High-speed train bearing cross-domain fault diagnosis method based on domain adversarial neural network

Abstract

The invention discloses a high-speed train bearing cross-domain fault diagnosis method based on a domain adversarial neural network, and the method comprises the steps: S1, collecting vibration signals of a source domain rack and a target domain real vehicle bearing, and carrying out the preprocessing of the vibration signals; s2, extracting signal multi-view features, performing joint training PCA dimension reduction and redundancy reduction, and constructing a migratable feature set; s3, screening a sample subset most similar to the distribution of the target domain from the source domain based on the maximum mean value difference; s4, combining the screened source domain sample and the target domain training signal, and obtaining a domain invariant fault diagnosis model through domain adversarial neural network training; s5, performing preprocessing and feature extraction on a to-be-diagnosed real vehicle signal, and inputting the signal into the model; according to the method, final diagnosis is output based on multi-segment result majority voting, high-precision cross-domain recognition of the high-speed train bearing fault is achieved, diagnosis precision, robustness and interpretability are achieved, and the method is suitable for online monitoring and intelligent operation and maintenance of high-speed train running components.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis and intelligent operation and maintenance technology, specifically involving a cross-domain fault diagnosis method for high-speed train bearings based on domain adversarial neural networks. Background Technology

[0002] High-speed trains are the core equipment of my country's passenger transportation system, and their operational safety is of paramount importance. As key components bearing the train's load and performing rotational motion, running gear bearings operate under harsh conditions of high speed, strong impact, and complex variable loads, making them high-risk components prone to failure. The health of the bearings directly affects the train's operational safety and efficiency. Therefore, developing accurate and reliable intelligent bearing fault diagnosis technology tailored to real-world operating environments has significant engineering value and practical implications.

[0003] Existing bearing fault diagnosis methods can be mainly divided into two categories: one is based on manual feature extraction and traditional machine learning, and the other is based on deep learning-based end-to-end diagnostic methods. The former relies on signal processing techniques and expert experience, constructing time-domain, frequency-domain, or time-frequency-domain features and combining them with classifiers such as support vector machines for state recognition. This type of method performs well under stable operating conditions, but its feature representation ability and model generalization face severe challenges under the complex noise interference and load fluctuations of real-world vehicles. The latter utilizes models such as convolutional neural networks to automatically learn fault features, reducing reliance on human experience to some extent, but its performance depends on the training and testing data following an independent and identically distributed distribution. This is often difficult to meet in engineering practice.

[0004] The core challenge lies in the fact that while controlled bench tests in the laboratory provide a readily available and comprehensive database of various bearing fault data (source domain), the bearing vibration data collected in the actual operation environment of high-speed trains (target domain) is not only scarce in labels but also significantly different from the clean bench data due to the influence of complex factors such as environmental vibration, track excitation, and sudden load changes. This difference in domain distribution leads to severe performance degradation, i.e., model generalization failure, when high-precision models trained on bench data are directly applied to real-world trains.

[0005] While traditional transfer learning methods attempt to mitigate this problem through feature alignment or sample reweighting, they often struggle to accurately learn domain-invariant robust features that can distinguish fault modes and are insensitive to domain changes in high-dimensional, nonlinear fault feature spaces, and also carry the risk of negative transfer. Furthermore, most deep transfer learning models are like "black boxes," lacking a connection between their decision-making process and physical mechanisms, making their diagnostic conclusions difficult for on-site operations personnel to understand and trust, thus hindering the practical application of the technology in security-critical scenarios. Summary of the Invention

[0006] This invention addresses the problems in high-speed train bearing fault diagnosis, such as large differences in cross-operating conditions, scarce target domain labels, and lack of interpretability of diagnostic results. It proposes a cross-domain intelligent diagnostic method based on domain adversarial neural networks. This method aims to improve the accuracy and robustness of cross-domain fault identification and enhance the interpretability of the diagnostic process, providing a scientific basis for the intelligent operation and maintenance of high-speed train bearings.

[0007] To achieve the above-mentioned objectives, the present invention employs the following technical solution, comprising the following steps:

[0008] S1. Collect the bearing vibration signal from the source domain bench test and the bearing vibration signal from the target domain actual vehicle operation, and preprocess the signals to obtain the source domain signal and the target domain training signal respectively.

[0009] S2. Based on the source domain signal and target domain training signal preprocessed in S1, extract multi-view features in the time domain, frequency domain, and time-frequency domain; jointly train the principal component analysis model based on the source domain signal and target domain training signal, reduce the dimensionality of the multi-view features, and remove redundant features by correlation reduction to form a transferable feature set.

[0010] S3. Based on the feature set obtained in S2, calculate the distribution similarity by the maximum mean difference, and select the sample subset that is most similar to the distribution of the target domain from the source domain;

[0011] S4. Based on the source domain sample subset filtered by S3 and the target domain training signal preprocessed by S1, the model is trained through a domain adversarial neural network containing a gradient inversion layer and a domain discriminator, so that the model learns the domain-invariant shared features and thus obtains the fault diagnosis model.

[0012] S5. Collect the vibration signal of the actual vehicle bearing in the target domain to be diagnosed and perform the preprocessing described in S1; use the feature processing model trained in S2 to extract its features and input them into the fault diagnosis model trained in S4 to obtain preliminary diagnosis results; based on the diagnosis results of multiple signal segments of the same bearing, generate the final fault diagnosis result through majority voting and output it.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] This invention eliminates the fundamental differences between the source and target domains in terms of sampling and baseline through a unified signal preprocessing process, and uses a domain adversarial neural network for feature learning. Through adversarial training with a gradient inversion layer, the model is forced to learn domain-invariant features that are sensitive to faults but insensitive to domain changes. Thus, even when the target domain lacks labels, it can still identify and locate bearing faults with high accuracy, effectively solving the model generalization problem caused by domain offset.

[0015] This invention innovatively employs a dual strategy combining feature-level screening and sample-level screening. At the feature level, a high-quality, low-redundancy transferable feature set is constructed by combining PCA dimensionality reduction and correlation analysis. At the sample level, the mechanism of using the maximum mean difference to select a subset of source domain samples that are most similar to the target domain distribution for training eliminates source domain data and redundant features with excessively large distribution differences that may cause interference from the source, significantly reducing the risk of negative transfer and ensuring the effectiveness and stability of the transfer process.

[0016] The shared multilayer sensor tag classifier designed in this invention can simultaneously determine both fault type and defect size. This not only enables qualitative fault identification but also quantitative assessment of fault severity, providing more granular and instructive information support for predictive maintenance and repair strategy development for trains.

[0017] This invention constructs a complete interpretability framework through feature distribution visualization, domain discriminator index quantification evaluation, and feature contribution analysis. This method can trace the model's diagnostic decisions back to specific time-domain, frequency-domain, or time-frequency-domain features and compare them with known bearing fault characteristic frequencies, making the diagnostic results verifiable and effectively alleviating the "black box" problem of deep models. This significantly improves the credibility and acceptability of diagnostic conclusions in engineering practice.

[0018] This invention relies solely on widely used acceleration vibration signals, and its preprocessing steps and network structure parameters can be flexibly adjusted according to actual data conditions. The solution clearly defines the hardware and software environment for system operation and provides technical indicators supporting high concurrency and long-term stable operation, demonstrating that this technical solution possesses good engineering potential and industrialization prospects. It is easily integrated with existing onboard health monitoring systems to achieve online intelligent monitoring and maintenance decision-making for high-speed train bearings. Attached Figure Description

[0019] To make the technical solutions in the embodiments of the present invention or the prior art clearer, the accompanying drawings referenced in the embodiments or the prior art are now briefly described. It should be understood that these drawings are only used to assist in illustrating some embodiments of the present invention, and those skilled in the art can obtain other forms of drawings based on these drawings without creative effort.

[0020] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In this embodiment, the high-speed train bearing cross-domain fault diagnosis method based on domain adversarial neural networks, such as... Figure 1 As shown, the procedure is as follows:

[0023] S1. Collect bearing vibration signals from the source domain bench test and bearing vibration signals from the target domain actual vehicle operation; preprocess the signals to obtain source domain signals and target domain training signals respectively; specifically:

[0024] S1.1. Collect bearing vibration signals from the source domain bench test platform under different speeds and loads, as well as bearing vibration signals collected during the actual operation of the high-speed train in the target domain.

[0025] S1.2. Select the sensor channel according to the fault location. For the fault of the drive end bearing, select the drive end acceleration channel first. For the fault of the fan end bearing, select the fan end acceleration channel first. When there is no exact location marking, select according to the priority of "drive end > fan end > bearing end".

[0026] S1.3 Perform detrending processing on the vibration signals of each channel. Obtain the detrending signal by fitting the signal using the least squares method and removing linear or slowly varying baselines.

[0027] S1.4 For source domain signals with sampling rates of 12kHz or 48kHz, frequency domain reconstruction is used to resample to a uniform sampling rate of 32kHz. Frequency domain zero-filling is performed on low sampling rate signals, band-limited truncation is performed on high sampling rate signals, and time domain waveforms are reconstructed through inverse fast Fourier transform.

[0028] S1.5 Based on the high-frequency energy ratio, the noise intensity is estimated, and the passband and order of the Butterworth bandpass filter are adaptively designed. Zero-phase forward and backward filtering is used to suppress environmental noise and structural noise. The segmented features are then subjected to robust normalization based on the median and the absolute deviation of the median to obtain the source and target domain signals after unified preprocessing.

[0029] The adaptive bandpass filtering described in S1 includes: dividing the noise intensity into three levels—high noise, medium noise, and low noise—based on the proportion of high-frequency energy. In the high noise level, the bandpass is set to 80–6000 Hz and a 6th-order Butterworth filter is used. In the medium noise level, it is set to 60–7200 Hz and a 5th-order filter is used. In the low noise level, it is set to 50–8000 Hz and a 4th-order filter is used. All of these settings employ forward and backward zero-phase filtering.

[0030] S2. Based on the source domain signal and target domain training signal preprocessed in S1, extract multi-view features in the time domain, frequency domain, and time-frequency domain; jointly train a principal component analysis model based on the source domain signal and target domain training signal, reduce the dimensionality of the multi-view features, and remove redundant features through correlation reduction to form a transferable feature set; specifically:

[0031] S2.1. The detrended, resampled, and denoised vibration signal is segmented by sliding window according to the preset window length and overlap rate to obtain segment-level signal segments;

[0032] S2.2 Calculate the time-domain characteristics for each signal segment, including:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042] S2.3. Perform Fast Fourier Transform and power spectrum estimation on each signal segment to extract frequency domain features, including dominant frequency, peak frequency, peak amplitude, total energy, frequency centroid, frequency variance, and mechanistic frequency neighborhood energies such as BPFI, BPFO, BSF, and FTF corresponding to the bearing structural parameters.

[0043]

[0044]

[0045]

[0046]

[0047] S2.4. Use short-time Fourier transform to generate time-frequency diagrams, and extract time-frequency domain features such as maximum amplitude and total energy for each time segment. The mathematical form is as follows:

[0048]

[0049] S2.5. The time-domain, frequency-domain, and time-frequency-domain features are concatenated into a high-dimensional handcrafted feature vector. Principal component analysis is then performed on the source and target domain samples to reduce dimensionality, compressing them to a preset dimension while ensuring the cumulative variance contribution rate. The principal component analysis dimensionality reduction involves concatenating the source and target domain samples in the sample dimension and learning the same linear subspace. The samples from both domains are then simultaneously projected onto this subspace to improve the stability of cross-domain feature similarity measurement and subsequent maximum mean difference estimation.

[0050] S2.6 Calculate the Spearman rank correlation coefficient matrix among the dimensionality-reduced features, remove highly correlated redundant features, and retain representative features that are highly correlated with the fault mechanism but have low correlation, forming a high-quality transferable feature set. The Spearman correlation coefficients are:

[0051]

[0052] S3. Based on the feature set obtained in S2, calculate the distribution similarity using the maximum mean difference, and select the subset of samples from the source domain that is most similar to the distribution of the target domain; specifically:

[0053] S3.1. In the unified PCA subspace, a radial basis function kernel is used to construct the kernel space, mapping the samples from the source and target domains to the reproducing kernel Hilbert space. The mathematical form of the radial basis function kernel is:

[0054]

[0055] S3.2 Calculate the distribution similarity between each source domain signal segment and all target domain signal segments using the maximum mean difference index, and obtain the similarity vector of the source domain samples:

[0056]

[0057] S3.3 Sort the source domain samples in descending order of similarity. Based on the preset retention ratio threshold, select the high-quality source domain samples that are most similar to the target domain distribution to form a set of transfer training source domains, so as to suppress the negative transfer caused by samples with large distribution differences.

[0058] In the source domain sample selection process based on the maximum mean difference, a radial basis function kernel is used to construct the kernel function. It adaptively adjusts parameters based on the overall noise level difference between the source and target domains. This is to mitigate the estimation bias caused by kernel function saturation in high-dimensional feature spaces. The parameter γ can be adaptively adjusted based on the overall distribution differences (such as noise levels) between the source and target domain data to optimize the measurement of distribution similarity.

[0059] S4. Based on the source domain sample subset filtered in S3 and the target domain training signal preprocessed in S1, a domain adversarial neural network containing a gradient inversion layer and a domain discriminator is trained to learn domain-invariant shared features, thereby obtaining a fault diagnosis model; specifically:

[0060] S4.1 Construct a one-dimensional convolutional neural network feature extractor. The preprocessed source and target domain signal segments are used as inputs. The signal segments are processed through multiple layers of one-dimensional convolution, nonlinear activation and pooling operations. Finally, a fixed-dimensional fault representation vector is obtained through adaptive pooling and L2 normalization.

[0061] S4.2 Construct a shared multilayer perceptron as a label classifier on the representation vector. The output of the shared layer is connected to the fault type classification head and the defect size classification head respectively, so as to realize the joint discrimination of normal, inner ring fault, outer ring fault and rolling element fault as well as each defect size. The size classification head only participates in the loss calculation on faulty samples.

[0062] S4.3. A gradient inversion layer is connected to the output of the feature extractor and a domain discriminator is connected to perform binary classification of the source domain and target domain samples. The gradient inversion layer inverts the gradient sign of the domain discriminator during backpropagation to achieve adversarial feature learning.

[0063] S4.4. A classification loss is constructed using labeled samples from the source domain, and a domain discriminant loss is constructed using samples from both the source and target domains. The classification loss and the domain loss are weighted and summed according to preset weights to obtain the total loss. An adaptive learning rate optimization algorithm is used to jointly train the parameters of the feature extractor, label classifier, and domain discriminator, enabling the network to learn a shared feature space that combines fault separability and domain invariance. The classification loss is as follows: , softmax The size branch only accounts for the loss for abnormal class samples: Total category loss: .

[0064] The one-dimensional convolutional neural network feature extractor includes four convolutional layers, each with a kernel length of 9. Padding is used to maintain the sequence length. The first three convolutional layers are followed by a max-pooling layer to compress the time dimension proportionally. The last convolutional layer is followed by an adaptive max-pooling layer, mapping sequences of arbitrary length to a single time step output. The shared multilayer perceptron in S4 includes two fully connected layers with dimensions of 128→256→128. Dropout is applied between the two layers and at the output to suppress overfitting. The output of the shared multilayer perceptron is simultaneously connected to a fault type classification head and a defect size classification head. Both classification heads use linear layers to output four-dimensional logits, corresponding to normal states and three fault types or four defect sizes. The domain adversarial training in S4 is implemented using a gradient inversion layer. During forward propagation, the input to the domain discriminator remains unchanged, while during backward propagation, the gradient of the domain discriminator is multiplied by a negative value and scaled, achieving adversarial optimization between the feature extractor and the domain discriminator. The domain loss weight gradually increases with the training progress according to an increasing function.

[0065] S5. Collect the vibration signal of the actual vehicle bearing in the target domain to be diagnosed and perform the preprocessing described in S1; extract its features using the feature processing model trained in S2 and input it into the fault diagnosis model trained in S4 to obtain preliminary diagnosis results; based on the diagnosis results of multiple signal segments of the same bearing, generate the final fault diagnosis result through majority voting and output it, specifically as follows:

[0066] S5.1 Perform the same preprocessing and segmentation on the unlabeled bearing vibration signal in the target domain according to the process in steps 1 and 2.

[0067] S5.2 Utilize the feature extractor and label classifier trained by domain adversarial training to output the predicted probability of fault type and defect size for each target domain signal segment, and use file-level majority voting or confidence weighting rules to generate the final fault diagnosis result of the target bearing.

[0068] S5.3 Output the diagnostic results and their confidence levels to the train maintenance system for real-time monitoring, health assessment, and maintenance decisions. The target domain diagnosis employs a "segment-by-segment discrimination + file-level majority voting" method. It statistically analyzes the prediction results of all signal segments belonging to the same bearing file. When the proportion of segments or the average confidence level of a certain type of label exceeds a preset threshold, that type of label is taken as the final diagnostic result for that bearing.

[0069] S6. Based on the shared feature space and diagnostic model trained in S4, the reliability and consistency of the diagnostic results with the underlying mechanism are verified through distribution visualization, domain discriminator performance evaluation, and feature contribution analysis; specifically including:

[0070] S6.1 Before and after transfer training, visualize the distribution of source and target domain samples in the shared feature space using t-SNE or PCA to compare the inter-class clustering and the degree of overlap between the source and target domains.

[0071] S6.2 Calculate the receiver operating characteristic area (AUC) of the domain discriminator during the testing phase. When the AUC is close to 0.5, it is determined that the feature distributions of the source domain and the target domain are sufficiently aligned.

[0072] S6.3. Using feature contribution analysis methods such as SHAP, the diagnostic results of the target domain samples are interpreted post-hocly. The positive and negative contributions of each feature to the prediction results are quantified at the time segment or frequency interval level, and compared and verified with the fault mechanism characteristics of bearing inner ring, outer ring, rolling elements, etc., to form a traceable diagnostic basis.

[0073] The method of this invention can be summarized as follows:

[0074] 1. Design a unified signal preprocessing workflow to perform channel optimization, detrending processing, frequency domain resampling, and adaptive bandpass filtering on source domain bench data and target domain vehicle data to eliminate sampling differences and low-frequency drift, and improve the comparability of features.

[0075] 2. Multi-perspective handcrafted features are constructed from the three perspectives of time domain, frequency domain, and time-frequency domain. Dimensional compression and redundant feature removal are achieved through PCA and Spearman rank correlation coefficient, resulting in a feature set that is closely related to the bearing failure mechanism and has good transferability.

[0076] 3. Using the maximum mean difference index based on the RBF kernel, the distribution similarity between the source domain and the target domain samples is calculated in a unified subspace. High-quality source domain samples that are more similar to the target domain are selected to reduce the negative transfer caused by strong mismatch samples between domains.

[0077] 4. Construct a domain adversarial neural network consisting of a one-dimensional convolutional neural network feature extractor, a shared multilayer perceptron label classifier, and a domain discriminator. Use a gradient inversion layer to achieve adversarial training between the feature extractor and the domain discriminator, so that the model can learn domain-invariant features that have discriminative power for both the source and target domains while maintaining the source domain fault classification performance.

[0078] 5. In the diagnostic output stage, the final fault type and defect size of the target domain bearing are determined by combining segmented prediction and document-level majority voting strategies. The migration process and decision criteria are quantitatively explained through t-SNE visualization, domain discriminator AUC and SHAP feature contribution analysis.

[0079] In this invention

[0080] PCA is principal component analysis, MMD is maximum mean difference, DANN is domain adversarial neural network, AUC is area under the receiver operating characteristic curve, SHAP is SHapley Additive exPlanations, t-SNE is t-distributed random neighborhood embedding, ReLU is corrected linear unit, Dropout is random deactivation, and logits is the unnormalized prediction score. All of these are existing technologies.

[0081] The effects of this invention are as follows:

[0082] 1. Strong cross-domain robustness: Through unified preprocessing and domain adversarial training, this invention can effectively alleviate the distribution differences between bench and real vehicle data in terms of sampling rate, noise intensity and operating conditions, and can still achieve high fault identification accuracy even when the target domain lacks labels.

[0083] 2. Fully utilize source domain information and suppress negative transfer: By combining PCA and MMD in the source domain sample selection strategy, this invention selects only high-quality source domain samples with distributions similar to the target domain to participate in transfer training, thereby weakening the influence of severely mismatched samples and improving the stability and effectiveness of the transfer process.

[0084] 3. Multi-task joint identification: Simultaneously completes multi-task identification of fault type and defect size in a shared feature space, which improves the model's ability to distinguish different fault modes and severity, and provides more granular information support for subsequent maintenance strategy formulation.

[0085] 4. High interpretability: Through structured design, feature space visualization and SHAP feature contribution analysis, this invention can provide criteria corresponding to the actual fault mechanism at the time domain, frequency domain or time-frequency domain feature level, alleviate the "black box" problem of deep models and improve the credibility of diagnostic results.

[0086] 5. Engineering Deployment Friendly: This invention is based solely on acceleration vibration signals. The preprocessing and network structure parameters can be flexibly adjusted according to actual sampling conditions, making it easy to integrate with existing on-board monitoring systems to achieve online health monitoring and intelligent operation and maintenance decisions for high-speed train bearings.

[0087] The present invention also protects an electronic device comprising: a processor coupled to a memory; the memory being used to store a computer program; and the processor being used to invoke the computer program stored in the memory to implement the above-described method.

[0088] A readable storage medium includes a program or instructions that, when executed on a computer, cause the above-described method to be performed. Furthermore, the system's hardware environment meets the requirements of a CPU with ≥4 cores, ≥4GB of memory, and ≥32GB of storage. The software supports Windows, Linux, and the open-source Euler system. Testing solutions, user manuals, and technical training are provided. This system supports concurrent management of 300 devices and can run continuously for 30*24 hours without any abnormalities under stress testing.

[0089] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this document.

[0090] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0091] In the embodiments provided herein, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, apparatuses, or units, or they may be electrical, mechanical, or other forms of connection.

[0092] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments described herein, depending on actual needs.

[0093] Furthermore, the functional units in the various embodiments of this document can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0094] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this paper, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this paper. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0096] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks, characterized in that, Includes the following steps: S1. Collect the bearing vibration signal from the source domain bench test and the bearing vibration signal from the target domain actual vehicle operation, and preprocess the signals to obtain the source domain signal and the target domain training signal respectively. S2. Based on the source domain signal and target domain training signal preprocessed in S1, extract multi-view features in the time domain, frequency domain, and time-frequency domain. Based on the joint training of the source domain signal and the target domain training signal, the principal component analysis model is trained, the multi-view features are dimensionality reduced, and redundant features are eliminated by correlation reduction to form a transferable feature set. S3. Based on the feature set obtained in S2, calculate the distribution similarity by the maximum mean difference, and select the sample subset that is most similar to the distribution of the target domain from the source domain; S4. Based on the source domain sample subset filtered by S3 and the target domain training signal preprocessed by S1, the model is trained through a domain adversarial neural network containing a gradient inversion layer and a domain discriminator, so that the model learns the domain-invariant shared features and thus obtains the fault diagnosis model. S5. Collect the vibration signal of the actual vehicle bearing in the target domain to be diagnosed and perform the preprocessing described in S1; use the feature processing model trained in S2 to extract its features and input them into the fault diagnosis model trained in S4 to obtain preliminary diagnosis results; based on the diagnosis results of multiple signal segments of the same bearing, generate the final fault diagnosis result through majority voting and output it.

2. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 1, characterized in that, Also includes: S6. Based on the shared feature space and diagnostic model trained in S4, the reliability and consistency of the diagnostic results with the underlying mechanism are verified through distribution visualization, domain discriminator performance evaluation, and feature contribution analysis; specifically including: S6.1 Before and after transfer training, visualize the distribution of source and target domain samples in the shared feature space using t-SNE or PCA to compare the inter-class clustering and the degree of overlap between the source and target domains. S6.2 Calculate the area of ​​the receiver operating feature curve of the domain discriminator during the testing phase. When the area of ​​the operating feature curve is close to 0.5, it is determined that the feature distributions of the source domain and the target domain are fully aligned. S6.

3. Using the SHAP feature contribution analysis method, the diagnostic results of the target domain samples are interpreted post-hocly. The positive and negative contributions of each feature to the prediction results are quantified at the time segment or frequency interval level, and compared and verified with the failure mechanism characteristics of the bearing inner ring, outer ring, and rolling element to form a traceable diagnostic basis.

3. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 1, characterized in that, The preprocessing includes detrending, resampling, and noise reduction.

4. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 1, characterized in that, S2 specifically includes: S2.

1. The preprocessed signal in S1 is segmented by sliding window according to the preset window length and overlap rate to obtain segment-level signal segments; S2.2 Calculate the time-domain characteristics for each signal segment, including: Mean: Standard deviation: Root mean square: Peak value: Skewness: kurtosis: Peak factor: Pulse factor: Margin factor: S2.

3. Perform Fast Fourier Transform and power spectrum estimation on each signal segment to extract frequency domain features, including dominant frequency, peak frequency, peak amplitude, total energy, frequency centroid, frequency variance, and the frequency neighborhood energy corresponding to the bearing structural parameters (BPFI, BPFO, BSF, FTF). Outer ring failure frequency: Inner ring failure frequency: Rolling element failure frequency: Cage failure frequency: S2.

4. Use short-time Fourier transform to generate time-frequency diagrams, and extract the time-frequency domain features of the maximum amplitude and total energy of each time segment. The mathematical form is as follows: S2.

5. The above time domain, frequency domain and time-frequency domain features are concatenated into a high-dimensional handmade feature vector. Principal component analysis is performed on the source domain and target domain samples to reduce the dimensionality, compressing it to the preset dimension while ensuring the cumulative variance contribution rate. S2.6 Calculate the Spearman rank correlation coefficient matrix among the features after dimensionality reduction, remove highly correlated redundant features, and retain representative features that are highly correlated with the fault mechanism but have low correlation, forming a high-quality transferable feature set; where the Spearman correlation coefficient is:

5. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 4, characterized in that, Principal component analysis dimensionality reduction involves concatenating the source and target domain samples along the sample dimension to learn the same linear subspace, and then simultaneously projecting the samples from both domains into this subspace.

6. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 1, characterized in that, S3 specifically includes: S3.

1. In the unified PCA subspace, a radial basis function kernel is used to construct the kernel space, mapping the samples from the source and target domains to the reproducing kernel Hilbert space. The mathematical form of the radial basis function kernel is: K(x,y)=exp(-γ·∥xy∥ 2 ) (16) S3.2 Calculate the distribution similarity between each source domain signal segment and all target domain signal segments using the maximum mean difference index, and obtain the similarity vector of the source domain samples: S3.3 Sort the source domain samples in descending order of similarity. Based on the preset retention ratio threshold, select the high-quality source domain samples that are most similar to the target domain distribution to form a set of transfer training source domains, so as to suppress the negative transfer caused by samples with large distribution differences.

7. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 1, characterized in that, S4 specifically includes: S4.1 Construct a one-dimensional convolutional neural network feature extractor. The preprocessed source and target domain signal segments are used as inputs. The signal segments are processed through multiple layers of one-dimensional convolution, nonlinear activation and pooling operations. Finally, a fixed-dimensional fault representation vector is obtained through adaptive pooling and L2 normalization. S4.2 Construct a shared multilayer perceptron as a label classifier on the representation vector. The output of the shared layer is connected to the fault type classification head and the defect size classification head respectively, so as to realize the joint discrimination of normal, inner ring fault, outer ring fault and rolling element fault as well as each defect size. The size classification head only participates in the loss calculation on faulty samples. S4.

3. A gradient inversion layer is connected to the output of the feature extractor and a domain discriminator is connected to perform binary classification of the source domain and target domain samples. The gradient inversion layer inverts the gradient sign of the domain discriminator during backpropagation to achieve adversarial feature learning. S4.4 A classification loss is constructed using labeled samples from the source domain, and a domain discriminant loss is constructed using samples from both the source and target domains. The classification loss and domain loss are weighted and summed according to preset weights to obtain the total loss. An adaptive learning rate optimization algorithm is used to jointly train the parameters of the feature extractor, label classifier, and domain discriminator, enabling the network to learn a shared feature space that possesses both fault separability and domain invariance. The classification loss is as follows: The size branch only accounts for the loss for abnormal class samples: Total category loss:

8. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 7, characterized in that, The one-dimensional convolutional neural network feature extractor includes four convolutional layers, each with a kernel length of 9. Padding is used to keep the sequence length constant. The first three convolutional layers are followed by a max pooling layer to compress the time dimension proportionally. The last convolutional layer is followed by an adaptive max pooling layer, mapping sequences of arbitrary length to a single time step output. The shared multilayer perceptron includes two fully connected layers with dimensions of 128→256→128. Dropout is applied between the two layers and at the output to suppress overfitting. The output of the shared multilayer perceptron is simultaneously connected to a fault type classification head and a defect size classification head. Both classification heads use linear layers to output four-dimensional logits, corresponding to normal states and three fault types or four defect sizes. The domain adversarial training is implemented using a gradient inversion layer. That is, the input of the domain discriminator is kept constant during forward propagation, and the gradient of the domain discriminator is multiplied by a negative value and scaled during backward propagation to achieve adversarial optimization between the feature extractor and the domain discriminator. The domain loss weight gradually increases with the training progress according to an increasing function.

9. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 1, characterized in that, S5 specifically includes: S5.1 Perform the same preprocessing and segmentation on the unlabeled bearing vibration signal in the target domain according to the procedures in S1 and S2. S5.2 Utilize the feature extractor and label classifier trained by domain adversarial training to output the predicted probability of fault type and defect size for each target domain signal segment, and use file-level majority voting or confidence weighting rules to generate the final fault diagnosis result of the target bearing. S5.3 Output the diagnostic results and their confidence levels to the train maintenance system for real-time monitoring, health assessment, and maintenance decision-making.

10. The method for cross-domain fault diagnosis of high-speed train bearings based on domain adversarial neural networks according to claim 9, characterized in that, The diagnosis of the target domain adopts the method of "segment-by-segment discrimination and file-level majority voting". The prediction results of all signal segments belonging to the same bearing file are statistically analyzed. When the proportion of segments of a certain type of label or the average confidence exceeds a preset threshold, the label of that type is taken as the final diagnosis result of the bearing.

Images