Physics-guided resampling and asymmetric fusion for ultra-high-frequency bolt-loosening detection Authors: Zihao Luo, Xinrui Shi, Kai Xie, Chang Wen, Sixiang Liu, Wei Zhang Published: 2026 ABSTRACT AbstractReliable detection of bolt loosening in safety-critical infrastructure requires monitoring that captures microsecond-scale transients. However, processing 1 MHz vibration signals at the edge presents a fundamental dilemma: standard downsampling can smear sparse diagnostic impulses, whereas full-bandwidth processing is often computationally prohibitive. Here we propose PGRF-Net, a physics-guided deep learning framework that reconciles transient preservation with extreme compression. PGRF-Net integrates (1) a physics-guided resampling operator for high-rate vibration signals (e.g., 150 : 1 compression); (2) a heterogeneous gradient–spectral tensor representation that augments time–frequency information with morphology-aware channels; and (3) an asymmetric fusion module over two representations derived from the same signal (pseudo-image representation + waveform representation). On a six-class 1 MHz PVDF vibration dataset (109,668 segments; temporal-split test = 16,131), PGRF-Net reaches a best-run clean test accuracy of 95.12% under a block-wise temporal split. A strict file-disjoint split is also reported as a cross-scene hold-out protocol. On an independent Zenodo benchmark (3-fold file-disjoint CV, 9 runs per experiment), we evaluate the proposed pipeline together with three feature-engineering controls (Exp 501–503). These results support a practical compression–learning pipeline for industrial monitoring where both transient fidelity and computational efficiency are required.Subject terms: Engineering, Mathematics and computing Introduction Bolted connections are ubiquitous in engineering structures and mechanical assemblies because they are convenient, serviceable, and load-bearing. However, bolt loosening under dynamic loading and environmental influences remains a major cause of degradation and failure in bridges, wind-turbine towers, pipelines, and manufacturing equipment1. The mechanics of self-loosening—a coupled process involving preload loss, frictional slip, and dynamic excitation—have been studied extensively, showing a progressive reduction in clamping force and eventual loss of joint integrity1. Reliable and timely detection of bolt loosening is therefore essential for structural health monitoring (SHM) and condition-based maintenance. A broad spectrum of detection techniques have been proposed. Reviews of bolted-connection monitoring commonly group approaches into three families2: (1) physics-based inspection using ultrasonic, guided-wave, acoustic, or electromechanical impedance methods3–5; (2) vibration-based modal and signal-processing techniques that examine shifts in natural frequency, damping, or energy distribution6,7; and (3) vision-based and data-driven methods that exploit images or learned features for automated assessment8–10. Each family offers distinct advantages: ultrasonic and guided-wave methods provide physically interpretable metrics for clamping force or contact state3,4, whereas vibration-based methods can be implemented with modest sensor infrastructure and are sensitive to contact nonlinearities11,12. In field deployments, however, performance can degrade due to coupling variability, environmental noise, and the subtlety of early-stage loosening signatures. Vision and image-based solutions have attracted increasing interest because they can non-intrusively capture geometric and surface cues associated with loosening (e.g., head rotation, gap formation) and can be scaled using modern imaging hardware8,13. Several studies demonstrate effective bolt rotation detection and localization using classical image processing and machine learning pipelines8,13, and more recent work integrates deep convolutional neural networks and graphical models to improve robustness under viewpoint and illumination changes10,14,15. Despite these advances, purely visual methods are inherently limited in detecting micro-loosening or internal preload loss when no apparent surface change occurs; therefore, they are best suited as part of a multimodal inspection strategy. Ultrasonic and piezoelectric techniques remain indispensable for quantitative assessment of bolt preload and contact state. Combined longitudinal/transverse ultrasonic probing, time-of-flight, and velocity estimation have been used to infer clamping force and detect loosening with high sensitivity under controlled conditions3,4. Piezo-impedance and actuator–sensor schemes enable local condition monitoring and residual-life estimation by tracking electromechanical coupling variations5,16,17. Vibro-acoustic modulation and time-reversal approaches further enhance the detection of weak nonlinear contact features associated with early looseness17,18. Nevertheless, these methods typically require careful sensor coupling and calibration, and their performance can degrade in complex, noisy operational environments. Vibration-signal analysis and data-driven models address these limitations by leveraging temporal patterns and statistical signatures in measured responses. Time–frequency representations (e.g., STFT, wavelets) and energy-concentration measures provide compact descriptors for transient and cyclostationary phenomena associated with loosened joints11,12. Sparse factorization and tensor methods can emphasize impulsive or localized spectral content relevant to contact faults19. Recent tool-condition and milling-fault studies further strengthen this direction, including degradation-state modeling for tool wear and RUL prediction, ViT-based semi-supervised milling diagnosis, and burst-informed acoustic-emission diagnostics20–22. For industrial assemblies, impact-response monitoring in closed pipeline networks and wearable impedance plus multi-channel fusion in steel frames highlight progress toward deployable field diagnostics23,24. In bolt monitoring, however, signatures that separate fine-grained loosening states can be weak and easily masked by ambient noise or domain shift, which can limit generalization. Two recurring challenges emerge from the literature. First, feature imbalance and dominance: when representations are fused naively, high-dimensional cues can dominate and suppress weaker but diagnostically important transient information3,10. Second, scale sensitivity and temporal sparsity: micro-loosening can manifest as sparse impulsive events at multiple temporal scales, and fixed-scale encoders or single-resolution spectrograms may under-sample these events11,19,25. Domain shifts (sensor drift, sampling differences, and environmental perturbations) further degrade the stability of learned representations. To advance bolted-joint monitoring toward robust, fine-grained condition classification, recent studies emphasize stronger time–frequency feature engineering and fusion across representations or sensing channels rather than relying on a single handcrafted descriptor3,6,10,23,24,26–28. In practical settings, fusion must be designed to preserve weak but diagnostically crucial transient cues while leveraging more stable complementary cues (e.g., waveform morphology vs. time–frequency textures), and evaluation protocols must minimize leakage due to temporal autocorrelation and overlapping windows. On the feature front, higher-order and sparsity-aware decompositions (e.g., low-rank + sparse separation) and phase-sensitive descriptors have shown potential to disentangle intrinsic structural components from transient interference19. Additionally, multi-resolution training paradigms and hybrid encoder stacks (e.g., CNN Bi-LSTM attention/Transformer) have been proposed to capture both local impulsive events and long-range temporal dependencies29–32. Despite these advancements, a critical gap remains in processing ultra-high-frequency (1 MHz) vibration signals for bolt loosening detection. While high-rate sampling is important for capturing micro-transients indicative of early-stage loosening, it substantially increases computational and memory costs. Conventional averaging-based downsampling can attenuate sparse impulses, whereas relying on lower-rate sensing can miss key micro-transient cues. This creates a practical dilemma: balancing computational efficiency with transient preservation—a challenge that existing methods11,12,19 do not directly address. Motivated by these observations, we introduce PGRF-Net (Physics-Guided Resampling and Asymmetric Fusion Network), a deep learning framework designed for ultra-high-frequency bolt loosening detection. To avoid ambiguity with PDE-constrained physics-informed neural networks, we use the term physics-guided for this signal-processing-driven design. PGRF-Net builds on prior work in ultrasonic/vibration monitoring3–5,16–18, time–frequency analysis11,12,19, and modern fault diagnosis architectures29–33, while explicitly targeting the challenges of 1 MHz signals. The main contributions are: 1. Physics-guided adaptive resampling: a resampling strategy that compresses 1 MHz vibration signals by 150 : 1 while preserving impulse-like transients via a hybrid peak/energy operator. 2. Heterogeneous gradient–spectral representation: a 5-channel tensor that combines time–frequency magnitude, time/frequency gradients, a normalized energy map, and a time-domain embedding to retain complementary cues. 3. Asymmetric attention-based fusion: an attention mechanism that fuses pseudo-image and waveform representations derived from the same signal to mitigate representation imbalance and improve robustness. In addition to the model, we provide an end-to-end pipeline and a C#-based desktop application for real-time inference, illustrating practical feasibility for edge deployments. Methods Experimental setup and data acquisition We collect ultra-high-frequency vibration signals using a PVDF film sensor (LDT0-028K) attached to a bolted joint. Signals are sampled at 1 MHz. Six bolt states are created using controlled torque settings (Severe Loose, Loose, Critical, Transition, Secure, Over-Tight). Under the main setting (compression 150 : 1), this yields 109,668 segments in total, with a temporal-split test set of 16,131 segments. For each state, five independent acquisition files (run01–run05) are recorded under different scene conditions (environment and setup-state variations), so each file is treated as a scene-level unit in split design. Details and test-split class statistics are summarized in Table 1, and the experimental setup is shown in Fig. 1. Table 1Class statistics and diagnostic characteristics (temporal-split test set, compression 150 : 1, ). Torque is reported relative to a reference tightening level (); values above indicate intentional over-tightening.ClassTorqueSamplesPhysical characteristicsDiagnostic challengeSevere loose3679Macro-impact transients, low dampingDetection reliability (safety-critical)Loose2547Multiple micro-impacts, increasing dampingMagnitude discriminationCritical1981Damped oscillations, resonance shiftsBoundary classificationTransition3113Subtle transients, gradient featuresDynamic state recognitionSecure2547Damped vibrations, stable resonancesDistinguishing from over-tightOver-tight2264High stiffness, suppressed vibrationsSubtle parameter changesTotal–16,131––The bold values in Table 1 represent the total sample size of this experiment. Diff erent data cropping methods willlead to inconsistent sample quantities, so the sample size corresponding to this experiment is specially specified . Class statistics and diagnostic characteristics (temporal-split test set, compression 150 : 1, ). Torque is reported relative to a reference tightening level (); values above indicate intentional over-tightening. The bold values in Table 1 represent the total sample size of this experiment. Diff erent data cropping methods willlead to inconsistent sample quantities, so the sample size corresponding to this experiment is specially specified . Experimental setup and data collection process. Panels (a), (b) show the bolt structure and LDT0-028K PVDF sensor placement; panels (c), (d) show torque application and measurement using a digital torque wrench to generate labeled loosening states. Physics-guided adaptive resampling Bolt loosening induces intermittent micro-impacts and stick-slip events that manifest as sparse high-amplitude transients. Let denote the raw vibration sequence sampled at 1 MHz, and let denote the compressed sequence (here , corresponding to 150 : 1 compression). Averaging-based downsampling uses where . For sparse impulsive events, this operation can suppress peak magnitudes and may remove diagnostically useful transients. To preserve impulse peaks while retaining local energy trends, we use a hybrid peak/energy operator: with . In the main setting, we use and overlap ratio . A qualitative comparison against common operators is shown in Fig. 7. Impulse preservation under aggressive compression (150 : 1). We compare common operators (decimation, mean pooling, max pooling) against the proposed hybrid peak/energy operator (). The hybrid operator better retains the transient magnitude in the compressed domain, supporting the physical motivation that loosening-related impacts are sparse and high-amplitude. Multi-representation learning and fusion Although the architecture contains two branches, both are derived from the same vibration signal: (i) a pseudo-image branch built from time–frequency and gradient-enhanced representations; and (ii) a 1D signal branch that processes the resampled waveform. We therefore describe PGRF-Net as multi-representation learning rather than true multi-sensor multimodal fusion. Heterogeneous gradient-spectral representation After resampling, we transform the 1D sequence into a 5-channel tensor (with in our implementation). The channels are designed to preserve complementary frequency, gradient, energy, and morphology cues: Log-amplitude spectrogram , computed with Hann window length , hop size , and . Log-amplitude spectrogram , computed with Hann window length , hop size , and . Time gradient and frequency gradient , obtained by finite differences on to emphasize transient onsets and resonance-ridge shifts. Time gradient and frequency gradient , obtained by finite differences on to emphasize transient onsets and resonance-ridge shifts. Normalized energy map E, derived from squared STFT magnitude to emphasize high-energy regions. Normalized energy map E, derived from squared STFT magnitude to emphasize high-energy regions. Time-domain embedding, built by uniformly sampling the 1D sequence and tiling it across width to retain coarse waveform morphology. Time-domain embedding, built by uniformly sampling the 1D sequence and tiling it across width to retain coarse waveform morphology. We resize the four time–frequency maps to and stack them with the time-domain embedding to construct . For explicit reproducibility, we define the STFT channel generation as (Fig. 2) and then compute gradient/energy channels via Five-channel heterogeneous representation used by PGRF-Net: log-amplitude spectrogram, time gradient, frequency gradient, normalized energy map, and time-domain embedding. PGRF-Net architecture and training objective PGRF-Net follows a 2-2-2 configuration: Image Encoder Type 2 (ResNet-101), Signal Encoder Type 2 (Hybrid CNN-LSTM-Transformer), and Fusion Type 2 (Multi-Head Attention). For ablation IDs Exp 022–222, we use the code tuple (i, s, f), where i denotes pseudo-image encoder type (0 = Light CNN, 1 = ResNet-50, 2 = ResNet-101), s denotes signal encoder type (0 = RNN-like, 1 = LSTM-like, 2 = Hybrid), and f denotes fusion type (0 = concatenation, 1 = simple attention, 2 = multi-head attention) (Figs. 3, 4 and Table 2). Overall workflow of PGRF-Net for bolt-loosening detection, from raw 1 MHz vibration signals through physics-guided adaptive resampling, heterogeneous representation construction, dual-branch encoding, cross-modal attention fusion, and label-smoothed classification. Dual-branch feature encoding module. The pseudo-image branch (ResNet-101 stages) and the 1D signal branch (CNN-LSTM-Transformer) produce visual and signal embeddings for subsequent fusion. Table 2Method summary for the primary configuration (Exp 222, , compression 150 : 1). The pseudo-image branch and waveform branch are two complementary representations extracted from the same vibration stream.BlockInput OutputKey settingsResampling (1 MHz, length L) (length 8192)Hybrid peak/RMS pooling, , overlap , compression 150 : 15-channel representation = {log-STFT, time-grad, freq-grad, energy map, time-embedding}Pseudo-image encoderResNet-101, first conv adapted to 5 channels; projection to 256-dSignal encoderMulti-scale 1D CNN stem + Bi-LSTM + Transformer encoder; projection to 256-dFusionAsymmetric multi-head attention + residual + normalizationClassifier2-layer MLP with dropoutTraining lossCombined label-smoothing CE () + focal loss (), mixing Method summary for the primary configuration (Exp 222, , compression 150 : 1). The pseudo-image branch and waveform branch are two complementary representations extracted from the same vibration stream. To keep notation traceable without reintroducing an equation-heavy narrative, we provide a compact formula map in Table 3. Table 3Compact summary of core mathematical definitions used in PGRF-Net. This table serves as a notation reference that complements the detailed method description.ComponentFormulaRole in pipelineLinear downsamplingReference operator; may suppress sparse impulses.Hybrid resamplingPreserves peak transients and local energy trends under aggressive compression.STFT constructionTime–frequency transform before 5-channel representation.Log-spectrogramMain spectral-intensity channel.Gradient channelsEmphasizes onset dynamics and resonance-ridge shifts.Energy mapNormalized energy-distribution channel.Cross-modal attention Fuses pseudo-image and waveform embeddings with residual normalization.Label smoothingReduces over-confidence and improves calibration.Loss termsBalances class imbalance and hard-sample emphasis.Final objectiveOptimization target for PGRF-Net training. Compact summary of core mathematical definitions used in PGRF-Net. This table serves as a notation reference that complements the detailed method description. The fusion module performs attention-based interaction between pseudo-image and waveform features, followed by residual addition and normalization to stabilize cross-representation integration. For one attention head, the core fusion equations are (Figs. 5, 6) Cross-modal attention fusion module. Visual and signal embeddings are normalized and fused by multi-head cross-attention to obtain a fused embedding. Classification head and loss computation. The fused embedding is projected to logits by a linear classifier, while ground-truth labels are converted to smoothed targets for label-smoothed training loss. To address class imbalance and improve calibration, we first smooth hard labels and define two complementary losses: and optimize the combined objective: where , label smoothing , and focal parameter . Evaluation protocol and data integrity High-rate vibration data exhibit strong temporal autocorrelation; naive random shuffling or intra-file overlap can leak near-duplicate windows across train/val/test. We therefore use a block-wise temporal split within each file (temporal split) as the main protocol: each measurement file is split into contiguous train/val/test time blocks. We additionally report a file-disjoint split (no file appears in more than one split), which serves as a cross-scene hold-out test because each file corresponds to a distinct acquisition scene. Overlapping windows are used within each split only. For the main setting of Exp 222, we report repeated runs with three seeds (256, 2026, and 3407); mean ± standard deviation are reported where applicable. For scene-level validation, we adopt a Leave-One-Scene-Out (LOSO-5) protocol in which one scene (run01–run05) is used for testing, one distinct scene is used for validation, and the remaining three scenes are used for training in each fold; final performance is reported as mean ± standard deviation across folds. Accordingly, lower performance under the file-disjoint split is interpreted as scene-transfer difficulty (background, excitation path, and coupling shifts) rather than as evidence of data leakage. Implementation and reproducibility All models are implemented in PyTorch (2.8.0) with Python 3.12 and CUDA 12.8. Experiments are run on an NVIDIA RTX 5090 GPU. In addition to the in-house ablation suite (Exp 022–222), we report a Zenodo external suite on the public Lab Structure Dynamic Test dataset (record 15516419)34: classical reference models Exp 401–404 and feature-engineering controls Exp 501–503. The Zenodo suite follows strict file-disjoint 3-fold CV with three seeds per fold (9 runs per experiment). Scripts for generating all figures and tables are included in the released code repository. Results Resampling operator evaluation The operator definition is given in Methods (Eqs. 1–4). Here we report empirical evidence that hybrid peak/energy compression preserves impulsive transients more faithfully than decimation, mean pooling, or max pooling alone (Fig. 7). Sensitivity analyses show that performance is stable across and compression ratios (Figs. 11, 12, Table 7, and Table 8), supporting and 150 : 1 as practical default settings. Sensitivity analysis for the resampling coefficient (Exp 222, temporal split, compression 150 : 1, seed = 3407). Sensitivity analysis for compression ratio (Exp 222, temporal split, , seed = 3407). Table 7Sensitivity analysis for the hybrid resampling coefficient (clean test split). All runs are Exp 222 (temporal split), compression 150 : 1, seed = 3407.Acc. (%)0.594.850.795.120.995.00 Sensitivity analysis for the hybrid resampling coefficient (clean test split). All runs are Exp 222 (temporal split), compression 150 : 1, seed = 3407. Table 8Sensitivity analysis for compression ratio (clean test split). All runs are Exp 222 (temporal split), , seed = 3407. Note that the number of test windows varies with the compression ratio due to the sliding-window extraction rule.Compression ratioAcc. (%)#Test windows100:194.1724,510150:195.1216,131200:194.7011,970 Sensitivity analysis for compression ratio (clean test split). All runs are Exp 222 (temporal split), , seed = 3407. Note that the number of test windows varies with the compression ratio due to the sliding-window extraction rule. Overall performance and benchmark comparison Table 4 compares the reproduced baseline (Exp 022) and PGRF-Net (Exp 222) on the clean test split under the temporal split protocol. Table 4Clean test accuracy comparison for the reproduced baseline (Exp 022) and the proposed PGRF-Net (Exp 222) under the temporal split.MethodAcc (%)Baseline (Exp 022)68.33PGRF-Net (Exp 222)95.12 Clean test accuracy comparison for the reproduced baseline (Exp 022) and the proposed PGRF-Net (Exp 222) under the temporal split. PGRF-Net (Exp 222) achieves 95.12% clean test accuracy, compared with 68.33% for the reproduced baseline (Exp 022) under the same temporal split. Because both branches are derived from the same vibration signal, the improvement reflects the combined contributions of transient-preserving compression, heterogeneous representation, and fusion design rather than introducing an additional sensor. External deep-learning baselines under a reference protocol To broaden external comparisons across CNN-based, 1D, and sequence models, we trained seven external reference models (Exp 301–307) under a unified reference protocol that excludes the proposed hybrid resampling and 5-channel representation pipeline. Waveform baselines use linearly resampled 1D inputs, whereas image baselines use standard RGB log-spectrograms. Because this protocol uses a larger clean test set under the temporal split ( vs. 16, 131), these results are presented as supplementary comparisons and are not directly comparable with Table 4. Among the external baselines, the strongest model is the InceptionTime-like 1D encoder (Exp 302, 88.96% accuracy), followed by WDCNN (Exp 301, 88.21%) and CNN-BiLSTM-Attention (Exp 307, 87.55%). ResNet1D and the Transformer encoder achieve 81.25% and 78.52%, respectively. On spectrogram inputs, EfficientNet-B0 reaches 80.85%, whereas ResNet-18 reaches 23.33%, indicating that standard RGB spectrogram conversion alone is insufficient for this dataset. These results suggest that the gains of PGRF-Net are not explained by pretrained ResNet transfer alone, because the reference spectrogram CNN baselines remain substantially weaker. Owing to the protocol and test-size mismatch, these values are treated as supplementary comparisons rather than a protocol-matched ranking (Fig. 8 and Table 5). External baseline comparison on the clean test split under the unified reference protocol. Waveform baselines use linear resampling and spectrogram baselines use standard RGB log-spectrograms; all results are reported on the temporal split with test windows. Table 5External deep-learning baseline comparison under a unified baseline protocol that excludes the proposed resampling and representation design. All runs use the temporal split and seed = 3407. Waveform baselines use linear resampling, and image baselines use standard RGB log-spectrograms. Because this protocol uses a different clean test-set size () than the main PGRF-Net setting, these values are reported as supplementary evidence and are not directly comparable with Table 4.Exp.MethodInputAcc (%)Macro F1Weighted F1301WDCNNLinear-resampled waveform88.210.8800.882302InceptionTime-likeLinear-resampled waveform88.960.8900.889303ResNet1DLinear-resampled waveform81.250.8110.812304ResNet-18Standard RGB log-spectrogram23.330.2210.208305EfficientNet-B0Standard RGB log-spectrogram80.850.8060.808306Transformer encoderLinear-resampled waveform78.520.7810.783307CNN-BiLSTM-AttentionLinear-resampled waveform87.550.8740.875 External deep-learning baseline comparison under a unified baseline protocol that excludes the proposed resampling and representation design. All runs use the temporal split and seed = 3407. Waveform baselines use linear resampling, and image baselines use standard RGB log-spectrograms. Because this protocol uses a different clean test-set size () than the main PGRF-Net setting, these values are reported as supplementary evidence and are not directly comparable with Table 4. Component ablation study We conduct systematic ablations (Exp 022–222) and sensitivity analyses to justify key hyperparameters ( and compression ratio) and to quantify seed variability and split-control behavior (Fig. 9 and Table 6): Main ablation on the clean test split for Exp 022–222 under the temporal split (, compression 150 : 1, seed = 3407). Table 6Main ablation on the clean test split under the proposed temporal (block-wise intra-file) split. All runs use , compression 150 : 1, and seed = 3407. Experiment code is decoded as (i, s, f): i (pseudo-image encoder type: 0 = Light CNN, 1 = ResNet-50, 2 = ResNet-101), s (signal encoder type: 0 = RNN-like, 1 = LSTM-like, 2 = Hybrid), and f (fusion type: 0 = concatenation, 1 = simple attention, 2 = multi-head attention).Exp.Configuration (high-level)Acc. (%)022Light CNN pseudo-image encoder + hybrid signal encoder + multi-head attention fusion68.33122ResNet-50 pseudo-image encoder + hybrid signal encoder + multi-head attention fusion94.59202ResNet-101 pseudo-image encoder + RNN-like signal encoder + multi-head attention fusion94.23212ResNet-101 pseudo-image encoder + LSTM-like signal encoder + multi-head attention fusion93.47220ResNet-101 pseudo-image encoder + hybrid signal encoder + concat fusion94.55221ResNet-101 pseudo-image encoder + hybrid signal encoder + simple attention94.68222PGRF-Net (ResNet-101 + hybrid signal encoder + multi-head fusion)95.12 Main ablation on the clean test split under the proposed temporal (block-wise intra-file) split. All runs use , compression 150 : 1, and seed = 3407. Experiment code is decoded as (i, s, f): i (pseudo-image encoder type: 0 = Light CNN, 1 = ResNet-50, 2 = ResNet-101), s (signal encoder type: 0 = RNN-like, 1 = LSTM-like, 2 = Hybrid), and f (fusion type: 0 = concatenation, 1 = simple attention, 2 = multi-head attention). Sensitivity analyses and leakage-avoidance controls As specified in the Methods section, file-disjoint results are interpreted as strict cross-scene stress-test outcomes, and performance gaps relative to the temporal split are analyzed as scene-transfer difficulty (Tables 7, 8, 9, 10). Table 9Seed variability for the main setting of Exp 222 (temporal split, , compression 150 : 1) on the clean test split. Mean and standard deviation are computed across the three reported seeds (256, 2026, and 3407).SeedAcc. (%)#Test windows25694.6116,131202695.0216,131340795.1216,131Mean ± Std94.91 ± 0.27– Seed variability for the main setting of Exp 222 (temporal split, , compression 150 : 1) on the clean test split. Mean and standard deviation are computed across the three reported seeds (256, 2026, and 3407). Table 10Leakage-avoidance control via file-disjoint split. Exp 222 with file split (seed = 3407) is compared to the temporal split main result (seed = 3407) on the clean test split.Split modeAcc. (%)#Test windowsTemporal (block-wise intra-file)95.1216,131File-disjoint (control)19.0631,104 Leakage-avoidance control via file-disjoint split. Exp 222 with file split (seed = 3407) is compared to the temporal split main result (seed = 3407) on the clean test split. The file-disjoint gap in Fig. 10 is therefore interpreted as cross-scene transfer difficulty rather than data inconsistency. Across the five scene-level files, acquisition background and preload evolution differ; moreover, some segments are intentionally near operating boundaries, where adjacent states naturally share similar signatures. This heterogeneity improves dataset realism, and the five scene-level files jointly cover a broader range of realistic deployment conditions. Leakage-avoidance control on the clean test split: temporal (block-wise intra-file) split vs. strict file-disjoint split for Exp 222. The file-disjoint test uses scene-level file separation (run01–run05), where files are collected under different environment and preload trajectories. Impact of the pseudo-image encoder (Pretrained ResNet) Under the same temporal split and main settings (, compression 150 : 1, seed=3407), the reproduced baseline Exp 022 remains in the 68–69% clean-accuracy range, while Exp 122 reaches 94.59%. This large gain (about +25 percentage points) confirms that the 5-channel pseudo-image representation contributes strong discriminative information. In safety terms, this branch is particularly important around critical-loosening boundaries, where decision errors are costlier than in clearly separated states (Figs. 11, 12, 13). Robustness to controlled AWGN injection across SNRs for Exp 222 under the temporal split (reported run, seed = 3407). Fusion strategy insights (attention vs. concatenation) For fusion variants, Exp 220 (concatenation) achieves 94.55%, Exp 221 (simple attention) achieves 94.68%, and Exp 222 (multi-head attention) achieves 95.12% under the same clean setting. Exp 222 therefore delivers the strongest clean performance among fusion variants and the most resilient low-SNR behavior (Fig. 13, Table 11). Table 11Comprehensive robustness comparison under controlled AWGN injection for Exp 022–222 (temporal split, , compression 150 : 1, seed = 3407). Values are test accuracy (%).Exp.Clean20 dB10 dB5 dB0 dB−5 dB02268.3367.8633.6315.8915.7915.7912294.5990.4519.1914.0414.0420.5520294.2385.1641.2320.1714.0414.0421293.4791.4645.0125.5219.3019.3222094.5590.6218.8314.2814.0516.1922194.6891.6455.4124.7718.4915.7922295.1292.0773.6267.2758.0130.83 Comprehensive robustness comparison under controlled AWGN injection for Exp 022–222 (temporal split, , compression 150 : 1, seed = 3407). Values are test accuracy (%). Error analysis Figure 14 (Fig. 14 in the manuscript sequence) shows that the largest residual confusion appears in the Secure vs. Over-Tight pair, while most other classes stay close to diagonal dominance. Relative to the baseline matrix, PGRF-Net reduces off-diagonal spread for Severe Loose, Loose, and Critical, supporting the claim that transient-preserving compression and heterogeneous representation improve separability in boundary-adjacent states. Test-split confusion matrices (clean) for the baseline (Exp 022) and PGRF-Net (Exp 222) under the temporal split. The matrices are computed on the held-out test split only (not the full dataset). Despite the strong overall accuracy, class-level diagnostics from a representative seed = 3407 run (Acc = 94.68%) highlight the remaining challenges (Table 12): Table 12Per-class performance for PGRF-Net (Exp 222) on the clean test split under the temporal split (representative seed = 3407 run, Acc = 94.68%). The best-run overall accuracy is reported in Table 4.ClassPrecisionRecallF1-ScoreSevere Loose0.99970.93310.9653Loose0.95860.94580.9522Critical0.92140.98230.9509Transition0.98410.97270.9784Secure0.89170.91830.9048Over-tight0.89080.93330.9116 Per-class performance for PGRF-Net (Exp 222) on the clean test split under the temporal split (representative seed = 3407 run, Acc = 94.68%). The best-run overall accuracy is reported in Table 4. Transition states: The Transition class reaches an F1-score of 0.9784, suggesting that the gradient-enhanced representation captures distinctive cues for intermediate torque conditions. Secure vs. over-tight boundary Secure (F1 = 0.9048) and Over-Tight (F1 = 0.9116) are more confusable than other classes. This is consistent with a saturation-like behavior at high preload: once contact stiffness approaches an upper regime, additional torque may produce smaller changes in the vibration signature. Improving over-tight detection may therefore benefit from complementary sensing (e.g., strain-based measurements) in future work. Severe loose class The Severe Loose state has precision 0.9997 and recall 0.9331 on the clean test split, indicating a low false-alarm rate for this safety-critical condition while still recovering most severe-loose samples. Generalization and robustness evaluation We evaluate robustness by injecting controlled AWGN at different SNRs during evaluation. A comprehensive comparison across Exp 022–222 is reported in Table 11, and the Exp 222 curve (reported run, seed = 3407) is shown in Fig. 13. Exp 222 remains highest from clean to dB, while all models degrade as SNR decreases. External cross-condition generalization on public data To evaluate cross-condition transfer, we tested both the proposed pipeline and feature-engineering controls on the public Lab Structure Dynamic Test dataset (Zenodo record 15516419)34. The suite includes classical references (Exp 401–404) and feature-engineering controls (Exp 501–503), all under strict file-disjoint 3-fold CV with three seeds per fold. Following the requested A/B/C design logic, A and B are a same-model downsampling control pair (A = Exp 502, B = Exp 501), while C covers non-PGRF alternatives through both an alternative feature pipeline control (Exp 503) and other feature/model families (Exp 401–404) (Fig. 15). Clean fold-test macro-F1 on the public Zenodo dataset for Exp 401–404 and Exp 501–503 (3-fold file-disjoint CV, three seeds per fold). Error bars indicate standard deviation over all runs. Across all seven experiments, clean fold-test macro-F1 ranges from 0.757 to 0.849, and the d15_cfg22 fold remains the most difficult across methods. Figures 16 and 17 show that Exp 402 and Exp 503 exhibit wider run-level dispersion than Exp 401 and Exp 404, consistent with their larger cross-run standard deviations in Table 13. In the strict same-model A/B control, A (Exp 502, regular averaging) and B (Exp 501, physics-guided adaptive downsampling) show close clean performance (macro-F1 difference 0.010). For C, Exp 503 together with Exp 401–404 provides a broader non-PGRF set with alternative feature/model designs; this C group spans clean macro-F1 values of 0.757–0.849, indicating wider variability than the A/B pair. Under noisy evaluation, both Exp 501 and Exp 404 degrade substantially toward 0 dB, indicating that condition-shift robustness remains an open challenge despite strong clean performance on easier folds (Figs. 18, 19 and Tables 14, 15). Run-level macro-F1 distribution over all nine runs per experiment (3 folds 3 seeds). This view complements mean ± std statistics by showing dispersion and outlier behavior. Run-level accuracy distribution over all nine runs per experiment (3 folds 3 seeds). This view complements mean ± std statistics by showing dispersion and outlier behavior. Table 13Zenodo external benchmark with strict file-disjoint 3-fold CV and three seeds per fold (total 9 runs per experiment). Values are fold-test mean ± std across all runs.Exp.MethodClean Macro-F1Clean acc (%)Worst-fold Macro-F1401WDCNN0.849 ± 0.15886.37 ± 14.370.644 (d15_cfg22)402InceptionTime0.757 ± 0.24883.34 ± 14.700.436 (d15_cfg22)403ResNet1D0.814 ± 0.17583.81 ± 14.830.585 (d15_cfg22)404CNN-BiLSTM-Attn0.839 ± 0.14584.84 ± 14.090.649 (d15_cfg22)501PGRF-Net (proposed pipeline)0.809 ± 0.11582.07 ± 11.150.664 (d15_cfg22)502Avg-Downsample (A)0.819 ± 0.11783.19 ± 11.090.666 (d15_cfg22)503Decimate-based pipeline (C)0.825 ± 0.16085.37 ± 12.810.614 (d15_cfg22) Zenodo external benchmark with strict file-disjoint 3-fold CV and three seeds per fold (total 9 runs per experiment). Values are fold-test mean ± std across all runs. Fold-wise clean macro-F1 (mean over three seeds) for Exp 401–404 and Exp 501–503. The d15_cfg22 fold is consistently more difficult across all tested pipelines. Noise robustness on the public Zenodo dataset (fold-test macro-F1, mean over folds and seeds) for Exp 501 and Exp 404. Table 14Feature-engineering A/B/C mapping on the Zenodo dataset. Group A and Group B share the same deep architecture and isolate only the downsampling effect; Group C covers alternative feature/model designs via Exp 503 and Exp 401–404.Design groupExperimentClean Macro-F1Clean acc (%)A: Regular averaging downsampling + same modelExp 5020.819 ± 0.11783.19 ± 11.09B: Physics-guided adaptive downsampling + same modelExp 5010.809 ± 0.11582.07 ± 11.15C1: Alternative feature pipeline control (decimate)Exp 5030.825 ± 0.16085.37 ± 12.81C2: Other feature/model families (non-PGRF)Exp 401–4040.757–0.84983.34–86.37 Feature-engineering A/B/C mapping on the Zenodo dataset. Group A and Group B share the same deep architecture and isolate only the downsampling effect; Group C covers alternative feature/model designs via Exp 503 and Exp 401–404. Table 15Noise robustness on Zenodo fold-test split (macro-F1). Values are mean ± std over folds and seeds.MethodClean10 dB5 dB0 dB501 PGRF-Net (proposed)0.809 ± 0.1150.716 ± 0.1470.545 ± 0.1950.435 ± 0.172404 CNN-BiLSTM-Attn0.839 ± 0.1450.695 ± 0.1890.607 ± 0.2240.506 ± 0.246 Noise robustness on Zenodo fold-test split (macro-F1). Values are mean ± std over folds and seeds. Computational efficiency To provide deployment-relevant evidence, we additionally performed a CPU-only benchmark on a local workstation with an AMD Ryzen 5 9600X processor (6 cores/12 threads), Windows 11, and PyTorch 2.7.1, using the trained Exp 222 checkpoint. The summary is shown in Table 16. Table 16CPU-only efficiency benchmark for the Exp 222 checkpoint on AMD Ryzen 5 9600X. Latency statistics are measured over 200 timed iterations after warm-up.MetricValueCPU platformAMD Ryzen 5 9600X (6C/12T)PyTorch CPU threads6Total parameters51.76 MApproximate FLOPs per sample (PyTorch profiler)16.66 GMean latency (batch = 1)51.17 msP50/P95 latency (batch = 1)50.77/55.88 msThroughput (batch = 1)19.54 samples/sMean latency (batch = 4)145.54 msThroughput (batch = 4)27.48 samples/sProcess RSS delta during benchmark111.84 MB CPU-only efficiency benchmark for the Exp 222 checkpoint on AMD Ryzen 5 9600X. Latency statistics are measured over 200 timed iterations after warm-up. Representation weighting analysis Average weighting Figure 20 shows that the average attention weights are strongly biased toward the waveform representation. This is expected because the waveform retains higher temporal resolution than the pseudo-image representation produced by time–frequency transforms and resizing. This average trend does not imply that the pseudo-image branch is dispensable; branch importance is sample-dependent and pseudo-image cues become critical on hard boundary samples. Average attention weights under the temporal split. For many samples, the model assigns higher weight to the waveform representation than to the pseudo-image representation derived from STFT features. Outlier cases However, averages can obscure behavior on challenging boundary cases. The outlier analysis in Fig. 21 shows that attention can shift toward the pseudo-image branch for hard samples, especially near critical operating boundaries where misclassification risk is highest. Outlier analysis of attention weights. While the waveform representation dominates on average, the pseudo-image branch can receive substantially higher weight for boundary cases (e.g., the highest-ranked Severe Loose outlier), indicating complementary cues from the time–frequency representation. For hard samples—particularly near the boundaries of Severe Loose/Critical/Transition/Secure—the attention mechanism assigns substantially higher weight to the pseudo-image representation (up to in our analysis). This supports the role of the pseudo-image branch as a safety-relevant complementary channel in ambiguity-prone operating regions. We note that this analysis is descriptive; further controlled studies are required to quantify boundary-risk reduction under deployment shifts. Discussion PGRF-Net targets three practical challenges in ultra-high-frequency bolt loosening detection: (1) computational efficiency versus transient preservation, (2) representation imbalance between time–frequency textures and waveform cues, and (3) fine-grained state discrimination under domain shift and noise. The hybrid resampling and heterogeneous representation aim to retain sparse impact signatures in the compressed domain (Fig. 7), while the fusion module aggregates complementary cues from two representations derived from the same signal. Regarding data integrity, the file-disjoint split control (Table 10 and Fig. 10) shows that cross-scene generalization is substantially more difficult than intra-file temporal generalization, plausibly due to scene-dependent background and acquisition differences across scene-level files. Robustness under controlled AWGN injection is summarized in Table 11 and Fig. 13; Exp 222 is strongest across all reported SNR levels, but all models still degrade in low-SNR regimes. In the independent Zenodo benchmark (Table 13 and Figs. 15, 16, 17, 18, 19), all tested pipelines exhibit a common challenging fold (d15_cfg22), suggesting persistent condition-shift sensitivity beyond the primary platform. Finally, the remaining ambiguity between Secure and Over-Tight is consistent with smaller vibration changes at high preload, which can limit separability using vibration alone. Study limitations Two limitations should be considered when interpreting these results. First, although the dataset contains five scene-level files per state and file-disjoint evaluation therefore measures within-platform cross-scene generalization, all data are still collected on one bolt platform with one primary sensor configuration. Consequently, current evidence does not yet establish cross-sensor or cross-device generalization. A dedicated multi-sensor and multi-device dataset for this task is under construction. Second, although we report two external suites (Exp 301–307 on our dataset and Exp 401–404/501–503 on the public Zenodo dataset), this is not yet an exhaustive protocol-matched benchmark against all recent methods. Specifically, the Zenodo experiments emphasize strict split control and a controlled A/B/C feature-engineering comparison, while broader cross-device retuning and domain adaptation remain future work. Therefore, these results are best interpreted as complementary evidence rather than a definitive universal ranking. Real-world deployment feasibility Practical monitoring systems often require on-device inference under limited compute budgets. To support deployment, we implemented a standalone desktop application in C# (.NET 10.0) that integrates an ONNX-exported model for real-time inference and USB-based data acquisition. On AMD Ryzen 5 9600X, the CPU-only benchmark reports a mean latency of 51.17 ms per sample (batch = 1, about 19.54 samples/s) and 27.48 samples/s at batch = 4 (Table 16), providing a reproducible software-side efficiency reference for this deployment setting. Conclusion PGRF-Net addresses a practical bottleneck in bolt loosening monitoring: learning from 1 MHz vibration signals under extreme compression while retaining transient signatures that correlate with contact dynamics. The proposed pipeline integrates a physics-guided hybrid resampling operator, a heterogeneous gradient–spectral representation, and an asymmetric fusion module over two complementary representations derived from the same signal. Under the main block-wise temporal split, PGRF-Net reaches a best-run clean test accuracy of 95.12% and a three-seed clean mean of 94.91% ± 0.27%. A strict file-disjoint split is additionally reported as a cross-scene hold-out control and highlights the difficulty of scene transfer even within the same hardware setup. On the Zenodo benchmark with 3-fold file-disjoint CV and three seeds per fold, clean macro-F1 spans 0.757–0.849 across Exp 401–404 and 0.809–0.825 across Exp 501–503, while the d15_cfg22 fold and low-SNR settings remain difficult. Overall, these findings support the proposed design within the evaluated settings, while broader cross-device validation and fully protocol-matched benchmarking remain important future work. Acknowledgements This research was supported by the National Natural Science Foundation of China (Grant No. 62373372) and the National Innovative Entrepreneurship Undergraduate Training Programme (Grant Nos. 202510489026 and 202410489004). Author contributions Z.L.: Conceptualization, Methodology, Software, Validation, Writing - Original Draft. X.S.: Data Curation, Investigation, Visualization. K.X.: Writing - Review & Editing, Supervision. C.W.: Formal analysis. S.L.: Conceptualization. W.Z.: Resources. All authors reviewed and approved the manuscript. Data availability The in-house dataset collected in this study (released as data. zip) and all source code are available at https://github. com/bolt-phm/PIR-Net. The external benchmark data used for cross-condition evaluation are publicly available from Zenodo (record 15516419) . Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Zihao Luo, Xinrui Shi have contributed equally to this work. References Associated Data Data Availability Statement The in-house dataset collected in this study (released as data. zip) and all source code are available at https://github. com/bolt-phm/PIR-Net. The external benchmark data used for cross-condition evaluation are publicly available from Zenodo (record 15516419) .