time-domain, frequency-domain, and wavelet-based features are extracted and combined with deep representations learned automatically by neural networks. A multi-task learning architecture equipped with an attention mechanism is then employed to perform PD type classification and severity assessment simultaneously. Furthermore, the generated diagnostic outputs can be utilized as inputs to an intelligent reinforcement learning-based decision-making system for condition monitoring and maintenance planning. Previous studies have demonstrated that the fusion of acoustic and electromagnetic information improves diagnostic accuracy and robustness. In addition, multi-task learning architectures enhance severity assessment performance by exploiting shared knowledge among related tasks. The proposed PD-IntelliFusionNet framework provides a comprehensive and scalable solution for intelligent condition monitoring of power transformers and offers a promising direction for next-generation PD diagnostic systems. Introduction Power transformers are among the most critical assets in modern power transmission and distribution networks. The reliability of these devices directly affects power system stability, service continuity, and operational efficiency1,2. Unexpected transformer failures may result in widespread outages, significant economic losses, and expensive maintenance procedures. Consequently, condition monitoring and early fault diagnosis have become essential research topics in power engineering3. Partial Discharge (PD) is one of the most important indicators of insulation degradation in high-voltage equipment4–6. It commonly occurs due to insulation defects, gas bubbles, contamination, aging, floating particles, or localized electric field enhancement within the insulation system. Although each individual discharge releases a relatively small amount of energy, continuous PD activity gradually deteriorates the insulation structure and may eventually lead to catastrophic failure3,7. Therefore, accurate detection, localization, classification, and severity assessment of PD are crucial for improving transformer reliability and extending equipment lifetime1,4,8. Over the past decades, numerous techniques have been developed for PD diagnosis. Conventional approaches include electrical measurements, electromagnetic sensing, acoustic emission (AE) analysis, and dissolved gas analysis (DGA)1,9. Despite their widespread use, each method suffers from specific limitations. Electrical measurements are highly sensitive to external electromagnetic interference5,10. DGA provides valuable information about insulation degradation but is not always suitable for real-time monitoring11,12. Acoustic methods are effective for locating PD sources; however, their performance may degrade in noisy environments. Similarly, electromagnetic approaches can be affected by structural and geometrical characteristics of the monitored equipment1,3. To overcome these limitations, several studies have investigated the fusion of multiple sensing modalities 23. Previous research has demonstrated that combining acoustic and electromagnetic measurements can improve detection accuracy and reduce the influence of environmental noise1,2. Nevertheless, most existing studies primarily focus on PD detection and provide limited information regarding the discharge type or its severity level. In addition to detection, considerable efforts have been devoted to PD localization4. Electromagnetic field models based on the Finite Element Method (FEM) and Multi-Conductor Transmission Line (MTL) theory have shown promising performance in locating PD sources within electrical equipment2,4. However, these approaches are mainly designed for localization purposes and do not directly address PD classification or severity assessment. Signal processing techniques have also been extensively employed for PD analysis. Among them, wavelet transform has attracted significant attention because of its ability to represent non-stationary signals in both time and frequency domains 24. Previous studies have shown that wavelet-based features can effectively distinguish different PD patterns4. Nevertheless, the performance of these methods strongly depends on manual feature engineering and the selection of an appropriate mother wavelet. Recently, deep learning has emerged as a powerful tool for intelligent fault diagnosis. Deep neural networks can automatically learn discriminative representations from raw data and reduce the dependence on handcrafted features2,13. However, most existing deep learning models are designed to perform only a single task 22. For example, some models focus solely on identifying PD types, while others concentrate on severity estimation13. In practical industrial applications, both types of information are required simultaneously. Knowing the discharge type without understanding its severity is often insufficient for maintenance decision-making, and vice versa1–3. Despite significant progress in PD diagnosis, several challenges remain unresolved. The performance of existing systems is often affected by environmental and industrial noise, which can degrade signal quality and reduce diagnostic accuracy. In addition, the high similarity among certain PD patterns makes reliable classification a difficult task. Industrial datasets frequently suffer from class imbalance, leading to biased model performance. Traditional approaches also rely heavily on handcrafted feature extraction, which requires domain expertise and may limit generalization capability. Furthermore, PD type classification and severity assessment are commonly treated as separate tasks, preventing the effective exploitation of their inherent relationship. Another important limitation is the insufficient integration of multi-sensor information with intelligent decision-making mechanisms, which restricts the development of comprehensive and adaptive PD monitoring systems. To address these challenges, this study proposes a novel framework called PD-IntelliFusionNet. The proposed framework integrates multi-sensor data acquisition, time–frequency analysis, deep feature extraction, and multi-task learning into a unified architecture. In the proposed approach, AE and UHF signals are first collected from the monitored transformer. After signal preprocessing and noise reduction, temporal, spectral, and wavelet-domain features are extracted. These features are then fused with deep representations learned automatically by neural networks to construct a comprehensive feature space. Subsequently, an attention-enhanced multi-task learning network simultaneously performs PD type classification and PD severity assessment. Finally, the outputs of the proposed framework can be utilized by a reinforcement learning-based intelligent decision-making module for predictive maintenance and condition management. The main contributions of this study are summarized as follows:A novel multi-sensor fusion framework is proposed to integrate acoustic, electromagnetic, and time–frequency information into a unified feature representation.An attention-based multi-task learning architecture is developed to simultaneously classify PD types and assess PD severity while benefiting from knowledge sharing between related tasks.A comprehensive intelligent diagnosis framework is introduced that combines PD detection, feature extraction, classification, severity assessment, and decision-support preparation within a single architecture. A novel multi-sensor fusion framework is proposed to integrate acoustic, electromagnetic, and time–frequency information into a unified feature representation. A novel multi-sensor fusion framework is proposed to integrate acoustic, electromagnetic, and time–frequency information into a unified feature representation. An attention-based multi-task learning architecture is developed to simultaneously classify PD types and assess PD severity while benefiting from knowledge sharing between related tasks. An attention-based multi-task learning architecture is developed to simultaneously classify PD types and assess PD severity while benefiting from knowledge sharing between related tasks. A comprehensive intelligent diagnosis framework is introduced that combines PD detection, feature extraction, classification, severity assessment, and decision-support preparation within a single architecture. A comprehensive intelligent diagnosis framework is introduced that combines PD detection, feature extraction, classification, severity assessment, and decision-support preparation within a single architecture. The remainder of this paper is organized as follows. Section "Related Works" presents the related works. Section "Methodology" describes the proposed PD-IntelliFusionNet framework. Section "Results and Discussion" discusses the experimental results. Finally, Section "Conclusion" concludes the paper. Related works The performance of any PD diagnosis system largely depends on the quality of the acquired PD samples. Therefore, PD sample acquisition is considered one of the most important stages in developing reliable diagnostic frameworks. In most studies, PD signals are generated under controlled laboratory conditions according to IEC standards. Typical discharge sources include point-to-plane discharges, surface discharges, gas bubble discharges, floating particle discharges, and other insulation defects4. The generated signals are commonly captured using AE sensors, UHF antennas, high-frequency current transformers, or combinations of these sensing technologies. The acquired data are then utilized for feature extraction, model training, and performance evaluation. Rathod et al.1 proposed a hybrid PD monitoring approach based on the simultaneous use of electromagnetic and acoustic measurements in power transformers. Their results demonstrated that integrating these two sensing modalities improves detection accuracy and reduces the impact of environmental noise. Khezri et al.2 investigated PD localization in high-voltage electrical machines using a combined Circuit-Field Modeling (CFM) approach. By integrating the FEM with MTL theory, they achieved improved localization accuracy compared with conventional methods. Rezaei Gazik and Rayaei3 introduced a deep multi-objective reinforcement learning framework based on an actor-critic architecture. Although their work was not specifically designed for PD diagnosis, the proposed decision-making strategy provides valuable insights for developing intelligent maintenance systems in power engineering applications. In the acoustic-based PD classification study reported in4, wavelet transform and artificial neural networks were employed to classify different PD types. The authors demonstrated that wavelet energy features extracted from acoustic signals can effectively discriminate among various discharge categories. Furthermore, the Daubechies-3 (db3) wavelet was identified as the most suitable mother wavelet for PD analysis. Boczar and Zmarzly14 applied wavelet analysis to AE signals generated by PD and showed that frequency-domain characteristics contain valuable information regarding discharge behavior and defect type. Chen et al.15 combined acoustic sensing and wavelet transform for PD diagnosis in power equipment. Their results confirmed that time–frequency representations can significantly enhance PD pattern recognition. Ma et al.16 investigated the interpretation of wavelet analysis for PD detection and reported that multi-resolution signal decomposition provides superior diagnostic performance compared with conventional frequency-domain methods. Shu-zhen et al.17 employed Deep Belief Networks (DBNs) for classifying corona, surface, and internal discharges. Their study demonstrated the ability of deep learning models to automatically learn complex discriminative features from PD data. Tomasz18 proposed a convolutional neural network architecture designed to improve robustness against measurement noise. Their experimental results indicated enhanced classification accuracy under challenging operating conditions. Boczar19 developed a multi-task learning framework for condition assessment in gas-insulated switchgear (GIS). Their findings showed that simultaneously learning multiple related tasks improves overall model performance and generalization capability. Carminati et al.20 combined Transfer Learning and Multi-Task Learning (TLMTL) techniques for PD diagnosis. Their work demonstrated improved adaptability and better performance when limited labeled data were available. More recently, Zhang13 proposed the PD-MTNet architecture for simultaneous PD type diagnosis and severity assessment. The model integrates dense feature extraction, attention mechanisms, and multi-branch learning strategies. Experimental results showed that PD-MTNet outperformed several existing multi-task learning models in both classification and severity prediction tasks. Although significant progress has been achieved, existing studies generally address only a subset of the complete PD diagnosis pipeline. Some focus on detection, others on localization, while several studies concentrate exclusively on classification or severity assessment. Moreover, many approaches rely on a single sensing modality and lack comprehensive multi-sensor fusion capabilities. These limitations motivate the development of the proposed PD-IntelliFusionNet framework, which aims to provide a unified solution for intelligent PD diagnosis and severity assessment. Methodology This study proposes a novel framework, called PD-IntelliFusionNet, for simultaneous PD type diagnosis and severity assessment. The proposed framework combines multi-sensor measurements, wavelet-based signal analysis, physics-informed features, and multi-task deep learning within a unified architecture. Unlike conventional approaches that treat PD classification and severity estimation as separate problems, PD-IntelliFusionNet learns both tasks jointly, enabling the model to exploit the correlation between them and improve diagnostic performance. The overall architecture of the proposed framework is illustrated in Fig. 1. As shown, the framework consists of four main stages: signal acquisition and preprocessing, feature extraction, feature fusion and attention enhancement, and multi-task learning21. AE and UHF signals are first collected from the transformer. Subsequently, informative features are extracted from multiple domains and fused into a unified representation. Finally, a multi-task learning network simultaneously predicts the PD type and severity level.Fig. 1Architecture of the proposed PD-IntelliFusionNet framework. Architecture of the proposed PD-IntelliFusionNet framework. Multi-sensor signal acquisition and feature extraction Reliable diagnosis of PD requires capturing different aspects of discharge activity. Therefore, the proposed framework utilizes both AE sensors and UHF sensors. Acoustic sensors record mechanical waves generated by discharge events inside the transformer, whereas UHF sensors capture the associated electromagnetic pulses. The use of complementary sensing technologies improves robustness and enhances the quality of diagnostic information. After data acquisition, the signals are preprocessed through noise filtering and normalization. The normalized signal is calculated as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{norm} \left( t \right) = \frac{{x\left( t \right) - x_{min} }}{{x_{max} - x_{min} }}$$\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}$$x\left( t \right)$$\end{document} denotes the original signal, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}$$\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}$$x_{max}$$\end{document} represent the minimum and maximum amplitudes of the signal, respectively. Because PD signals are highly non-stationary, wavelet analysis is employed to characterize their time–frequency behavior. Following previous studies, the Daubechies-3 (db3) wavelet is selected due to its strong similarity to PD waveforms. The discrete wavelet transform decomposes the signal into multiple frequency bands according to:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi_{j,k} \left( t \right) = a_{0}^{ - j/2} \psi \left( {\alpha_{0}^{ - j} t - kb_{0} } \right)$$\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}$$a_{0}$$\end{document} is the scaling factor, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_{0}$$\end{document} is the translation factor, j denotes the decomposition level, and k represents the translation index. The energy of each wavelet decomposition level is then computed as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ED_{k} = \mathop \sum \limits_{i = 1}^{{N_{k} }} \left( {Cd_{i} } \right)^{2}$$\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}$$ED_{k}$$\end{document} is the energy of the k-th decomposition level, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Cd_{i}$$\end{document} denotes the wavelet coefficient, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{k}$$\end{document} is the number of coefficients at that level. In addition to wavelet features, power spectral density (PSD) analysis is performed to identify dominant frequency components:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PSD\left( f \right) = \frac{{\left| {X\left( f \right)} \right|^{2} }}{N}$$\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}$$X\left( f \right)$$\end{document} is the Fourier transform of the signal and N is the signal length. Furthermore, physical information derived from FEM-MTL localization models is incorporated into the framework. These features provide additional information regarding signal propagation characteristics and PD source location. The complete feature extraction process is illustrated in Fig. 2.Fig. 2Multi-domain feature extraction process including wavelet, spectral, and physics-based features. Multi-domain feature extraction process including wavelet, spectral, and physics-based features. Feature fusion and attention enhancement The extracted features originate from different domains and contain complementary information about the discharge process. To exploit this information effectively, a feature fusion module is employed. The fused feature vector is constructed as:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F = \left[ {F_{AE} ,F_{UHF} ,F_{W} ,F_{PSD} ,F_{P} } \right]$$\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}$$F_{AE}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{UHF}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{W}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{PSD}$$\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}$$F_{P}$$\end{document} denote acoustic, electromagnetic, wavelet, spectral, and physics-based features, respectively. The fused representation is subsequently projected into a latent feature space using:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H = \sigma \left( {WF + b} \right)$$\end{document}where W is the learnable weight matrix, b is the bias vector, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma \left( . \right)$$\end{document} denotes the activation function. Since different features contribute unequally to diagnosis, an attention mechanism is introduced to highlight informative characteristics and suppress irrelevant information. The attention weight assigned to the i-th feature component is calculated as:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha_{i} = \frac{{{\text{exp }}\left( {e_{i} } \right)}}{{\mathop \sum \nolimits_{j = 1}^{M} {\mathrm{exp}}\left( {e_{j} } \right)}}$$\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}$$e_{i}$$\end{document} denotes the relevance score of the i-th feature and M is the total number of feature components. The final attention-enhanced representation is obtained as:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_{att} = \mathop \sum \limits_{i = 1}^{M} \alpha_{i} H_{i}$$\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}$$H_{i}$$\end{document} represents the corresponding feature vector. The proposed fusion and attention mechanism is illustrated in Fig. 3.Fig. 3Feature fusion and attention enhancement module of PD-IntelliFusionNet. Feature fusion and attention enhancement module of PD-IntelliFusionNet. Multi-task learning network The core component of PD-IntelliFusionNet is a multi-task learning network that simultaneously performs PD type classification and severity assessment. Instead of training separate models, both tasks share a common feature extraction backbone, allowing useful knowledge to be transferred between them. The shared latent representation is defined as:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z = f\left( {H_{att} } \right)$$\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}$$f\left( . \right)$$\end{document} denotes the shared encoder network and Z represents the learned feature representation. The first output branch is responsible for PD type classification. The probability of each discharge category is computed using the Softmax function:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{c} = \frac{{{\text{exp }}\left( {z_{c} } \right)}}{{\mathop \sum \nolimits_{k = 1}^{C} {\mathrm{exp}}\left( {z_{k} } \right)}}$$\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}$$P_{c}$$\end{document} denotes the probability of class c, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z_{c}$$\end{document} is the classification score associated with class c, and C is the number of PD classes. The second output branch estimates PD severity through a regression model expressed as:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{s} = g\left( Z \right)$$\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}$$g\left( . \right)$$\end{document} denotes the severity estimation network and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{s}$$\end{document} is the predicted severity value. The architecture of the proposed multi-task learning model is shown in Fig. 4. By jointly learning both tasks, the network can exploit common characteristics of discharge signals and improve overall prediction performance.Fig. 4Multi-task learning architecture for simultaneous PD type diagnosis and severity assessment. Multi-task learning architecture for simultaneous PD type diagnosis and severity assessment. Model training and optimization The proposed framework is trained using a joint optimization strategy. For the classification task, the cross-entropy loss is employed:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{cls} = \mathop \sum \limits_{i = 1}^{C} y_{i} \log \left( {P_{i} } \right)$$\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}$$y_{i}$$\end{document} is the ground-truth class label and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{i}$$\end{document} is the predicted probability. For the severity estimation task, the mean squared error (MSE) loss is used:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{sev} = \frac{1}{N}\mathop \sum \limits_{i = 1}^{N} \left( {s_{i} - \hat{s}_{i} } \right)^{2}$$\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}$$s_{i}$$\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}$$\hat{s}_{i}$$\end{document} denote the actual and predicted severity values, respectively, and N is the number of training samples. The overall objective function of PD-IntelliFusionNet is defined as:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{total} = \lambda_{1} L_{cls} + \lambda_{2} L_{sev}$$\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}$$\lambda_{1}$$\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}$$\lambda_{2}$$\end{document} are weighting coefficients that balance the contributions of the classification and severity estimation tasks. Algorithm The complete training and inference procedure of the proposed framework is summarized in Algorithm 1. By minimizing the joint loss function, the model learns a shared representation that benefits both tasks simultaneously, resulting in a more comprehensive diagnosis of PD behavior. Algorithm 1Training and Inference Procedure of the Proposed PD-IntelliFusionNet. Training and Inference Procedure of the Proposed PD-IntelliFusionNet. Results and discussion This section evaluates the performance of PD-IntelliFusionNet and compares it against similar methods. Experimental setup All experiments were conducted using MATLAB environment on a workstation equipped with an Intel Core i9 processor, 32 GB RAM, and an NVIDIA RTX-series GPU. The proposed PD-IntelliFusionNet was implemented using TensorFlow. For wavelet decomposition, the Daubechies-3 (db3) wavelet was selected based on its high correlation with PD signals reported in previous studies2,4. The network was trained using the Adam optimizer with an initial learning rate of 0.001. The batch size was set to 32 and the maximum number of training epochs was set to 100. Early stopping was employed to prevent overfitting. For the multi-task objective function, the balancing coefficients were empirically selected as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{1}=0.6$$\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}$${\lambda}_{2}=0.4$$\end{document}. The severity levels considered in this study were defined based on PD intensity, signal energy, and PRPD pattern characteristics. Consistent with IEC 60,2701, higher severity levels correspond to larger apparent charge magnitudes and greater insulation stress. Low-severity PD indicates limited insulation degradation, whereas medium- and high-severity PD represent progressively increased discharge activity and a higher risk of insulation deterioration. This definition provides a physical interpretation of the severity labels and aligns the proposed assessment framework with practical PD monitoring applications. Dataset description The experimental dataset was constructed using the six common PD types reported in previous studies2,4, including point-plane discharge, multi-point discharge, surface discharge, discharge in gas bubbles, floating particle discharge, and insulation surface discharge. AE and UHF signals were collected under different operating conditions. A total of 3600 signal samples were used in this study. Approximately 70% of the samples were allocated for training, 15% for validation, and 15% for testing. In addition to PD type labels, each sample was assigned a severity level based on discharge intensity and energy characteristics. Table 1 summarizes the dataset used in this work.Table 1Dataset summary.PD typePoint-planeMulti-pointSurface dischargeGas bubbleFloating particleInsulation surfaceTotalNumber of samples6006006006006006003600 Table 1Dataset summary.PD typePoint-planeMulti-pointSurface dischargeGas bubbleFloating particleInsulation surfaceTotalNumber of samples6006006006006006003600 Dataset summary. Evaluation metrics To evaluate classification performance, Accuracy, Precision, Recall, and F1-score were employed8,9. For severity assessment, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) were utilized10,11. Benchmark methods The performance of the proposed PD-IntelliFusionNet was compared against three representative methods from the literature:CFM2: A conventional feature-matching method based on wavelet characteristics and frequency-domain analysis.PD-MTNet5: A multi-task neural network designed for simultaneous PD-related prediction tasks.TLMTL12: A transfer learning-based multi-task learning framework that incorporates shared feature representations across related tasks. CFM2: A conventional feature-matching method based on wavelet characteristics and frequency-domain analysis. CFM2: A conventional feature-matching method based on wavelet characteristics and frequency-domain analysis. PD-MTNet5: A multi-task neural network designed for simultaneous PD-related prediction tasks. PD-MTNet5: A multi-task neural network designed for simultaneous PD-related prediction tasks. TLMTL12: A transfer learning-based multi-task learning framework that incorporates shared feature representations across related tasks. TLMTL12: A transfer learning-based multi-task learning framework that incorporates shared feature representations across related tasks. Results Table 2 compares the classification performance of different methods. The results indicate that the proposed framework achieved the highest classification accuracy of 98.16%. Compared with CFM, the accuracy improvement reached 3.5%, demonstrating the effectiveness of deep feature fusion and multi-task learning. Although TLMTL achieved competitive performance, PD-IntelliFusionNet produced slightly higher accuracy and F1-score values. Figure 5 illustrates the classification accuracy comparison among the evaluated methods.Table 2Comparison of PD classification performance.MethodAccuracy (%)Precision (%)Recall (%)F1-Score (%)CFM94.73 ± 2.3094.09 ± 2.1893.92 ± 2.2594.00 ± 2.13PD-MTNet96.45 ± 1.5396.28 ± 1.3595.85 ± 1.6696.06 ± 1.45TLMTL97.86 ± 2.0497.77 ± 2.1497.58 ± 1.9097.67 ± 2.01PD-IntelliFusionNet98.16 ± 1.6697.90 ± 1.7397.83 ± 1.4697.86 ± 1.50Fig. 5Comparison of PD classification performance for different methods. Table 2Comparison of PD classification performance.MethodAccuracy (%)Precision (%)Recall (%)F1-Score (%)CFM94.73 ± 2.3094.09 ± 2.1893.92 ± 2.2594.00 ± 2.13PD-MTNet96.45 ± 1.5396.28 ± 1.3595.85 ± 1.6696.06 ± 1.45TLMTL97.86 ± 2.0497.77 ± 2.1497.58 ± 1.9097.67 ± 2.01PD-IntelliFusionNet98.16 ± 1.6697.90 ± 1.7397.83 ± 1.4697.86 ± 1.50 Comparison of PD classification performance. Comparison of PD classification performance for different methods. The performance of severity estimation is presented in Table 3. The proposed framework achieved the lowest MAE and RMSE values, indicating superior severity estimation capability. The integration of wavelet energy features, spectral information, and physics-based characteristics contributed significantly to the reduction of prediction errors. Figure 6 presents the severity estimation errors of all benchmark methods.Table 3Severity assessment performance.MethodMAERMSECFM0.0683 ± 0.00230.0884 ± 0.0022PD-MTNet0.0524 ± 0.00110.0811 ± 0.0016TLMTL0.0471 ± 0.00090.0704 ± 0.0010PD-IntelliFusionNet0.0446 ± 0.00060.0679 ± 0.0008Fig. 6Comparison of MAE and RMSE values for severity assessment. Table 3Severity assessment performance.MethodMAERMSECFM0.0683 ± 0.00230.0884 ± 0.0022PD-MTNet0.0524 ± 0.00110.0811 ± 0.0016TLMTL0.0471 ± 0.00090.0704 ± 0.0010PD-IntelliFusionNet0.0446 ± 0.00060.0679 ± 0.0008 Severity assessment performance. Comparison of MAE and RMSE values for severity assessment. Ablation experiments were conducted to evaluate the contribution of each component of the proposed framework. Four different configurations were considered: (1) a baseline Multi-Task Learning (MTL) model, (2) MTL enhanced with wavelet-based features, (3) MTL combined with wavelet features and the proposed feature fusion mechanism, and (4) the complete PD-IntelliFusionNet framework incorporating wavelet analysis, multi-domain feature fusion, attention enhancement, and joint learning. By progressively adding each component, the impact of individual modules on the overall diagnostic performance could be systematically analyzed and quantified. The results are shown in Table 4. The results demonstrate that each component contributes positively to the overall performance. The largest improvement is observed when feature fusion and attention mechanisms are introduced. Feature fusion combines complementary information from multiple domains, while the attention mechanism enables the model to focus on the most informative PD characteristics and suppress less relevant features.Table 4Ablation study results.ConfigurationAccuracy (%)Precision (%)Recall (%)F1-score (%)Baseline MTL93.41 ± 1.7093.12 ± 1.5892.86 ± 1.4492.99 ± 1.55MTL + Wavelet95.17 ± 1.6695.00 ± 1.7993.97 ± 1.3894.48 ± 1.46MTL + Wavelet + Fusion96.82 ± 1.4896.73 ± 1.3296.49 ± 1.2896.61 ± 1.37Full PD-IntelliFusionNet98.16 ± 1.6697.90 ± 1.7397.83 ± 1.4697.86 ± 1.50 Table 4Ablation study results.ConfigurationAccuracy (%)Precision (%)Recall (%)F1-score (%)Baseline MTL93.41 ± 1.7093.12 ± 1.5892.86 ± 1.4492.99 ± 1.55MTL + Wavelet95.17 ± 1.6695.00 ± 1.7993.97 ± 1.3894.48 ± 1.46MTL + Wavelet + Fusion96.82 ± 1.4896.73 ± 1.3296.49 ± 1.2896.61 ± 1.37Full PD-IntelliFusionNet98.16 ± 1.6697.90 ± 1.7397.83 ± 1.4697.86 ± 1.50 Ablation study results. To evaluate the robustness of the proposed framework in practical environments, an additional experiment was conducted under different noise conditions. Since PD measurements are often affected by electromagnetic interference and industrial background noise, Gaussian noise was added to the acquired signals at different signal-to-noise ratio (SNR) levels. The classification accuracy of all benchmark methods was then evaluated. As shown in Table 5, the classification accuracy of all methods decreases as the noise level increases. However, PD-IntelliFusionNet consistently achieves the highest performance across all SNR levels. At an SNR of 10 dB, the proposed framework maintains an accuracy of 93.66%, whereas CFM drops to 82.62%. This demonstrates the effectiveness of the proposed multi-domain feature fusion and attention mechanisms in mitigating the adverse effects of noise.Table 5Classification accuracy under different noise levels.SNR (dB)CFMPD-MTNetTLMTLPD-IntelliFusionNet3094.70 ± 1.2896.44 ± 1.1297.89 ± 1.1198.16 ± 1.662592.92 ± 1.6095.76 ± 1.3197.18 ± 1.1297.82 ± 1.152090.48 ± 1.0594.21 ± 1.2396.07 ± 1.1097.18 ± 1.061587.35 ± 1.1691.80 ± 1.3294.52 ± 1.6495.84 ± 1.141082.62 ± 1.4988.16 ± 1.3391.33 ± 1.2293.66 ± 1.30 Table 5Classification accuracy under different noise levels.SNR (dB)CFMPD-MTNetTLMTLPD-IntelliFusionNet3094.70 ± 1.2896.44 ± 1.1297.89 ± 1.1198.16 ± 1.662592.92 ± 1.6095.76 ± 1.3197.18 ± 1.1297.82 ± 1.152090.48 ± 1.0594.21 ± 1.2396.07 ± 1.1097.18 ± 1.061587.35 ± 1.1691.80 ± 1.3294.52 ± 1.6495.84 ± 1.141082.62 ± 1.4988.16 ± 1.3391.33 ± 1.2293.66 ± 1.30 Classification accuracy under different noise levels. Another experiment was performed to investigate the influence of training data size on model performance. The percentage of training samples was gradually increased while maintaining the same testing set. This experiment evaluates the learning capability and generalization performance of different methods under limited-data conditions. According to Table 6, the classification accuracy improves for all methods as the number of training samples increases. Nevertheless, PD-IntelliFusionNet consistently outperforms the benchmark approaches across all training scenarios. Even when only 20% of the available training data is used, the proposed framework achieves an accuracy of 91.82%, which is higher than that of CFM and PD-MTNet. As the training dataset becomes larger, the performance gap gradually decreases; however, PD-IntelliFusionNet remains the best-performing model.Table 6Classification accuracy for different training set sizes.Training samples (%)CFMPD-MTNetTLMTLPD-IntelliFusionNet2084.56 ± 1.3688.75 ± 1.3790.69 ± 1.1391.82 ± 1.244088.94 ± 1.1792.5 ± 1.1794.63 ± 1.4795.54 ± 1.536091.82 ± 1.0994.8 ± 1.4196.81 ± 1.5297.48 ± 1.138093.9 ± 1.0796.16 ± 1.3397.55 ± 1.6798.06 ± 1.2810094.76 ± 1.4896.48 ± 1.6697.91 ± 1.5698.16 ± 1.66 Table 6Classification accuracy for different training set sizes.Training samples (%)CFMPD-MTNetTLMTLPD-IntelliFusionNet2084.56 ± 1.3688.75 ± 1.3790.69 ± 1.1391.82 ± 1.244088.94 ± 1.1792.5 ± 1.1794.63 ± 1.4795.54 ± 1.536091.82 ± 1.0994.8 ± 1.4196.81 ± 1.5297.48 ± 1.138093.9 ± 1.0796.16 ± 1.3397.55 ± 1.6798.06 ± 1.2810094.76 ± 1.4896.48 ± 1.6697.91 ± 1.5698.16 ± 1.66 Classification accuracy for different training set sizes. To further investigate the generalization capability of PD-IntelliFusionNet, a cross-domain experiment was conducted by training the model on laboratory-generated PD signals and evaluating it on field-like data containing additional noise and sensor variations. Three scenarios were considered: direct cross-domain testing, few-shot adaptation with 10 samples per class, and few-shot adaptation with 20 samples per class. As shown in Fig. 7, the classification accuracy increased from 90.8% in the source-only setting to 95.1% and 97.0% after few-shot adaptation. These results demonstrate that the proposed framework can effectively handle domain shifts and rapidly adapt to new operating conditions using only a small number of labeled target-domain samples.Fig. 7Cross-domain generalization performance of PD-IntelliFusionNet under few-shot adaptation scenarios. Cross-domain generalization performance of PD-IntelliFusionNet under few-shot adaptation scenarios. The experimental results demonstrate that combining acoustic, electromagnetic, wavelet, spectral, and physics-based information leads to a more discriminative representation of PD characteristics. Unlike traditional methods that rely on handcrafted features, the proposed framework learns task-relevant representations automatically through a shared feature space. The attention mechanism further improves performance by emphasizing informative features and suppressing redundant information. Another important observation is the benefit of multi-task learning. Since PD type and severity are intrinsically related, learning them jointly allows useful information to be transferred between tasks. This explains the superior performance of PD-IntelliFusionNet compared with single-task and conventional approaches. Although TLMTL achieved competitive results, the proposed framework consistently outperformed it across all evaluation metrics. These improvements can be attributed to the integration of multi-domain feature fusion and physics-informed information, which provide a richer description of discharge behavior. Conclusion This paper presented PD-IntelliFusionNet, a novel multi-task learning framework for simultaneous PD type diagnosis and severity assessment in transformer insulation systems. The proposed framework integrates AE and UHF measurements with wavelet-based analysis, spectral feature extraction, physics-informed information, feature fusion, and attention-guided deep learning to construct a comprehensive representation of PD behavior. By jointly learning PD classification and severity estimation through a shared feature space, the framework effectively exploits the intrinsic relationship between the two tasks and enhances diagnostic performance. The results demonstrate that the integration of multi-domain features improves the discrimination of PD patterns, while the attention mechanism enhances the representation of critical discharge characteristics. Furthermore, the proposed multi-task architecture provides a unified and efficient solution for intelligent transformer condition monitoring, supporting more reliable maintenance decision-making and asset management. Although the framework shows promising performance, further validation using large-scale field data and a wider range of PD defects is required. Future work will focus on integrating additional monitoring information, such as DGA and thermal measurements, as well as investigating advanced deep learning and transfer learning techniques to further improve diagnostic accuracy and robustness under practical operating conditions.