High-speed train bearing cross-domain fault diagnosis method based on transfer learning

Abstract

The invention discloses a high-speed train bearing cross-domain fault diagnosis method based on transfer learning, and relates to the technical field of train bearing fault diagnosis, and the method comprises the steps: carrying out the secondary combined noise reduction of a bearing vibration acceleration original signal; extracting time domain statistical features, frequency domain statistical features and associated time-frequency domain features, and constructing a high-dimensional feature set; carrying out dimension reduction processing on the high-dimensional feature set; constructing a multi-task neural network reference model for pre-training; constructing a CDAN migration diagnosis model based on feature adaptation; carrying out adversarial training and pseudo-label iterative optimization to obtain a trained migration diagnosis model; and carrying out actual diagnosis. According to the method, through two-stage combined noise reduction, multi-dimensional multi-domain feature extraction, dimension reduction processing, a multi-task neural network reference model and a feature adaptation-based CDAN migration diagnosis model, the accuracy and robustness of cross-domain fault diagnosis of the high-speed train bearing are improved by combining adversarial training and pseudo-tag iterative optimization.

Description

Technical Field

[0001] This application relates to the field of train bearing fault diagnosis technology, and in particular to a cross-domain fault diagnosis method for high-speed train bearings based on transfer learning. Background Technology

[0002] As a crucial component of modern transportation, the safety of the running gear in high-speed trains is paramount. Bearings, as critical rotating components, operate under high-speed, heavy-load, and complex alternating load conditions for extended periods, making them highly susceptible to pitting, spalling, and other faults. While data-driven intelligent diagnostic methods have made significant progress, they still face challenges such as data distribution discrepancies, strong noise interference, and unlabeled target domains. Therefore, further research is needed on train bearing fault diagnosis technology.

[0003] In the prior art, Chinese patent CN118013331A discloses a cross-device migration diagnosis method and system for train traction motor bearing condition detection. Based on a wavelet domain adaptive network with physical information, interpretable wavelet knowledge is integrated into a dual-stream convolutional layer with independent weights to cope with cross-machine diagnosis tasks. The first layer weights of the CNN are updated with optimized Laplace or Morlet wavelet weights containing rich information. The scale factor and translation factor with specific physical interpretation are constrained by the convolution kernel parameters, and a smoothing auxiliary scale factor is considered to ensure consistency with the neural network weights.

[0004] However, the aforementioned existing technologies rely solely on wavelet domain adaptive networks and dual-stream convolutional layers to achieve cross-device migration. They do not optimize the preprocessing process for the strong interference characteristics of bearing vibration signals, and the cross-domain feature adaptation mechanism is relatively simple. When faced with the complex service conditions of high-speed trains, the migration effect is limited, and the accuracy and robustness of cross-domain fault diagnosis of high-speed train bearings need to be improved. Summary of the Invention

[0005] This application provides a method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning, which aims to address the problem that the accuracy and robustness of cross-domain fault diagnosis of high-speed train bearings in the prior art need to be improved.

[0006] On the one hand, this application provides a method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning, including the following steps: Step 1: Obtain the original bearing vibration acceleration signals from the source and target domains. Perform two-stage joint noise reduction on the original bearing vibration acceleration signals using wavelet soft thresholding and adaptive filtering to obtain a clean vibration signal.

[0007] Step 2: Extract the time-domain statistical features, frequency-domain statistical features, and associated time-frequency domain features of the pure vibration signal to construct a high-dimensional feature set.

[0008] Step 3: Principal component analysis is used to reduce the dimensionality of the high-dimensional feature set to obtain the source domain feature vector and the target domain feature vector.

[0009] Step 4: Construct a multi-task neural network benchmark model. Use the source domain feature vector to pre-train the multi-task neural network benchmark model to obtain a pre-trained feature extractor and a pre-trained classifier.

[0010] Step 5: Construct a CDAN transfer diagnostic model based on feature adaptation, including a feature dimension adapter, a transfer feature extractor, a transfer classifier, and a conditional domain discriminator. The transfer feature extractor and the transfer classifier use the parameters of the pre-trained feature extractor and the pre-trained classifier as initial weights.

[0011] Step 6: Input the source domain feature vector and the adapted target domain feature vector into the CDAN transfer diagnostic model for adversarial training and pseudo-label iterative optimization to obtain the trained transfer diagnostic model.

[0012] Step 7: Collect vibration data of the high-speed train bearing to be diagnosed. After the corresponding processing in Steps 1 to 3, input the data into the trained transfer diagnostic model and output the fault type label and diagnostic confidence.

[0013] In one possible implementation, step one, the wavelet soft thresholding includes: The original bearing vibration acceleration signal is decomposed by wavelet decomposition, the noise standard deviation is estimated, a general threshold is calculated based on the noise standard deviation, a soft threshold function is constructed based on the general threshold, and the wavelet coefficients are nonlinearly contracted using the soft threshold function.

[0014] In one possible implementation, step one, the adaptive filtering includes: The least mean square adaptive filter is used. The signal that has passed through the wavelet soft threshold is used as the main input, and the synchronous pulse signal obtained by the speed sensor is used as the reference input. The output of the least mean square adaptive filter is subtracted from the main input to obtain the error signal, which is the pure vibration signal.

[0015] In one possible implementation, in step two, the time-domain statistical features include: mean, standard deviation, variance, root mean square value, skewness, peak value, minimum value, peak-to-peak value, peak factor, kurtosis, impulse factor, margin factor, and waveform index.

[0016] The frequency domain statistical features include: spectral centroid, frequency variance, spectral entropy, spectral root mean square, root mean square frequency, frequency standard deviation, spectral energy, and dominant frequency.

[0017] The associated time-frequency domain features include: wavelet packet energy distribution in 8 frequency bands, wavelet packet energy entropy, total energy of short-time Fourier transform, energy entropy of short-time Fourier transform, energy variance of short-time Fourier transform, average instantaneous frequency of Hilbert-Huang transform, and standard deviation of instantaneous frequency of Hilbert-Huang transform.

[0018] In one possible implementation, in step four, the multi-task neural network baseline model includes an input layer, a hard parameter shared hidden layer, and several task-specific output heads connected in sequence.

[0019] The hard parameter shared hidden layer consists of four fully connected layers. The number of neurons in the four fully connected layers from the input end to the output end are 256, 128, 64, and 32, respectively. The ReLU activation function is used between the layers.

[0020] Several task-specific output heads include: a first output head for identifying the data source, a second output head for identifying the fault type, and a third output head for identifying the fault size.

[0021] In one possible implementation, in step five, the feature dimension adapter is to add a linear transformation layer before the transfer feature extractor to map the target domain features in the target domain feature vector to the source domain dimension space, so as to adapt the transfer feature extractor and the transfer classifier.

[0022] In one possible implementation, in step five, the feature vector output by the transfer feature extractor is multilinearly mapped to the class probability distribution output by the transfer classifier to obtain a joint feature representation. The joint feature representation is then used as the input to the conditional domain discriminator, which is used to determine whether the sample corresponding to the joint feature representation comes from the source domain or the target domain.

[0023] In one possible implementation, in step five, a gradient inversion layer is provided between the multilinear mapping and the conditional domain discriminator to invert the gradient during backpropagation, thereby enabling adversarial training.

[0024] In one possible implementation, step six, the pseudo-label iterative optimization, employs a dynamic pseudo-label strategy, including: Set the initial training rounds and confidence threshold for pseudo-label intervention.

[0025] When the number of training rounds is greater than the number of initial training rounds, the feature vector of the target domain is predicted. If the maximum probability of the prediction is greater than the confidence threshold, the corresponding predicted category is used as a pseudo-label, the pseudo-label loss is calculated, and it participates in backpropagation.

[0026] The high-speed train bearing cross-domain fault diagnosis method based on transfer learning in this application has the following advantages: By employing two-stage joint noise reduction, multi-dimensional and multi-domain feature extraction, dimensionality reduction processing, a multi-task neural network benchmark model, and a feature-adaptive-based CDAN transfer diagnostic model, combined with adversarial training and pseudo-label iterative optimization, the accuracy and robustness of cross-domain fault diagnosis for high-speed train bearings have been improved.

[0027] Two-stage joint noise reduction of the original bearing vibration acceleration signal is performed using wavelet soft thresholding and adaptive filtering. Wavelet soft thresholding utilizes the time-frequency localization characteristic of wavelet transform to smoothly suppress high-frequency random noise through a soft thresholding function, avoiding pseudo-Gibbs oscillations caused by hard thresholding methods and effectively preserving transient impact components caused by faults. Adaptive filtering employs a least mean square adaptive filter, introducing the synchronization pulse signal obtained from the speed sensor as a reference to accurately cancel strong periodic interference such as shaft rotation frequency and its harmonics.

[0028] By extracting time-domain statistical features, frequency-domain statistical features, and associated time-frequency domain features, a high-dimensional feature set is constructed, which improves the comprehensiveness of feature extraction and data utilization.

[0029] By employing principal component analysis to reduce the dimensionality of high-dimensional feature sets, adaptive data compression is achieved while preserving physical information.

[0030] By adding a linear transformation layer before the transfer feature extractor, the target domain features in the target domain feature vector are mapped to the source domain dimension space, thus achieving physical alignment of heterogeneous features.

[0031] By performing a multilinear mapping between the feature vector output by the transfer feature extractor and the class probability distribution output by the transfer classifier, a joint feature representation is obtained and used as the input to the conditional domain discriminator, thereby capturing the complex distribution differences under multimodal structures and avoiding mode collapse.

[0032] By employing a dynamic pseudo-label strategy, pseudo-labels for the target domain are introduced during the stable training phase, and high-confidence samples are used to assist the model in refining the classification boundary in the unlabeled domain. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart illustrating the cross-domain fault diagnosis method for high-speed train bearings based on transfer learning provided in this application embodiment; Figure 2A time-domain waveform comparison diagram of the original bearing vibration acceleration signal and the pure vibration signal provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of a multi-task neural network baseline model provided in an embodiment of this application; Figure 4 This is a schematic diagram of the overall architecture of the CDAN migration diagnostic model provided in the embodiments of this application; Figure 5 A t-SNE visualization distribution of the target domain fault type label classification results provided in this application embodiment. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] like Figure 1 As shown in the figure, this application provides a method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning, including the following steps: Step 1: Obtain the original bearing vibration acceleration signals from the source and target domains. Perform two-stage joint noise reduction on the original bearing vibration acceleration signals using wavelet soft thresholding and adaptive filtering to obtain a clean vibration signal.

[0037] Step 2: Extract the time-domain statistical features, frequency-domain statistical features, and associated time-frequency domain features of the pure vibration signal to construct a high-dimensional feature set.

[0038] Step 3: Principal component analysis is used to reduce the dimensionality of the high-dimensional feature set to obtain the source domain feature vector and the target domain feature vector.

[0039] Step 4: Construct a multi-task neural network benchmark model. Use the source domain feature vector to pre-train the multi-task neural network benchmark model to obtain a pre-trained feature extractor and a pre-trained classifier.

[0040] Step 5: Construct a CDAN transfer diagnostic model based on feature adaptation, including a feature dimension adapter, a transfer feature extractor, a transfer classifier, and a conditional domain discriminator. The transfer feature extractor and the transfer classifier use the parameters of the pre-trained feature extractor and the pre-trained classifier as initial weights.

[0041] Step 6: Input the source domain feature vector and the adapted target domain feature vector into the CDAN transfer diagnostic model for adversarial training and pseudo-label iterative optimization to obtain the trained transfer diagnostic model.

[0042] Step 7: Collect vibration data of the high-speed train bearing to be diagnosed. After the corresponding processing in Steps 1 to 3, input the data into the trained transfer diagnostic model and output the fault type label and diagnostic confidence.

[0043] Specifically, in this embodiment, the original bearing vibration acceleration signal in the source domain is selected from the vibration acceleration data collected by the laboratory bearing fault simulation bench. The sampling frequency is 12kHz, covering four types: inner ring fault, outer ring fault, rolling element fault, and normal state, and includes different fault sizes (0.007 inches to 0.028 inches). The original bearing vibration acceleration signal in the target domain is selected from the online monitoring data of a certain type of high-speed train under actual operating conditions. This data is unlabeled and contains a lot of background noise.

[0044] For example, in step one, the wavelet soft thresholding includes: The original bearing vibration acceleration signal is decomposed by wavelet decomposition, the noise standard deviation is estimated, a general threshold is calculated based on the noise standard deviation, a soft threshold function is constructed based on the general threshold, and the wavelet coefficients are nonlinearly contracted using the soft threshold function.

[0045] Specifically, in this embodiment, the Daubechies 4 wavelet basis is selected to perform 4-level wavelet decomposition on the original bearing vibration acceleration signal, and the noise standard deviation estimate is calculated for the wavelet coefficients of each level. And according to the formula Calculate the general threshold Where N represents the length of the original bearing vibration acceleration signal. The expression for the soft threshold function is as follows: .

[0046] in, Represents the soft threshold function; Represents the contraction function; This represents the wavelet decomposition coefficients.

[0047] For example, in step one, the adaptive filtering includes: The least mean square adaptive filter is used. The signal that has passed through the wavelet soft threshold is used as the main input, and the synchronous pulse signal obtained by the speed sensor is used as the reference input. The output of the least mean square adaptive filter is subtracted from the main input to obtain the error signal, which is the pure vibration signal.

[0048] Specifically, let the main input be... The reference input is The order of the least mean square adaptive filter is (In this embodiment, we take 32), then the first... Output of the minimum mean square adaptive filter at time t and error signal for: , .

[0049] in, This represents the filter weights at time n.

[0050] Weight vector of least mean square adaptive filter The iterative update formula is: .

[0051] in, This is the step size factor (0.01 in this example). This is the reference input vector.

[0052] like Figure 2 As shown, the upper part is the time-domain waveform of the original bearing vibration acceleration signal before two-stage joint noise reduction, and the lower part is the time-domain waveform of the pure vibration signal after two-stage joint noise reduction. It can be seen that the two-stage joint noise reduction effectively suppresses the messy spikes in the original bearing vibration acceleration signal, while significantly enhancing the periodic impact pulse characteristics related to the fault, resulting in a significant improvement in the signal-to-noise ratio.

[0053] For example, in step two, the time-domain statistical features include: mean, standard deviation, variance, root mean square value, skewness, peak value, minimum value, peak-to-peak value, peak factor, kurtosis, impulse factor, margin factor, and waveform index.

[0054] The frequency domain statistical features include: spectral centroid, frequency variance, spectral entropy, spectral root mean square, root mean square frequency, frequency standard deviation, spectral energy, and dominant frequency.

[0055] The associated time-frequency domain features include: wavelet packet energy distribution in 8 frequency bands, wavelet packet energy entropy, total energy of short-time Fourier transform, energy entropy of short-time Fourier transform, energy variance of short-time Fourier transform, average instantaneous frequency of Hilbert-Huang transform, and standard deviation of instantaneous frequency of Hilbert-Huang transform.

[0056] Specifically, in this embodiment, in step three, the rank of the feature matrix differs due to the different data volume and signal-to-noise ratio between the source and target domains. This embodiment employs principal component analysis to reduce the dimensionality of the high-dimensional feature sets in both the source and target domains. Source domain dimensionality reduction: Principal components with a cumulative contribution rate of over 95.6% are retained, and 11 principal components are extracted through calculation, which are the source domain feature dimensions. ; Target domain dimensionality reduction: Retaining the same cumulative contribution rate, nine principal components were extracted, which are the feature dimensions of the target domain. This step eliminates redundant correlations between features and exposes engineering problems related to inconsistent cross-domain feature dimensions, providing a basis for subsequent feature dimension adapter design.

[0057] like Figure 3 As shown, exemplarily, in step four, the multi-task neural network baseline model includes an input layer, a hard parameter shared hidden layer, and several task-specific output heads connected in sequence.

[0058] The hard parameter shared hidden layer consists of four fully connected layers. The number of neurons in the four fully connected layers from the input end to the output end are 256, 128, 64, and 32, respectively. The ReLU activation function is used between the layers.

[0059] Several task-specific output heads include: a first output head for identifying the data source, a second output head for identifying the fault type, and a third output head for identifying the fault size.

[0060] Specifically, in this embodiment, the number of nodes in the input layer of the multi-task neural network benchmark model is 11, which corresponds to the feature dimension of the source domain.

[0061] Hard parameter shared hidden layer: It contains 4 fully connected layers. The number of neurons in the 4 fully connected layers from the input end to the output end are 256, 128, 64 and 32 respectively. The ReLU activation function is used between layers to increase non-linearity, and a Dropout layer is set with a ratio of 0.5 to prevent overfitting.

[0062] Task-specific output heads: First output head (Head1, identifies data source, output dimension 2: drive end / fan end), second output head (Head2, identifies fault type, output dimension 4: normal / inner race fault / outer race fault / rolling element fault), third output head (Head3, identifies fault size, output dimension 5). The multi-task neural network benchmark model was supervised and trained for 200 epochs on the source domain training set. Experimental results show that the multi-task neural network benchmark model achieves an accuracy of 88.97% in fault type identification and an F1 score of 0.91 on the source domain test set, demonstrating the effectiveness of the feature extraction structure.

[0063] like Figure 4 As shown, exemplarily, in step five, the feature dimension adapter (corresponding to...) Figure 4 The feature adapter in the transfer feature extractor (corresponding to) Figure 4A linear transformation layer is added before the shared feature extractor to map the target domain features in the target domain feature vector to the source domain dimension space, so as to adapt the transfer feature extractor and the transfer classifier.

[0064] Specifically, in this embodiment, the mathematical model of the feature dimension adapter is a linear transformation. ,in The weight matrix is ​​a learnable matrix. This represents the characteristics of the target domain before linear transformation. This represents the target domain features after linear transformation, used to solve the feature dimension mismatch problem when the feature dimension of the source domain is not equal to the feature dimension of the target domain.

[0065] For example, in step five, the feature vector output by the transfer feature extractor is multilinearly mapped to the class probability distribution output by the transfer classifier to obtain a joint feature representation. The joint feature representation is used as the input of the conditional domain discriminator, which is used to determine whether the sample corresponding to the joint feature representation comes from the source domain or the target domain.

[0066] Specifically, in this embodiment, the migration feature extractor (corresponding to...) Figure 4 The feature vector output by the shared feature extractor in the middle With transfer classifier (corresponding) Figure 4 The class probability distribution output by the classifier in the classifier. (Softmax output) undergoes multilinear mapping (outer product operation) to obtain joint feature representation. Conditional domain discriminator (corresponding) Figure 4 The domain discriminant in the model uses joint features to represent... As input, the joint feature indicates whether the corresponding sample comes from the source domain or the target domain.

[0067] For example, in step five, a gradient inversion layer is provided between the multilinear mapping and the conditional domain discriminator to invert the gradient during backpropagation, thereby enabling adversarial training.

[0068] Specifically, in step six, the source domain feature vector (labeled) and the adapted target domain feature vector (unlabeled) are input into the CDAN transfer diagnostic model for joint training.

[0069] Specifically, in this embodiment, the hyperparameter settings of the CDAN transfer diagnostic model are as follows: 100 training rounds, batch size of 64, Adam optimizer selected, and initial learning rate of 0.0001.

[0070] Loss function of CDAN transfer diagnostic model: A composite loss function is used. .in, The source domain classification loss is used to ensure the source domain classification accuracy; To combat the loss, adversarial learning is used to confuse the domain distribution; To minimize the entropy loss in the target domain, the principle of entropy minimization is used to make the target domain prediction more certain. For pseudo-label loss; These are the weight coefficients for adversarial loss, target domain entropy loss, and pseudo-label loss, respectively.

[0071] For example, in step six, the pseudo-label iterative optimization employs a dynamic pseudo-label strategy, including: Set the initial training rounds and confidence threshold for pseudo-label intervention.

[0072] When the number of training rounds is greater than the number of initial training rounds, the feature vector of the target domain is predicted. If the maximum probability of the prediction is greater than the confidence threshold, the corresponding predicted category is used as a pseudo-label, the pseudo-label loss is calculated, and it participates in backpropagation.

[0073] Specifically, in this embodiment, the initial training rounds for pseudo-labeling are set to 20, and the confidence threshold is 0.6. Only when the maximum probability of the CDAN transfer diagnostic model's prediction of the target domain feature vector is greater than 0.6 is the corresponding predicted category used as a pseudo-label, and the pseudo-label loss is calculated. It also participates in backpropagation to update model parameters. This strategy effectively utilizes high-confidence samples and corrects the classification boundary.

[0074] like Figure 5 The figure shown is a t-SNE visualization distribution of the target domain fault type label classification results. It can be seen that in the target domain containing outer ring faults, inner ring faults and rolling element faults, the method of this application can distinguish the three fault types in the unlabeled data quite well.

[0075] The embodiments of this application improve the accuracy and robustness of cross-domain fault diagnosis of high-speed train bearings by using two-level joint noise reduction, multi-dimensional and multi-domain feature extraction, dimensionality reduction processing, multi-task neural network benchmark model, and feature-adaptive CDAN transfer diagnostic model, combined with adversarial training and pseudo-label iterative optimization.

[0076] Two-stage joint noise reduction of the original bearing vibration acceleration signal is performed using wavelet soft thresholding and adaptive filtering. Wavelet soft thresholding utilizes the time-frequency localization characteristic of wavelet transform to smoothly suppress high-frequency random noise through a soft thresholding function, avoiding pseudo-Gibbs oscillations caused by hard thresholding methods and effectively preserving transient impact components caused by faults. Adaptive filtering employs a least mean square adaptive filter, introducing the synchronization pulse signal obtained from the speed sensor as a reference to accurately cancel strong periodic interference such as shaft rotation frequency and its harmonics.

[0077] By extracting time-domain statistical features, frequency-domain statistical features, and associated time-frequency domain features, a high-dimensional feature set is constructed, which improves the comprehensiveness of feature extraction and data utilization.

[0078] By employing principal component analysis to reduce the dimensionality of high-dimensional feature sets, adaptive data compression is achieved while preserving physical information.

[0079] By adding a linear transformation layer before the transfer feature extractor, the target domain features in the target domain feature vector are mapped to the source domain dimension space, thus achieving physical alignment of heterogeneous features.

[0080] By performing a multilinear mapping between the feature vector output by the transfer feature extractor and the class probability distribution output by the transfer classifier, a joint feature representation is obtained and used as the input to the conditional domain discriminator, thereby capturing the complex distribution differences under multimodal structures and avoiding mode collapse.

[0081] By employing a dynamic pseudo-label strategy, pseudo-labels for the target domain are introduced during the stable training phase, and high-confidence samples are used to assist the model in refining the classification boundary in the unlabeled domain.

[0082] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0083] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning, characterized in that, Includes the following steps: Step 1: Obtain the original bearing vibration acceleration signals from the source and target domains, and perform two-stage joint noise reduction on the original bearing vibration acceleration signals using wavelet soft thresholding and adaptive filtering to obtain a clean vibration signal; Step 2: Extract the time-domain statistical features, frequency-domain statistical features, and associated time-frequency-domain features of the pure vibration signal to construct a high-dimensional feature set; Step 3: Principal component analysis is used to reduce the dimensionality of the high-dimensional feature set to obtain the source domain feature vector and the target domain feature vector. Step 4: Construct a multi-task neural network benchmark model. Use the source domain feature vectors to pre-train the multi-task neural network benchmark model to obtain a pre-trained feature extractor and a pre-trained classifier. Step 5: Construct a CDAN transfer diagnostic model based on feature adaptation, including a feature dimension adapter, a transfer feature extractor, a transfer classifier, and a conditional domain discriminator. The transfer feature extractor and the transfer classifier use the parameters of the pre-trained feature extractor and the pre-trained classifier as initial weights. Step 6: Input the source domain feature vector and the adapted target domain feature vector into the CDAN transfer diagnostic model for adversarial training and pseudo-label iterative optimization to obtain the trained transfer diagnostic model. Step 7: Collect vibration data of the high-speed train bearing to be diagnosed. After the corresponding processing in Steps 1 to 3, input the data into the trained transfer diagnostic model and output the fault type label and diagnostic confidence.

2. The method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning according to claim 1, characterized in that, In step one, the wavelet soft thresholding includes: The original bearing vibration acceleration signal is decomposed by wavelet decomposition, the noise standard deviation is estimated, a general threshold is calculated based on the noise standard deviation, a soft threshold function is constructed based on the general threshold, and the wavelet coefficients are nonlinearly contracted using the soft threshold function.

3. The method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning according to claim 1, characterized in that, In step one, the adaptive filtering includes: The least mean square adaptive filter is used. The signal that has passed through the wavelet soft threshold is used as the main input, and the synchronous pulse signal obtained by the speed sensor is used as the reference input. The output of the least mean square adaptive filter is subtracted from the main input to obtain the error signal, which is the pure vibration signal.

4. The high-speed train bearing cross-domain fault diagnosis method based on transfer learning according to claim 1, characterized in that, In step two, the time-domain statistical features include: mean, standard deviation, variance, root mean square value, skewness, peak value, minimum value, peak-to-peak value, peak factor, kurtosis, impulse factor, margin factor, and waveform index. The frequency domain statistical features include: spectral centroid, frequency variance, spectral entropy, spectral root mean square, root mean square frequency, frequency standard deviation, spectral energy, and dominant frequency. The associated time-frequency domain features include: wavelet packet energy distribution in 8 frequency bands, wavelet packet energy entropy, total energy of short-time Fourier transform, energy entropy of short-time Fourier transform, energy variance of short-time Fourier transform, average instantaneous frequency of Hilbert-Huang transform, and standard deviation of instantaneous frequency of Hilbert-Huang transform.

5. The method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning according to claim 1, characterized in that, In step four, the multi-task neural network baseline model includes an input layer, a hard parameter shared hidden layer, and several task-specific output heads connected in sequence. The hard parameter shared hidden layer consists of four fully connected layers. The number of neurons in the four fully connected layers from the input end to the output end are 256, 128, 64, and 32, respectively. The ReLU activation function is used between the layers. Several task-specific output heads include: a first output head for identifying the data source, a second output head for identifying the fault type, and a third output head for identifying the fault size.

6. The method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning according to claim 1, characterized in that, In step five, the feature dimension adapter adds a linear transformation layer before the transfer feature extractor to map the target domain features in the target domain feature vector to the source domain dimension space, so as to adapt the transfer feature extractor and the transfer classifier.

7. The method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning according to claim 1, characterized in that, In step five, the feature vector output by the transfer feature extractor is multilinearly mapped to the class probability distribution output by the transfer classifier to obtain a joint feature representation. The joint feature representation is then used as the input to the conditional domain discriminator, which is used to determine whether the sample corresponding to the joint feature representation comes from the source domain or the target domain.

8. The method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning according to claim 7, characterized in that, In step five, a gradient inversion layer is set between the multilinear mapping and the conditional domain discriminator to invert the gradient during backpropagation, thereby enabling adversarial training.

9. The method for cross-domain fault diagnosis of high-speed train bearings based on transfer learning according to claim 1, characterized in that, In step six, the pseudo-label iterative optimization adopts a dynamic pseudo-label strategy, including: Set the initial training rounds and confidence threshold for pseudo-label intervention; When the number of training rounds is greater than the number of initial training rounds, the feature vector of the target domain is predicted. If the maximum probability of the prediction is greater than the confidence threshold, the corresponding predicted category is used as a pseudo-label, the pseudo-label loss is calculated, and it participates in backpropagation.

Images