Heavy-load train bogie fault diagnosis method based on lightweight data physical fusion

Abstract

The invention discloses a heavy-load train bogie fault diagnosis method based on lightweight data physical fusion, and belongs to the heavy-load train bogie fault diagnosis technology. The method comprises the following steps: acquiring and preprocessing a multi-channel vibration signal of the bogie; constructing a physical prior vector containing time domain and frequency domain features; constructing a diagnosis network taking a G-DSC lightweight time sequence convolution unit as a core, wherein the unit integrates main-auxiliary feature generation and a DCF large-kernel deep convolution structure; a multi-task cooperative training strategy is adopted, fault classification and physical prior regression tasks are jointly optimized, and a weight updating mechanism of physical consistency constraint is introduced in training. According to the method, real-time fault diagnosis with high precision and low calculation cost is realized, and the physical interpretability and cross-working-condition robustness of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of heavy-haul train bogie fault diagnosis technology, and in particular to a lightweight data-physical fusion method for heavy-haul train bogie fault diagnosis. Background Technology

[0002] With the rapid development of railway transportation, especially the promotion of heavy-haul trains, ensuring the safety and reliability of train operation has become a top priority for the railway sector. In heavy-haul trains, the bogie, as a critical component, directly affects the train's safety, stability, and passenger comfort. Bogie failures, especially the failure of key components such as springs and dampers, can lead to major safety accidents such as excessive train vibration and derailment, and in severe cases, even endanger passenger lives.

[0003] However, existing fault diagnosis methods typically have several shortcomings. High computational resource requirements: While traditional deep learning methods can provide high diagnostic accuracy, they usually require substantial computational resources, especially in online diagnosis, where they often struggle to meet real-time requirements. Lack of physical interpretability: Most existing fault diagnosis methods rely heavily on data-driven black-box models, lacking the ability to interpret fault mechanisms and failing to fully consider prior physical knowledge in the engineering field. Challenges in handling complex and variable operating conditions: In practical applications, fault signals from heavy-load train bogies are often influenced by complex factors such as track conditions, train load, and speed. These factors lead to high signal noise and complex fault mode variations, posing significant challenges to fault diagnosis. Summary of the Invention

[0004] To address the problems of high computational resource requirements, lack of physical interpretability, and insufficient generalization ability and robustness under complex and variable operating conditions in existing heavy-haul train bogie fault diagnosis methods, this invention proposes a lightweight data-physical fusion method for heavy-haul train bogie fault diagnosis. Through a lightweight fault diagnosis network and a data-physical integration strategy, the method effectively improves the efficiency, accuracy, and interpretability of fault diagnosis, thus solving the aforementioned problems.

[0005] This application discloses a lightweight data-physical fusion method for diagnosing bogie faults in heavy-haul trains, including: S1. Acquire multi-channel vibration signals of heavy-load train bogies under multiple working conditions, and preprocess the vibration signals to obtain time-series sample data; S2. Extract time-domain features and frequency-domain features from the time-series sample data to obtain physical prior features; S3. Construct a lightweight temporal convolutional unit, including a main-auxiliary feature generation unit and a large kernel deep convolutional context enhancement unit. The output of the main-auxiliary feature generation unit is connected to the input of the large kernel deep convolutional context enhancement unit. The output of the main-auxiliary feature generation unit and the output of the large kernel deep convolutional context enhancement unit are connected through residuals and then point convolution is performed. S4. Construct a lightweight temporal fault diagnosis network, the network comprising an embedding layer, a point convolutional layer, a deep temporal feature extraction module, a global pooling layer and a fully connected layer connected in sequence, the physical prior features being input to the fully connected layer through another branch; S5. A multi-task collaborative training strategy is adopted to train the lightweight temporal fault diagnosis network. The multi-task collaborative training includes optimizing the network parameters using a multi-task loss function, which includes a classification loss function and a physical prior loss function. S6. During training, a digital-analog collaborative weight update mechanism with physical consistency constraints is adopted. Through a dual-path gradient update strategy and a parameter adjustment mechanism driven by physical consistency, the network parameter update is guided to conform to the dynamic laws of the bogie.

[0006] Preferably, obtaining vibration signals in S1 includes: simulating different speeds, loads, and fault states by using the real track spectrum as an excitation input to a multibody dynamics simulation model containing the bogie structure, and generating corresponding multi-channel acceleration time history response signals.

[0007] Preferably, the preprocessing includes filtering and noise reduction of the signal, slicing it into segments of fixed duration, multi-channel alignment, and normalization.

[0008] Preferably, the time-domain features described in S2 include root mean square, peak-to-peak value, skewness, and kurtosis; The frequency domain features include energy spectral density, spectral centroid, bandwidth, and spectral entropy obtained through fast Fourier transform; The extracted time-domain and frequency-domain features are combined to construct a multi-dimensional physical prior feature vector.

[0009] Preferably, the main-auxiliary feature generation unit adopts a dual-path feature generation structure, including a main branch and an auxiliary branch. The main branch generates main features through point convolution, and the auxiliary branch generates auxiliary features through depthwise separable convolution based on the main features output by the main branch. Finally, the outputs of the two paths are fused in the channel dimension.

[0010] Preferably, the large kernel deep convolutional context enhancement unit extracts continuous structural features across time periods for each channel of the input multi-channel temporal features, then compresses the continuous structural features across time periods into a more compact channel representation, and maps them back to the multi-channel output features while maintaining the integrity of the temporal structure.

[0011] Preferably, the deep temporal feature extraction module includes multiple cascaded lightweight temporal convolutional units.

[0012] Preferably, the classification loss function is the cross-entropy loss function, which measures the difference between the predicted category and the true label. The calculation formula is as follows:

[0013] in, For the sample size, Indicates the first Each sample in the true category The predicted probability; The physical prior loss function uses mean squared error to measure the difference between the predicted physical prior vector and the true physical prior vector. The calculation formula is as follows:

[0014] in, For the predicted physical prior vector, For the corresponding true priors, Let be the dimension of the prior vector; The multi-task loss function is as follows:

[0015]

[0016] in, To adjust the relative weights between classification accuracy and physical consistency, the weights increase linearly from 0 to 1 with each training epoch. T represents the current training epoch, while T represents the total number of epochs.

[0017] Preferably, the dual-path gradient update strategy is as follows: the total gradient during training is divided into a data-driven gradient path and a physically consistent gradient path. The data-driven gradient path is used to update the convolutional and deep temporal feature extraction modules of the network, while the physically consistent gradient path is only used to update the master-auxiliary feature generation unit and the large kernel deep convolutional context enhancement unit near the input end.

[0018] Preferably, the parameter adjustment mechanism driven by physical consistency controls the degree of participation of the physical gradient in each unit of the shallow network by setting a dynamic participation factor for the physical consistency gradient. The dynamic participation factor is as follows:

[0019] in, For the current round, The adjustment coefficient is calculated using the following formula:

[0020] in, Base adjustment value, To adjust the rate factor, The physical consistency deviation for the current training round is given by, where, Base adjustment value, To adjust the rate factor, The physical consistency deviation for the current training round. Indicates the first In each training epoch, the physical prior prediction vector output by the physical prior regression branch... This represents the true physical prior vector corresponding to the training sample.

[0021] The beneficial effects of this invention are: (1) By combining physical prior knowledge and lightweighting, the present invention significantly improves the accuracy and robustness of fault diagnosis of bogies for heavy-haul railway trains.

[0022] (2) The present invention significantly reduces computational complexity, can run efficiently in resource-constrained environments, and reduces inference latency, thus meeting the needs of real-time health monitoring.

[0023] (3) The introduction of physical priors in this invention enhances the interpretability of the diagnostic network, provides stronger physical background support, and improves credibility and transparency.

[0024] (4) The GD-TCN proposed in this invention has good generalization ability and can adapt to fault diagnosis under different track, speed and load conditions, and has broad application prospects. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the main-auxiliary feature generation unit structure according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the large kernel depthwise convolutional context enhancement unit structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the lightweight temporal convolutional unit structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a lightweight timing fault diagnosis network structure according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments.

[0027] This application discloses a lightweight data-physical fusion method for diagnosing faults in heavy-haul train bogies, including: S1. Acquire multi-channel vibration signals of heavy-haul train bogies under various working conditions, and preprocess the vibration signals to obtain time-series sample data.

[0028] To construct a fault diagnosis model suitable for heavy-haul railway bogies, it is first necessary to acquire bogie sensor data. Vibration signals from the train under different operating conditions are collected to generate a dataset containing both normal and various fault states. This embodiment uses a combination of "real track spectrum input + SIMPACK dynamic simulation" to generate training data. First, real track irregularity spectra of different track grades are obtained through on-site detection, including key track parameters such as longitudinal irregularities, gauge variations, and directional irregularities. Then, these real track spectra are used as excitation inputs into the SIMPACK multibody dynamics model, which includes structural units such as the car body, frame, wheelsets, primary suspension, secondary suspension, and typical damping elements. The wheel-rail contact force model, commonly used in practical engineering, is used for dynamic solution.

[0029] The simulation process included various typical operating conditions, such as different operating speeds (e.g., 40 km / h, 60 km / h, 80 km / h), different load states (unloaded, half-loaded, fully loaded), and different curve radius sections. To construct a sample set with separable faults, this embodiment simulated multiple representative fault states by adjusting suspension stiffness, damping parameters, axle box side vibration input, and local damage to the frame. Finally, the acceleration time history response of key components such as the frame, axle box, and side frame was recorded to form multi-channel vibration data covering normal conditions and various fault scenarios.

[0030] The multi-channel acceleration signals output by SIMPACK require further processing after format conversion to meet the input requirements of deep learning models. First, the original time-domain signals undergo filtering and noise reduction, including bandpass filtering, mean shift correction, and outlier removal, to eliminate numerical noise generated during simulation and retain effective dynamic characteristics. Then, the continuous vibration signals are sliced ​​into 1-second segments, each containing approximately 200 sampling points, while retaining multi-point channel information to ensure temporal consistency and spatial distribution characteristics of the samples.

[0031] To ensure the comparability of data from different operating conditions during network training, the data from each channel were further normalized to ensure consistent amplitude scales. Simultaneously, a unified time axis was used to align the channels of different sensors, thereby maintaining the complete correlation structure between the temporal features of multiple channels. The resulting samples have a well-structured and sufficiently suppressed noise profile, making them suitable as standard inputs for lightweight convolutional units and the GD-TCN network.

[0032] S2. Extract time-domain and frequency-domain features from time-series sample data to obtain physical prior features.

[0033] In fault diagnosis of heavy-haul train bogies, the characteristics of vibration signals include not only statistical properties but also physical features related to the mechanical structure. Therefore, a physical prior vector incorporating both time-domain and frequency-domain features is constructed to improve the interpretability and robustness of the model.

[0034] To reflect the dynamic response of heavy-duty bogies under track irregularity excitation, this embodiment extracts several engineering-significant time-domain features from the vibration signals obtained from SIMPACK simulations, including root mean square (RMS), peak-to-peak value, skewness, and kurtosis. These features describe the instantaneous energy level, impact intensity, asymmetry, and kurtosis of the signal, and are closely related to physical phenomena such as wheel-rail impact force changes, suspension system flexibility degradation, and local frame damage. The construction of these time-domain features provides the model with important reference values ​​reflecting fault development trends and short-term dynamic behavior.

[0035] To capture the frequency variations of vibration signals, this embodiment performs a Fast Fourier Transform (FFT) on each signal time window to obtain the spectral structure of the vibration response. Based on the dynamic characteristics of heavy-haul trains, four types of frequency domain features are further extracted: energy spectral density, spectral centroid, bandwidth, and spectral entropy. These features can characterize phenomena such as the main response frequency bands of the wheelset-track system, mode changes in the suspension system, and frame vibration mode shifts, and can effectively reflect the impact of different faults on the frequency distribution. Frequency domain features can help the network identify long-term changing spectral patterns during the modeling process, thereby improving the stability of fault identification.

[0036] To introduce dynamic constraints during training, this embodiment combines the aforementioned time-domain and frequency-domain features to construct an 8-dimensional physical prior vector. This vector not only includes the statistical characteristics of the vibration signal but also incorporates key dynamic indicators reflecting the health status of the mechanical structure. The physical prior vector serves as the target output for the auxiliary regression task during network training, enabling the model to simultaneously learn fault categories and their corresponding physical laws. By incorporating the physical prior, this embodiment effectively suppresses overfitting of deep models to noise or random patterns, improving the model's robustness and physical interpretability under different speed, load, and trajectory conditions.

[0037] S3. Construct a lightweight temporal convolution (G-DSC) unit, which includes a main-auxiliary feature generation unit and a large kernel deep convolutional context enhancement unit. The output of the main-auxiliary feature generation unit is connected to the input of the large kernel deep convolutional context enhancement unit. The output of the main-auxiliary feature generation unit and the output of the large kernel deep convolutional context enhancement unit are connected through residuals and then point convolution is performed.

[0038] To meet the requirements of real-time performance and limited onboard computing resources in fault diagnosis of heavy-haul trains, this embodiment constructs a lightweight temporal convolutional unit. By combining a "master-slave" feature generation mechanism with a depthwise separable convolutional structure, and supplementing it with a large kernel depthwise convolution (DCF) temporal enhancement strategy, sufficient feature representation capability is maintained while significantly reducing the number of parameters and computational cost.

[0039] like Figure 1 As shown, the main-auxiliary feature generation unit adopts a dual-path feature generation structure, including a main branch and an auxiliary branch. The main branch generates basic feature channels through point convolutions to express the main information of the input signal; the auxiliary branch generates a set of lightweight derived features based on the output of the main branch using depthwise separable convolutions. The outputs of the two paths are fused along the channel dimension to form a combined structure of "main features + derived features," and channel compression is achieved through subsequent point convolutions. This unit can achieve effective channel expansion with extremely low computational cost, significantly reducing the number of parameters and floating-point computations required for convolution operations.

[0040] Specifically, the main-auxiliary feature generation unit takes multi-channel one-dimensional temporal features as input. First, it fuses and remaps the input features through one-dimensional point convolution to obtain the "main features" that carry the main discriminative information. This process focuses on completing the linear combination across channels and adjusting the channel dimensions, so that the features have stronger representational capabilities while maintaining the same time length. Then, based on the main features, a one-dimensional deep convolution branch is introduced to perform independent local temporal modeling for each channel, thereby extracting "auxiliary features" that are closely related to the bogie's dynamic response, such as impact, vibration periodicity, and local frequency band changes. This branch supplements fine-grained temporal information with low computational cost. Finally, the main features and auxiliary features are fused (e.g., splicing or weighted fusion) to form the output result, so that the output simultaneously contains the global discriminative information after channel fusion and the local dynamic details captured by deep convolution, providing more stable, interpretable, and more robust input features for subsequent temporal networks.

[0041] To enhance the network's ability to model long-term dependencies, this embodiment introduces a large-kernel deep convolution structure into the lightweight temporal convolutional unit. This unit employs a deep convolution with a kernel size of 31 to significantly expand the network's receptive field in the temporal dimension, thereby improving its ability to capture long-range vibration features. Under different track conditions, vehicle speed variations, and complex load conditions, the DCF unit can effectively extract contextual information across time scales, providing a more stable temporal representation basis for fault mode recognition.

[0042] Specifically, such as Figure 2As shown, using multi-channel temporal features as input, the DCF unit effectively captures long-term vibration modes by applying one-dimensional deep convolutions with large receptive fields to each channel. The unit first uses large-kernel convolutions to independently scan within each channel, extracting continuous structural features across time periods from the input sequence. This enables it to identify slow-varying trends and broadband features caused by track irregularities, wheel-rail force variations, or coupled vibrations of the suspension system. Subsequently, the convolutional output is compressed into more compact channel representations, forming small blue feature blocks in the figure, which are then mapped back to the multi-channel output features while maintaining the integrity of the temporal structure. Through this "single-channel large-kernel convolution - feature compression - multi-channel reconstruction" approach, the DCF unit not only enhances the network's ability to perceive long-term dependencies but also significantly reduces computational cost. When processing vibration signals from heavy-load train bogies, it combines high efficiency, stability, and cross-condition feature extraction capabilities, providing physically meaningful deep feature inputs for subsequent temporal fault diagnosis networks.

[0043] This embodiment integrates the aforementioned main-auxiliary feature generation unit with the large-kernel deep convolutional context enhancement unit to form a complete lightweight temporal convolutional unit. Structurally, this unit sequentially performs channel expansion, long temporal dependency extraction, and channel compression, and fuses the input features with the output results through residual connections, thereby significantly improving the stability of feature flow and deep modeling capabilities. As a fundamental module for constructing the overall fault diagnosis network, the G-DSC unit combines lightweight characteristics with strong temporal modeling capabilities, enabling efficient and stable feature extraction in resource-constrained scenarios.

[0044] Specifically, such as Figure 3As shown, the G-DSC unit takes multi-channel temporal features as input. First, it performs channel fusion and local detail extraction on the original features through a "main-auxiliary" feature layer, enabling the input sequence to obtain a dual representation with both global discriminative ability and local dynamic features while maintaining the overall temporal structure. Then, the unit feeds the enhanced features into a large-kernel convolutional layer with a large receptive field to capture slow-changing trends and structural dynamic patterns across time scales. Residual connections are used to directly superimpose the features before and after convolution, thereby improving the stability and cross-condition adaptability of deep features. To further enhance the network's ability to perceive key temporal structures, a DCF module is embedded within the large-kernel convolutional layer. This module utilizes a combination of depthwise convolution, batch normalization, and activation functions to efficiently extract long-term dependent structures under channel-independent conditions and compress them into compact intermediate features. These intermediate features are then mapped back to multi-channel representations through point convolution to enhance semantic interaction between different channels and improve overall feature representation capabilities. Ultimately, the features output by this unit simultaneously contain the basic information preserved by the residual path and the deep dynamic structure extracted by the DCF module, enabling the G-DSC unit to achieve multi-scale temporal modeling at extremely low computational cost and providing more stable, interpretable and physically relevant feature inputs for subsequent network layers.

[0045] S4. Construct a lightweight temporal fault diagnosis network, which includes an embedding layer, a point convolutional layer, a deep temporal feature extraction module, a global pooling layer, and a fully connected layer connected in sequence. Physical prior features are input to the fully connected layer through another branch.

[0046] Based on the lightweight convolutional module built in S3, this embodiment further proposes a lightweight temporal fault diagnosis network. The network uses multiple lightweight temporal convolutional units (G-DSC) as its core structure, and achieves multi-scale temporal feature extraction of bogie vibration signals through deep stacking. Combined with a multi-task learning mechanism, it completes fault classification and physical prior regression tasks with low computational cost.

[0047] Lightweight timing fault diagnosis network structure such as Figure 4As shown, the diagnostic network in this embodiment adopts a hierarchical structure composed of multiple layers of G-DSC units stacked sequentially. This structure can effectively extract features from the multi-source excitation, strong non-stationarity, wide bandwidth response, and cross-timescale coupling characteristics of the vibration signals of heavy-haul train bogies. Each layer of G-DSC unit sequentially executes the processing flow of "main-auxiliary feature generation → large kernel deep convolution enhancement → channel compression and residual fusion," enabling the network to progressively learn the complex temporal structure formed by the combined effects of wheel-rail impact, frame-car body coupled vibration, and suspension system dynamic response at different depths. The shallow G-DSC module at the front end of the network can capture high-frequency transient features related to rapidly changing events such as short-term wheel-rail impact, local structural loosening, and foreign object impact; while the deep modules establish a larger temporal receptive field through large kernel convolution to extract low-frequency or cross-scale long-term temporal dependencies caused by the cumulative effect of track irregularities, frame swaying, and stiffness degradation of damping components. Through this "layer-by-layer deepening" feature extraction mechanism, the network can adaptively learn the dynamic behavior changes of the bogie under different speed, load, and track conditions. Finally, the temporal features extracted from each layer are pooled into a fixed-length feature vector through global average pooling (GAP). This not only compresses redundant information but also preserves the dynamic response patterns that are crucial for fault identification, providing a stable and reliable input for subsequent fault classification and physical prior regression tasks.

[0048] After extracting temporal features from the multi-layer G-DSC unit, the feature vectors obtained through global average pooling (GAP) are input to the fault classification branch. This branch maps the multi-scale dynamic features learned by the network at different depths to specific fault category spaces through a multi-layer fully connected network. Since typical faults of heavy-haul train bogies (such as primary / secondary suspension stiffness degradation, damper failure, frame cracks, and axle box abnormal vibrations) often exhibit identifiable but frequency-coupled dynamic response patterns, the classification branch needs to perform reliable decision mapping based on the high-dimensional fused feature vectors. This classification branch is trained under supervised supervision using the cross-entropy loss function, enabling the network to distinguish subtle vibration differences under different speeds, track grades, load variations, and wheel-rail excitation conditions, thereby achieving high-precision identification of various bogie fault types. Thanks to the lightweight structural design of the front-end G-DSC module, the classification branch significantly reduces the overall model's parameter scale and computational load while maintaining high diagnostic accuracy, allowing it to run in real-time on an onboard embedded processor, meeting the real-time requirements of online safety monitoring for heavy-haul railways.

[0049] To enhance the diagnostic model's ability to characterize the dynamic behavior of heavy-haul train bogies, this embodiment introduces a physical prior regression branch at the network output to predict time-domain and frequency-domain physical features derived from vibration signals. These physical priors typically include root mean square values, peak-to-peak values, spectral centroids, and frequency band energy distributions, reflecting key physical states such as suspension system stiffness changes, damper performance degradation, abnormal wheel-rail forces, and frame vibration mode shifts. By providing explicit engineering constraints to the network, it can actively approximate the real mechanical response mechanism during the feature learning stage. The physical prior regression task and the fault classification task are trained together in a multi-task manner, achieving collaborative optimization through a joint loss function. This mechanism not only enables the network to learn more stable statistical-spectral domain physical laws while identifying fault categories but also effectively improves the model's cross-condition adaptability under different speeds, track irregularity levels, and load levels. Ultimately, the introduction of physical priors gives the extracted deep features higher interpretability and physical consistency, and significantly enhances the model's robustness and generalization ability in complex operating scenarios.

[0050] The lightweight temporal fault diagnosis network generates two types of feature representations at different semantic levels at the output. One type is a classification feature used for fault identification, and the other is a priori feature used to characterize the bogie's dynamic state. The classification feature is obtained by mapping the temporal features output by the deep temporal feature extraction module through a fully connected layer, and is used to perform bogie fault type identification, corresponding to the network's classification task. The priori feature is obtained by the network's regression prediction of physical quantities in the time and frequency domains of the vibration signal, and is used to characterize the energy distribution, spectral structure, and dynamic change characteristics of the vibration response. In addition, during network training, the priori feature not only serves as the output target of the physical prior regression task, but also participates in the feature learning process by introducing physical constraints. This limits the degree to which the network feature representation deviates from the dynamic laws of the bogie, thereby jointly applying physical consistency guidance to the classification task at both the structural and training levels.

[0051] Therefore, the lightweight temporal fault diagnosis network proposed in this embodiment uses multi-channel temporal vibration signals as input. First, it performs channel mapping and local feature extraction on the original signal through embedding layers and point convolutional layers, converting the multidimensional input sequence into a primary feature representation with basic discriminative capabilities. Subsequently, the primary features are fed into a deep temporal feature extraction module composed of multiple G-DSC units connected in series. Through structures such as large-kernel deep convolution, point convolution, and residual connections, it captures cross-scale dynamic patterns and long-term dependencies layer by layer, thereby obtaining a stable temporal representation that can cover different speeds, loads, and orbital conditions. The output of this module is further compressed into a fixed-length high-dimensional vector in a global pooling unit to characterize the overall dynamic characteristics of the sample. Simultaneously, the physical prior features calculated in the preprocessing stage serve as another input path, inputting together with the semantic features generated by the deep network into a fully connected layer to achieve discriminative output of fault categories and regression prediction of physical features. Through this parallel fusion structure of "deep data features + physical prior features", the lightweight temporal fault diagnosis network can balance the data-driven recognition capability and the constraint of physical consistency in classification tasks, thereby achieving stronger generalization performance and interpretability, and realizing accurate identification of the fault status of heavy-haul train bogies.

[0052] S5. A multi-task collaborative training strategy is adopted to train the lightweight temporal fault diagnosis network. The multi-task collaborative training includes optimizing the network parameters using a multi-task loss function, which includes a classification loss function and a physical prior loss function.

[0053] To further improve the diagnostic accuracy and generalization ability of the lightweight temporal convolutional neural network (GD-TCN) based on the master-slave feature generation mechanism in this embodiment, a multi-task collaborative training strategy is adopted, combining fault classification and physical prior regression tasks. Through a weighted loss function, the network can perform balanced optimization between the two tasks, ensuring both improved classification accuracy and maintained physical consistency.

[0054] The classification loss function used is the cross-entropy loss function, which measures the difference between the predicted class and the true label. The calculation formula is as follows:

[0055] in, For the sample size, Indicates the first Each sample in the true category The predicted probability is calculated based on the loss. This loss factor encourages lightweight temporal fault diagnosis networks to output accurate classification results with high confidence.

[0056] The physical prior loss function uses mean squared error (MSE) to measure the difference between the predicted physical prior vector and the true physical prior vector. The calculation formula is as follows:

[0057] in, For the predicted physical prior vector, For the corresponding true priors, denoted as the dimension of the prior vector. This loss term enables the lightweight temporal fault diagnosis network to capture mechanistic statistical-spectral features while optimizing classification accuracy.

[0058] The two loss functions are jointly optimized using weighted coefficients to ensure a balance between the impact of the classification and physics regression tasks during training. During training, the weight of the physical prior is gradually increased to guide the network to learn feature representations that better conform to physical laws. The multi-task loss function is as follows:

[0059]

[0060] in, To adjust the relative weights between classification accuracy and physical consistency, the weights increase linearly from 0 to 1 with each training epoch. T represents the current training epoch, while T represents the total number of epochs.

[0061] S6. During training, a digital-analog collaborative weight update mechanism with physical consistency constraints is adopted. Through a dual-path gradient update strategy and a parameter adjustment mechanism driven by physical consistency, the network parameter update is guided to conform to the dynamic laws of the bogie.

[0062] Based on the multi-task loss function constructed using S5, this embodiment further proposes a mathematical-analog collaborative weight update mechanism with engineering physics guidance. This mechanism combines data-driven feature learning with the dynamic laws of heavy-haul train bogies during training, enabling the network to maintain stable and interpretable feature representations under different operating conditions. This mechanism does not change the loss function itself, but rather constrains and designs the gradient flow direction and parameter update method during its backpropagation phase.

[0063] To address the multi-scale dynamic characteristics of heavy-haul train bogie vibration signals, which are caused by the combined effects of multiple factors such as wheel-rail impact, suspension-coupled vibration, track irregularities, and load variations, this embodiment proposes a dual-path gradient update strategy based on backpropagation. During training, the total gradient is divided into two categories: "data-driven gradient path" and "physically consistent gradient path," which are applied to different layers of the network respectively. Specifically, during the backpropagation phase of network training, the corresponding gradient information is calculated based on the fault classification loss function and the physical prior loss function, respectively, and the objects of action of different gradients are differentiated and controlled during the parameter update phase. The data-driven gradient generated by the fault classification loss function serves as the global gradient signal and participates in the backpropagation update of all trainable parameters in the network; the physically consistent gradient generated by the physical prior loss function is constrained during backpropagation through parameter grouping and gradient masking mechanisms, and only applies to a pre-defined shallow parameter set. The data-driven gradient path updates all convolutional and deep temporal feature extraction modules of the network, enhancing the model's ability to distinguish different fault modes. The physically consistent gradient path only updates the shallow structures near the input, including the main-auxiliary feature generation units and large-kernel deep convolutional units, ensuring that these modules responsible for extracting low-level dynamic features continuously converge in a direction consistent with the dynamic response of suspension, vibration energy distribution, and structural modal changes during training. Through this dual-path gradient update method, the model can simultaneously achieve both "data-driven discriminative ability" and "physically guided feature consistency," thus realizing a collaborative training effect between the data and the model.

[0064] To ensure that the shallow feature extraction process gradually conforms to the dynamic response characteristics of heavy-haul train bogies, this invention introduces a physically consistent parameter adjustment mechanism during the backpropagation stage. This mechanism involves setting a dynamic participation factor for the physically consistent gradient. This controls the degree of physical gradient participation in each unit of the shallow network, enabling the model to gradually strengthen its adherence to physical laws during training.

[0065] This invention designs the adjustment process of the physical consistency gradient to increase progressively with the training progress. This allows the model to be primarily driven by data-driven feature learning in the early stages, while gradually enhancing the guiding role of physical constraints on shallow parameters in the later stages. Specifically, dynamic participation factors... The exponentially enhanced form is defined as follows:

[0066] in, For the current round, This is the adjustment coefficient, used for control. The growth rate of varies with the physical residuals of the training, and is calculated using the following formula:

[0067] in, Base adjustment value, To adjust the rate factor, This represents the physical consistency deviation for the current training round. Indicates the first In each training epoch, the physical prior prediction vector output by the physical prior regression branch of the lightweight temporal fault diagnosis network is used to characterize the time-domain and frequency-domain physical characteristics estimated by the network based on the current input vibration signal. This represents the true physical prior vector corresponding to the training sample. This true physical prior vector is calculated from the vibration signal using a preset time-domain and frequency-domain feature extraction method and serves as a reference benchmark for physical consistency constraints.

[0068] Specifically, when the physical characteristics (such as vibration energy distribution, spectral center, modal changes, etc.) predicted by the lightweight timing fault diagnosis network deviate significantly from the actual physical priors, Increase, make This allows the physical consistency gradient to approach 1 more quickly, thus strengthening its guiding effect on shallow parameters; as the physical deviation gradually decreases, Automatically reduce, The growth rate is slowed down to avoid over-constraining high-level semantic features.

[0069] Through this adaptive adjustment mechanism, the model is primarily data-driven in the early stages of training to learn discriminative fault mode features. In the later stages, it gradually enhances its adherence to the dynamics of heavy-duty bogies, enabling shallow feature representations to better reflect the physical laws governing wheel-rail force changes, suspension system dynamic response, and track excitation characteristics. Ultimately, this mechanism significantly improves the model's physical interpretability and cross-condition robustness, forming a collaborative feature learning system between the mathematical model and the training environment that can automatically adjust according to the training status.

[0070] This application's embodiments, by combining prior physical knowledge and lightweight design, significantly improve the accuracy and robustness of fault diagnosis for heavy-haul railway train bogies, achieving a diagnostic accuracy of up to 97.67%. Compared to traditional deep learning models, GD-TCN drastically reduces computational complexity, requiring only 31K parameters and 14.68M FLOPs, enabling efficient operation in resource-constrained environments. Furthermore, its inference latency is only 0.14 seconds, meeting the requirements of real-time health monitoring. In addition, the introduction of prior physical knowledge enhances the model's interpretability, providing stronger physical context support and improving its credibility and transparency. GD-TCN also exhibits good generalization ability, adapting to fault diagnosis under different track, speed, and load conditions, and has broad application prospects.

[0071] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A lightweight data-physical fusion method for fault diagnosis of heavy-haul train bogies, characterized in that, include: S1. Acquire multi-channel vibration signals of heavy-load train bogies under multiple working conditions, and preprocess the vibration signals to obtain time-series sample data; S2. Extract time-domain features and frequency-domain features from the time-series sample data to obtain physical prior features; S3. Construct a lightweight temporal convolutional unit, including a main-auxiliary feature generation unit and a large-kernel deep convolutional context enhancement unit. The output of the main-auxiliary feature generation unit is connected to the input of the large-kernel deep convolutional context enhancement unit. The output of the main-auxiliary feature generation unit and the output of the large-kernel deep convolutional context enhancement unit are connected through residuals and then point convolution is performed. S4. Construct a lightweight temporal fault diagnosis network, the network comprising an embedding layer, a point convolutional layer, a deep temporal feature extraction module, a global pooling layer and a fully connected layer connected in sequence, the physical prior features being input to the fully connected layer through another branch; S5. A multi-task collaborative training strategy is adopted to train the lightweight temporal fault diagnosis network. The multi-task collaborative training includes optimizing the network parameters using a multi-task loss function, which includes a classification loss function and a physical prior loss function. S6. During training, a digital-analog collaborative weight update mechanism with physical consistency constraints is adopted. Through a dual-path gradient update strategy and a parameter adjustment mechanism driven by physical consistency, the network parameter update is guided to conform to the dynamic laws of the bogie.

2. The lightweight data-physical fusion-based method for diagnosing faults in heavy-haul train bogies according to claim 1, characterized in that, The vibration signal acquisition in S1 includes: simulating different speeds, loads and fault states by using the real track spectrum as an excitation input to a multibody dynamics simulation model containing the bogie structure, generating corresponding multi-channel acceleration time history response signals.

3. The lightweight data-physical fusion-based method for diagnosing faults in heavy-haul train bogies according to claim 2, characterized in that, The preprocessing includes filtering and noise reduction of the signal, slicing it into segments of fixed duration, multi-channel alignment, and normalization.

4. The lightweight data-physical fusion method for fault diagnosis of heavy-haul train bogies according to claim 3, characterized in that, The time-domain features described in S2 include root mean square, peak-to-peak value, skewness, and kurtosis; The frequency domain features include energy spectral density, spectral centroid, bandwidth, and spectral entropy obtained through fast Fourier transform; The extracted time-domain and frequency-domain features are combined to construct a multi-dimensional physical prior feature vector.

5. The lightweight data-physical fusion method for fault diagnosis of heavy-haul train bogies according to claim 4, characterized in that, The main-auxiliary feature generation unit adopts a dual-path feature generation structure, including a main branch and an auxiliary branch. The main branch generates the main features through point convolution, and the auxiliary branch generates auxiliary features through depthwise separable convolution based on the main features output by the main branch. Finally, the outputs of the two paths are fused in the channel dimension.

6. The lightweight data-physical fusion method for fault diagnosis of heavy-haul train bogies according to claim 5, characterized in that, The large kernel deep convolutional context enhancement unit extracts continuous structural features across time periods for each channel of the input multi-channel temporal features, then compresses the continuous structural features across time periods into a more compact channel representation, and maps them back to the multi-channel output features while maintaining the integrity of the temporal structure.

7. The lightweight data-physical fusion-based method for diagnosing faults in heavy-haul train bogies according to claim 6, characterized in that, The deep temporal feature extraction module includes multiple cascaded lightweight temporal convolutional units.

8. The lightweight data-physical fusion method for fault diagnosis of heavy-haul train bogies according to claim 7, characterized in that, The classification loss function uses the cross-entropy loss function to measure the difference between the predicted category and the true label. The calculation formula is as follows: in, For the sample size, Indicates the first Each sample in the true category The predicted probability; The physical prior loss function uses mean squared error to measure the difference between the predicted physical prior vector and the true physical prior vector. The calculation formula is as follows: in, For the predicted physical prior vector, For the corresponding true priors, Let be the dimension of the prior vector; The multi-task loss function is as follows: in, To adjust the relative weights between classification accuracy and physical consistency, the weights increase linearly from 0 to 1 with each training epoch. T represents the current training epoch, while T represents the total number of epochs.

9. The lightweight data-physical fusion method for fault diagnosis of heavy-haul train bogies according to claim 8, characterized in that, The dual-path gradient update strategy is as follows: the total gradient during training is divided into a data-driven gradient path and a physically consistent gradient path. The data-driven gradient path is used to update the convolutional and deep temporal feature extraction modules of the network, while the physically consistent gradient path is only used to update the master-auxiliary feature generation unit and the large kernel deep convolutional context enhancement unit near the input end.

10. The lightweight data-physical fusion method for fault diagnosis of heavy-haul train bogies according to claim 9, characterized in that, The physical consistency-driven parameter adjustment mechanism controls the degree of participation of the physical gradient in each unit of the shallow network by setting a dynamic participation factor for the physical consistency gradient. The dynamic participation factor is as follows: in, For the current round, The adjustment coefficient is calculated using the following formula: in, Base adjustment value, To adjust the rate factor, The physical consistency deviation for the current training round. Indicates the first In each training epoch, the physical prior prediction vector output by the physical prior regression branch... This represents the true physical prior vector corresponding to the training sample.

Images