During service, bogie frames are subjected to combined low-frequency quasi-static loads and high-frequency wheel-rail excitations. They exhibit prominent frequency-band characteristics and are prone to fatigue damage. Traditional simulation methods rely heavily on boundary conditions, suspension parameters, and load inputs, making it difficult to achieve accurate stress prediction for critical locations under complex line operating conditions. To address this problem, this paper proposes a bogie frame stress prediction method that integrates finite element (FE) analysis, frequency-segmented sensor placement, and dual-band neural networks. FE analysis is performed to identify fatigue-vulnerable locations of the frame under low-frequency and high-frequency loads as stress monitoring positions. Combined with the optimal sensor placement method, this paper determines the layout scheme of vibration sensors for the frame under low-frequency and high-frequency excitation conditions. A dual-band stress prediction model for low and high frequencies is established based on synchronous acceleration-stress data acquired from line tests. A comparative study is conducted on long short-term memory (LSTM), temporal convolutional network, Transformer, and attention-enhanced models. Meanwhile, attention ablation and perturbation analyses are carried out to verify the model’s capability to capture key temporal segments. The results show that the LSTM with attention (LSTM-ATT) model achieves favorable overall performance in predicting low-frequency trend stress and high-frequency dynamic stress. The prediction error of the fatigue utilization factors at key positions is controlled within 10%, and an adequate safety margin is maintained for high-risk measuring points. The proposed method provides an effective solution for structural health monitoring and fatigue life evaluation of critical components of bogie frames.
Concealed voids of the cement-emulsified asphalt (CA) mortar layer in the slab track system pose considerable structural risks to high-speed railways, necessitating early detection. Finite element model updating methods offer global structural health assessment but are computationally demanding due to numerous simulations in high-dimensional optimization. To overcome these limitations, a novel damage identification method integrating a parameterized neural network surrogate model with time-domain sparse Bayesian learning is proposed. The surrogate model combines convolutional neural networks and long short-term memory networks, employing a dual-channel architecture for static structural parameters and dynamic impacts, further enhanced by position encoding and residual learning to predict acceleration sequences. Parameter optimization in this framework is performed with an improved particle swarm optimization algorithm with a local search strategy. The surrogate model achieved predictions with an average mean squared error of 0.0022 and an R -squared value of 0.913. The feasibility of the proposed method was validated on a scaled model of the slab track system. The numerical and experimental results demonstrate that the proposed method can successfully identify the location and severity of the CA mortar void and quantify the associated uncertainties. Furthermore, the proposed method significantly enhances computational efficiency, which offers a viable technical foundation for structural health monitoring of the slab track system.
Extracting weak and compound fault features from rotating machinery signals remains challenging because strong noise and multi-source modulation severely degrade the reliability of conventional decomposition methods. This study proposes an ordinal structure-guided adaptive ordinal pattern (OP) based mode decomposition framework, referred to as AOPMD, which introduces an ordinal structural entropy (OSE) criterion to quantitatively characterize the periodic regularity of impulsive fault transients. By exploiting anchor sequence consistency, the OSE criterion enables adaptive selection of key decomposition parameters and robust identification of fault-related modes, thereby alleviating the empirical parameter dependence inherent in conventional OP based mode decomposition. Numerical simulations and experimental studies on bearing faults, gear faults, and bearing and gear compound faults demonstrate that the proposed method achieves clearer fault feature separation and improved interpretability compared with representative decomposition techniques. These results indicate that the proposed AOPMD offers a robust and physically interpretable solution for compound fault diagnosis in rotating machinery under complex operating conditions.
Domain adaptation (DA) techniques have made significant advancements in the field of mechanical fault diagnosis. However, existing methods typically assume that source domain data is accessible during the DA phase. In real-world engineering scenarios, this assumption is often impractical due to limitations in data privacy, storage overheads, and transmission bandwidth. To address this issue, a novel source-free DA framework is proposed for rotating machinery fault diagnosis. First, the Progressive Pseudo-Labeling strategy is introduced, which gradually builds a reliable pseudo-label memory bank and dynamically updates it with historical information. This strategy effectively suppresses incorrect pseudo-labels. Then the Boundary Adversarial Calibration module is designed to incorporate low-confidence boundary samples into model training, enhancing feature discriminability. Furthermore, the Targeted Prototype Alignment constraint is introduced to promote intraclass compactness and interclass separation by pulling target samples toward their corresponding class prototypes while pushing them away from those of other classes. Extensive source-free cross-domain diagnostic experiments conducted on two rotating machinery datasets yielded average accuracies of 99.37 and 99.04%, respectively. The proposed framework achieves strong average performance and remains competitive across all evaluated transfer tasks, validating its feasibility and effectiveness in practical diagnostic scenarios.
To overcome the severe performance degradation and majority-class bias of traditional fault diagnosis methods under practical scenarios of extreme label scarcity and inherent class imbalance, this paper proposes a semi-supervised weighted stacked autoencoder with spectral peak significance constraints (SSWAEF). A nearest-neighbor consistency voting strategy coupled with a reciprocal-class-size sampling mechanism is first developed to construct a high-quality, class-balanced pseudo-labeled dataset. At the feature learning stage, considering the physical nature that mechanical faults typically manifest as energy concentrations at specific frequencies, a physics-informed weighted loss is designed. By incorporating power spectral density peak significance, this constraint amplifies gradients in critical frequency bands, forcing the network to preferentially extract discriminative fault structures rather than fitting broadband noise. For the fine-tuning stage, a dual-weighted cross-entropy loss is constructed, which integrates class-balancing weights and instance-confidence weights to ensure robust learning from minority classes without being misled by low-quality pseudo-labels. Extensive experiments on the Paderborn University dataset, a laboratory dataset, and an industrial field dataset validate the superiority of the proposed method. Under the extreme scenario with an imbalance and labeled rate of 0.2/0.2, SSWAEF maintains high accuracies of 99.63, 92.96, and 94.44% across the three datasets, respectively, demonstrating its exceptional robustness and diagnostic performance.
To address the difficulty in extracting fault features due to severe multi-source coupling and wide energy differences in composite faults, a sparse Ramanujan refined mode decomposition (SDR-SRRMD) method based on spectrum differential reconstruction is proposed. On the one hand, the SDR-SRRMD method distinguishes spectral lines based on the discrepancy between the original spectrum and the preprocessed spectrum derived via optimal weight impulse extraction. Based on this discrimination, the proposed method applies filtering to the original signal, which can achieve effective separation of the impulse components through reconstruction. On the other hand, the SDR-SRRMD method achieves accurate extraction of multi-period impulse components by constructing a sparse Ramanujan subspace and projecting the impulse components into their respective subspaces. Simulation and experimental signal analysis results demonstrate that this method can effectively separate and extract multi-period impulse components, serving as an effective solution for composite fault diagnosis.
Aiming at the difficulty of full-dimensional identification of multiform damages in aerospace and rail transit structures via single detection technology, this study proposes a heterogeneous data fusion detection method integrating ultrasonic guided wave and machine vision based on fuzzy D-S evidence theory. First, ultrasonic guided-wave signal centroid analysis and elliptical discretization imaging are used to preliminarily locate suspected damage areas. Guided by the positioning results, collaborative robots and machine vision technology (including Gaussian filtering, adaptive threshold segmentation, and morphological optimization) are employed for precise identification and quantification of surface damages. To address data randomness, fuzziness, and evidence conflicts, Gaussian membership functions are introduced to optimize evidence construction, and the Dempster rule is adopted for dual-source evidence fusion. A hidden damage location and quantification criterion based on spatial correlation hypothesis and nonlinear mapping function is also proposed. Experimental validation on 6061 aluminum alloy plate specimens shows that the method achieves 100% surface damage recognition rate and 98.3% hidden damage detection rate. The surface damage positioning error is ≤10.7 mm with size error ≤3.36%, and the hidden damage positioning error is ≤18.8 mm. The evidence conflict coefficient is below 0.3, avoiding the “belief paradox” in traditional D-S evidence theory. This method outperforms single detection technologies, providing an efficient solution for full-dimensional health monitoring of complex structures.
Reliable sensor data are fundamental to the effectiveness of long-term structural health monitoring (SHM) systems, in which measurement anomalies can compromise condition assessment, damage detection, and maintenance decision-making. This study presents a data separability-driven framework for automated anomaly classification of sensor measurements in bridge-based SHM applications. The framework integrates wavelet-based denoising, class balancing, guided data cleaning, multi-representation feature encoding, and deep-learning-based classification. To improve class distinction, 1-h signal segments are transformed into complementary representations, including time-frequency (TF) images, Gramian angular fields, and Markov transition fields. The framework was validated using month-long acceleration measurements collected from a full-scale cable-stayed bridge comprising 28,272 window-level samples from 38 accelerometers across six anomaly classes and one normal class. Quantitative analyses show that class balancing improves classification accuracy from 95.3 to 97.2%, while subsequent label refinement further improves accuracy to 99.2%. Feature fusion provides a further improvement, achieving 99.4% accuracy with Macro-F1 and Weighted-F1 scores exceeding 0.99. The empirical separability score increases from 0.33 for the original dataset to 0.70 for the fully processed dataset, demonstrating a strong relationship between enhanced class distinction and classification performance. The proposed framework provides a practical methodology for automated measurement-data quality assessment in long-term SHM systems. Although the results demonstrate strong performance on the adopted benchmark dataset, further validation using independent monitoring systems and grouped validation strategies is needed to establish cross-sensor and cross-site generalization.
Rail short-wave irregularities significantly influence wheel–rail interaction and rail integrity. In this study, we developed a three-dimensional finite element model that incorporates short-wave irregularities into the wheel–rail contact simulation. The model captures the transient contact forces and the resulting guided wave propagation along the rail. Our analysis shows that the excited signal propagates primarily as a low-frequency longitudinal guided wave mode with a speed of approximately 5263.16 m s −1 and an attenuation rate of 0.4 dB m −1 . Moreover, the characteristic frequency of the guided wave is found to correlate with both the wheel speed and the wavelength of the rail irregularities, demonstrating the potential for estimating irregularity dimensions from guided wave signals. Field experiments validate the fundamental wave propagation and attenuation characteristics, with an average measured attenuation rate of 0.55 dB m −1 . These findings suggest the feasibility of a nonintrusive technique for potential rail monitoring, offering valuable insights for enhancing railway safety and maintenance.
Carbon fiber composites are susceptible to external threats during service due to their linear elastic mechanical behavior and brittle fracture failure characteristics. Therefore, accurate, timely, and effective structural health monitoring of carbon fiber composites has become particularly important. Carbon fiber has good conductivity, making it possible to monitor damage using the electrical potential method. In this study, we set up three different diameter sizes of probes ( ∅ = 5, 10, and 15 mm) and solidified them with plain-woven laminates through co-cured technology. The ∅ = 15 mm probes can achieve accurate measurement of in-situ resistance by using the four-probe method. The sensitivity of the two-dimensional electrical potential method to the damage morphology of plain-woven composites under different current loading modes was analyzed by presetting defects in laminates. It was found that electrical current injection mode affects the accuracy of electrical potential changes. The damage positions in the oblique current injection mode are all near the electrical potential peak and valley. The electrical potential field disturbance caused by damage inside the plain-woven laminates is not significant. Compared to the oblique current injection modes, the accuracy of the damage area obtained by the potential-based damage monitoring in the direct current modes is higher. Under the direct current injection, the maximum electrical potential change rate reaches 200% after damage compared to the intact state. The offset length between the maximum value of the electrical potential change map and the damage area edge is 20 mm, indicating that the potential damage morphology is not completely consistent with the damage area. The electrical potential damage map obtained through the inversion algorithm can locate the damage location. These results provide new ideas for improving the sensitivity of the electrical potential method for structural health monitoring.
The integrity and safety of underwater bridge structures can be compromised by damage; therefore, timely detection and assessment are crucial. However, underwater damage detection is constrained by turbidity, low illumination, and multiple coexisting damage types, which complicates comprehensive automated safety assessment. This study proposes an automated framework for structural damage detection and safety assessment of underwater bridge structures. Multiple types of underwater damage are analyzed, and an underwater damage image dataset (UDID) is established. A modified linear unsharp masking method is used to adaptively enhance the high-frequency features of the UDID through multi-scale image fusion. A well-trained GoogLeNet model is used to automatically detect underwater structural damage. Based on the detection results, bridge safety is classified into five levels using the damage index method. A case study involving a concrete bridge in China demonstrates the effectiveness of the proposed framework. The GoogLeNet model achieves a detection accuracy of 97% in the large-scale test. The concrete bridge is assessed as level 3, which represents moderate damage and is consistent with field detection results. In underwater environments featured by a low signal-to-noise ratio, the detection accuracy of damage types that depend on texture and contrast features decreases significantly. This framework effectively addresses the limitations of existing underwater damage detection methods and enables automated safety assessment for underwater bridge structures.
Harmonic drives are widely used in industrial robots. However, due to the flexspline’s high contact ratio and the elastic deformation, conventional signal processing, such as synchronous averaging, struggle to extract information about flexspline tooth faults. To address this issue, this article proposes an instantaneous angular speed (IAS) signal energy-difference weighting (EDW) method for flexspline fault detection. First, the IAS signal is calculated by the outputs from the built-in encoder, and synchronous averaging is performed to reduce aperiodic components by using the full repeating cycle of the flexspline relative motion as the reference. Second, the energy difference weighting matrix is created using two angular positions separated by one quarter of the input cycle, with higher weights as signed to locations with greater energy variation. Third, a polar energy spectrum in the angular domain is introduced to depict the angular waveform and its energy distribution. The energy map is compared to the installation orientation of the defective flexspline, and the IAS signal’s crest factor (CF) is compared before and after EDW processing. These results support the use of the proposed method to detect flexspline tooth breakage and provide an interpretable representation.
Aiming at the limitations of existing multivariate signal decomposition methods such as fast multivariate empirical mode decomposition (FMEMD) and completely adaptive projection multivariate local characteristic-scale decomposition (CAPMLCD) for gear fault diagnosis, this paper proposes a fast multivariate all-time-scale decomposition (FMATD) method. FMATD incorporates the ATD as its one-dimensional kernel within an efficient “projection-decomposition-reconstruction” framework, preserving the mode separation capability and adaptivity of ATD. Meanwhile, the efficient decomposition framework enhances computational efficiency and avoids over-decomposition. Furthermore, a fast projection strategy is designed to select the projection vectors in real time based on the signal energy distribution, thereby enhancing computational efficiency and decomposition accuracy. Applying FMATD to gear simulation signals and real vibration signals from faulty face gears demonstrates that the proposed method can effectively extract fault modes from face gear signals. Compared with FMEMD, CAPMLCD, and multivariate variational mode decomposition, FMATD yields the component with the clearest fault features in the envelope spectrum. In terms of computational efficiency, FMATD outperforms both FMEMD and CAPMLCD.
Driven by global clean energy strategies, wind power develops rapidly. Bearings, core wind turbine transmission parts, govern system reliability and safety. Conventional diagnosis suffers three key practical limitations: single-sensor signals cannot fully characterize nonlinear composite faults; mainstream deep learning models act as opaque black boxes without clear diagnostic interpretability; highly coupled composite fault features cannot be separately extracted by existing algorithms. To address these challenges, this paper proposes an interpretable multi-sensor bearing nonlinear composite fault diagnosis method for wind power systems. Firstly, a multi-sensor dynamic frequency guided synchronous compressed wavelet transform is designed to precisely extract and unify multi-sensor non-stationary signal features via dynamic frequency matching, adaptive wavelet basis selection, and scale parameter optimization. Secondly, a dynamic calibration and feature enhancement network is constructed, including a dynamic dual-branch calibration fusion module for adaptive feature weighting and decoupling, and a wavelet attention feature enhancement network for sensitive feature enhancement and interpretability improvement. Finally, a Mahalanobis distance aware Krylov Transformer network is developed, integrating Mahalanobis distance to enhance early subtle fault sensitivity and an efficient global enhanced Krylov Transformer for deep feature modeling. Experiments across the three datasets yield average diagnostic accuracies of 98.80, 99.24, and 99.93%, respectively. Even under −4 dB noise interference, the proposed model retains an average accuracy above 90%. Component decoupling verification reveals that 96.0% of composite fault samples can be simultaneously identified via two independent fault channels, with the Pearson correlation coefficient between channel outputs as low as 0.18. Moreover, controlled sub-band masking tests show that masking the HH sub-band containing fault impulse information reduces the overall diagnostic accuracy from 98.45 to 76.89%, corresponding to a 21.56 percentage point drop, this sufficiently proves that the model’s inference relies heavily on high-frequency time–frequency features corresponding to fault impulses.
High-speed cameras have been widely adopted as a non-contact alternative to conventional contact sensors for mechanical condition monitoring. Traditional image spatial filtering faces two challenges: selecting a appropriate spatial frequency parameter, and phase wrapping caused by the arctangent function’s principal-value interval when encoding displacements from phase differences. To overcome these issues, a vision-based phase motion analysis method using a complex-valued steerable pyramid decomposition is presented for measuring bearing vibration signals. By constructing multiple sets of complex-valued filters to extract sequential image phase information at different spatial frequencies and orientations, and then jointly solving the phases, the problems of temporal wrapping and spatial discontinuity are resolved. To handle the unknown coupling relationships in compound faults of rolling bearings, a periodic-reconstruction-enhanced fast nonlinear blind deconvolution method is proposed. A periodic enhancement and reconstruction mechanism is incorporated into the fast nonlinear blind deconvolution algorithm to enhance impulse trains of each fault period, and then defect characteristic frequencies are obtained via envelope demodulation. For the two parameters in the blind deconvolution algorithm that are sensitive to the results, namely the filter length and the deconvolution period, the envelope autocorrelation function is used to analyze the impulse period. Then, under the determined deconvolution period, the optimal filter length is selected using the minimum multi-scale permutation entropy as the criterion. Through experiments, the inner race and outer race defect features were successfully separated and extracted from the vision-measured compound fault signal of the bearing. Additionally, the advantages of the proposed methods were fully validated through different illumination experiments and comparisons with various other methods. This work provides a solution for non-contact condition monitoring.
Accurate three-dimensional (3D) acoustic emission (AE) source localization is fundamental to structural health monitoring (SHM) and damage characterization in concrete structures. However, achieving high-precision localization remains challenging due to the inherent heterogeneity of concrete, complex boundary reflections, and the high cost of obtaining large-scale labeled datasets. This study proposes a novel physics-guided deep learning framework, termed Dual TDOA+PIR, which integrates dual Time Difference of Arrival (TDOA) representations with physics-informed regularization (PIR) for 3D AE localization in concrete prisms. Here, the physical knowledge is incorporated through algebraic and geometric constraints derived from the wave-propagation model, rather than by solving the underlying wave equation. The framework features: (1) a dual-representation strategy that fuses explicit raw TDOA features with Transformer-derived global dependencies from multi-threshold sequences; and (2) a physics-informed regularization scheme that imposes TDOA-consistency and geometric propagation constraints to ensure physical consistency under sparse data conditions. Experimental validation on a 100 × 400 × 100 mm concrete prism demonstrates that the proposed approach achieves a Mean Absolute Error (MAE) of 4.88 mm with an optimal physics weight ( λ = 10 − 5 ). This represents a 56.35% improvement over the K-nearest neighbors baseline (11.18 mm) and a 37.11% improvement over the non-physics-constrained dual-representation model (7.76 mm). Ablation studies further reveal that: (1) three-threshold TDOA extraction optimizes the balance between feature richness and generalization; (2) physics constraints substantially enhance Z-axis (depth) localization by 56.28%; and (3) the framework maintains high robustness across various physics-weight settings. This methodology provides a robust and high-precision solution for real-time monitoring of internal damage in concrete structures.
Cracking and compression damage of concrete face slabs directly influence the leakage condition and operational safety of concrete-faced rockfill dams (CFRDs). During operation, however, underwater face-slab damage is difficult to perceive and localize in a timely manner. Rockfill deformation is a dominant factor affecting the damage state of concrete face slabs. Accordingly, this paper proposes a diagnostic method for underwater damage zones in concrete face slabs of CFRDs by capturing the intrinsic relationship between seepage and deformation observations. First, Random Forest (RF) is combined with SHapley Additive exPlanations (SHAP) to reduce the dimensionality of the influencing factor set for seepage discharge, which is constructed from monitoring data on rockfill settlement, face-slab joint deformation, and environmental loads. A RIME-optimized Light Gradient Boosting Machine model is then developed to establish the seepage discharge–deformation mapping relationship. The model-calculated seepage discharge is subsequently compared with a seepage-discharge control threshold determined by a bootstrap-assisted empirical quantile method to identify abnormal deformation periods. Based on SHAP analysis of the seepage–deformation coupling relationship during these abnormal periods, underwater damage zones in concrete face slabs are diagnosed by identifying key abnormal deformation locations that contribute significantly to abrupt increases in seepage discharge. A case study of an in-service CFRD confirms the effectiveness and feasibility of the proposed method, with the identified underwater damage zones in concrete face slabs showing good agreement with inspection results. These findings highlight the practical value of the proposed method for rapid underwater damage diagnosis and structural health monitoring of CFRDs.
Early diagnosis of wind turbine blade aerodynamic imbalance remains challenging because early fault features are weak and easily obscured by operating-condition fluctuations and environmental noise. To improve structural health monitoring reliability, this study proposes a lightweight wide-spectrum bi-temporal fusion network (L-WS-BFN). Based on generator-speed data from an 8.35 MW wind turbine, a preprocessing pipeline with rated-condition constraints, sample-wise standardization, and online probabilistic augmentation was developed. This pipeline reduces the effects of start-up/shutdown transients, control switching, and outlier noise on model learning. Guided by the rotor aerodynamic-load—drivetrain torsional vibration mechanism, the proposed framework integrates wide-spectrum convolution, bi-temporal fusion, and lightweight decision-making for non-stationary 1P modulation. The Bi-Temporal Fusion Module (BFM) improves amplitude—phase representation of slowly varying 1P disturbances through adjacent-segment comparison and competitive attention fusion. L-WS-BFN achieved 95.4% training accuracy with only 0.01855 M parameters and good class balance in precision, recall, and F1-score. Comparative experiments, ablation studies, network-depth sensitivity analysis, and t-distributed stochastic neighbor embedding (t-SNE) visualization confirmed its noise robustness, generalization ability, and edge-deployment suitability. Two normal-state records from August 2025, collected from the target turbine and another turbine, were further used for external normal-only validation. The results support its seasonal specificity and cross-turbine false-alarm robustness, indicating an efficient and reliable edge-side diagnostic solution for wind turbines.
Distribution alignment is the foundation of cross-domain fault diagnosis and has a direct impact on the transfer diagnostic accuracy. Some representative joint distribution alignment methods reduce fine-grained distribution discrepancies between the source domain and the target domain by employing the maximum mean discrepancy metric and have achieved success in certain aspects. However, these methods neglect the interference of noise signals and do not consider the zero-mean characteristic of mechanical vibration signals, making them insufficient to fully characterize distribution discrepancies, thereby limiting the effectiveness of fine-grained distribution alignment. To this end, a discriminative fine-grained domain confusion (DFDC) framework is proposed in this article to achieve targeted fine-grained alignment. First, a novel reinforced memory discriminative feature extractor is proposed, which can extract fault-discriminative information from monitoring signals under noisy environments while overcoming the catastrophic forgetting problem of gated recurrent units. Then, a new robust fine-grained domain confusion mechanism is developed to enhance the discrepancy representation capability between source domain and target domain. Finally, the designed DFDC framework is validated through two scenarios, demonstrating excellent diagnostic performance.
To address the limitations of Kurtogram in handling high-amplitude impacts in wheel–rail noise and its inability to effectively identify composite faults in the wheelset-bearing system, this study introduces a novel Weight Kurtogram-based multi-demodulation band recognition strategy. This strategy firstly introduces a flexible frequency band division approach based on the fluctuation state of the Fourier spectrum. Additionally, a novel Weight Kurtosis indicator is designed to fully utilize the impulsiveness and periodicity of fault signatures, providing a great immunity to high-amplitude shocks. Furthermore, inspired by the observed multi-layer sub-band clustering in the Weight Kurtogram, a unique multi-resonant frequency band identification strategy is introduced to fully reveal all informative frequency bands. To validate the efficacy of this method, simulations and tests are conducted using real wheelset-bearing system vibration signals. The results indicate that the multi-band demodulation strategy is an effective method for detecting multiple source faults in the wheelset-bearing system.
