A cross-domain locomotive bearing fault diagnosis method and device based on a twin network

Abstract

The application discloses a cross-domain locomotive bearing fault diagnosis method and device based on a twin network, which comprises collecting vibration signals of a locomotive bearing and current signals of a traction motor, pre-processing the vibration signals and the current signals and obtaining a frequency spectrum; a twin network containing a weight sharing branch is constructed, the vibration signal frequency spectrum is input into the twin network, a vibration domain feature encoder is obtained by performing self-supervised training on an encoder based on a contrastive learning loss function; a classifier for fault type classification and speed prediction is constructed based on the vibration domain feature encoder, a vibration domain bearing fault diagnosis model is obtained by performing supervised fine-tuning on labeled vibration signal samples; the vibration domain feature encoder is migrated to a current signal domain, a current domain classifier is trained based on labeled current data, and a current domain bearing fault diagnosis model is obtained; current signals collected under a to-be-diagnosed working condition are input into the current domain bearing fault diagnosis model, and corresponding bearing fault types and speed information are output.

Description

Technical Field

[0001] This invention relates to the field of locomotive bearing fault diagnosis technology, specifically to a cross-domain locomotive bearing fault diagnosis method and device based on twin networks. Background Technology

[0002] As a critical component of the traction system, the operating condition of locomotive bearings directly affects the safety and reliability of the locomotive. Existing locomotive bearing fault diagnosis methods largely rely on vibration acceleration sensors installed on axle boxes or wheelsets. These sensors collect vibration signals and extract time-domain or frequency-domain features, then combine these signals with traditional machine learning or deep learning models for fault identification. However, existing technologies have at least several limitations.

[0003] First, sensor cost and installation conditions are limiting factors. Vibration sensor installation locations are restricted by space, installation environment, and maintenance conditions, making large-scale deployment difficult on some locomotives or existing lines. Furthermore, sensor procurement and wiring costs are high. Second, easily acquired signals have poor diagnostic capabilities. Compared to vibration signals, locomotive traction motor phase current signals are easier to collect on-site. However, fault characteristics in current signals are often masked by changes in operating conditions and background noise. Directly diagnosing bearing faults based on current signals has limited accuracy and insufficient robustness. Third, differences between the operating conditions and the signal domain lead to poor cross-domain performance of the model. Vibration signals collected on the test bench and current signals collected on-site differ significantly in distribution. Changes in speed and load under different operating conditions also cause "domain shifts," making it difficult for models trained in the vibration domain to be directly transferred to the current domain, resulting in poor generalization ability. Fourth, existing cross-domain and deep learning methods have limitations. Some existing fault diagnosis methods based on deep neural networks and domain adaptation usually require a large amount of labeled data in the target domain, or achieve feature alignment only through adversarial training and statistical moment matching. They fail to make full use of the correlation structure between the vibration domain and the current domain, and do not combine mechanisms such as Siamese networks, self-supervised representation learning and multi-task learning for joint design.

[0004] Therefore, there is an urgent need for a new diagnostic method and system that can fully utilize the cross-domain correlation between vibration signals and current signals, achieve high-precision diagnosis of locomotive bearing faults even when there are few or no labeled samples in the target domain, and take into account the model's cross-domain generalization ability and engineering feasibility. Summary of the Invention

[0005] This invention addresses the problems in existing locomotive bearing fault diagnosis, such as reliance on vibration signals, readily available but poorly diagnosable field signals, and weak cross-domain generalization ability. It proposes a cross-domain locomotive bearing fault diagnosis method and device based on Siamese networks. By performing representation learning and classification learning in the vibration signal domain and transferring the obtained feature codes to the current signal domain, high-precision bearing fault diagnosis based on current signals is achieved, thereby improving fault diagnosis accuracy and cross-domain generalization ability with less current domain labeled data.

[0006] To achieve the above objectives, this invention proposes a cross-domain locomotive bearing fault diagnosis method based on twin networks, the method comprising the following steps:

[0007] A cross-domain locomotive bearing fault diagnosis method based on twin networks, the method comprising:

[0008] S1. Collect vibration signals from locomotive bearings and current signals from traction motors, preprocess the vibration signals and current signals, and obtain a spectrum diagram;

[0009] S2. Construct a Siamese network with weight-sharing branches, input the vibration signal spectrum into the Siamese network, and perform self-supervised training on the encoder based on the contrastive learning loss function to obtain the vibration domain feature encoder.

[0010] S3. Based on the vibration domain feature encoder, a classifier for fault type classification and speed prediction is constructed. The classifier is then fine-tuned using labeled vibration signal samples to obtain a vibration domain bearing fault diagnosis model.

[0011] S4. Transfer the vibration domain feature encoder to the current signal domain, and train the vibration domain feature encoder and current domain classifier transferred to the current domain based on real labeled current data to obtain the current domain bearing fault diagnosis model.

[0012] S5. Input the current signal collected under the working condition to be diagnosed into the current domain bearing fault diagnosis model, and output the corresponding bearing fault type and speed information.

[0013] Furthermore, the spectrogram is generated using variational mode decomposition and fast Fourier transform, the methods of which include:

[0014] Variational mode decomposition is used to decompose the original time-domain signal into several intrinsic mode components by minimizing the following objective function:

[0015] ;

[0016] By alternately optimizing, each mode is iteratively updated. and its corresponding center frequency Minimize the objective function until the convergence condition is met, where For unit impact function, For Lagrange multipliers, The original signal to be decomposed. As a penalty factor, For time, It is the imaginary unit.

[0017] Furthermore, in step S2, two different data augmentation views of the same vibration signal sample are input into two branches of the Siamese network, respectively. The corresponding feature vectors are obtained through the encoder and the projection head. Positive and negative sample pairs are constructed using the contrastive learning loss function, and the feature vectors are subjected to self-supervised representation learning using NT-XENT loss to obtain the feature encoder. The feature encoder uses the introduction of channel and spatial attention mechanisms to perform weighted processing on the channel and spatial dimensions of the spectrogram respectively.

[0018] Furthermore, the model is trained by maximizing the similarity between sample pairs, as follows:

[0019] ;

[0020] , These are features output through a twin network; It's a temperature parameter. This refers to the number of samples in each batch during training; Sim is... , Cosine similarity between them:

[0021] ;

[0022] The final losses are as follows:

[0023] ;

[0024] Indicates to The backpropagation of the path feature gradient is interrupted.

[0025] Furthermore, during the self-supervised representation learning process of the feature vector using NT-XENT loss, the gradient backpropagation of one branch of the feature vector is cut off.

[0026] Furthermore, based on the feature encoder, a classifier for vibration signals is constructed. The classifier includes a classification branch for outputting fault types and a regression branch for outputting rotational speed parameters. Labeled vibration signal samples are used to perform supervised fine-tuning training on the feature encoder and classifier to classify fault categories.

[0027] Furthermore, the training loss is a weighted sum of the cross-entropy loss of the fault type and the regression loss of the rotational speed prediction.

[0028] Furthermore, the classifier distinguishes between rotational speed categories and fault categories, using the following training loss:

[0029] ;

[0030] ;

[0031] in Loss for classifying fault types, SupCon is the loss for classifying rotational speed types, and SupCon is the supervised comparison loss. The loss weights are for classifying rotational speed types. For supervised comparison of loss weights, It's a temperature parameter. For cosine similarity, It is a collection of samples of the same category. It is the temperature coefficient hyperparameter.

[0032] Furthermore, during supervised fine-tuning training, the parameters of the shallow convolutional layers of the feature encoder are frozen, while the parameters of the higher-level convolutional layers are unfrozen and trained.

[0033] A cross-domain locomotive bearing fault diagnosis device based on twin networks includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of any of the methods described above.

[0034] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0035] This invention provides a cross-domain locomotive bearing fault diagnosis method and device based on Siamese networks. It utilizes vibration signals to construct a source domain representation model and transfers it to the target domain of current signals, achieving accurate bearing fault identification relying solely on easily obtainable current signals, significantly reducing the number and cost of vibration sensors deployed on-site. Furthermore, this invention designs a Siamese network structure incorporating the CBAM attention mechanism and a three-stage training architecture for this cross-domain diagnosis problem. Under a unified framework of self-supervised comparative learning, supervised fine-tuning, and current domain transfer learning, it effectively improves the discriminative power and cross-domain generalization ability of feature representations. Simultaneously, through multi-task joint learning and encoder freeze / unfreeze strategies, it ensures fault identification accuracy while also addressing the modeling requirements for operating condition information such as rotational speed. This invention can guarantee the stability of the closed-loop diagnostic system and the effectiveness of fault identification even when the target domain has limited or no source domain data participating in online inference. Compared with existing locomotive bearing fault diagnosis technologies based on single-domain signals or traditional domain adaptive methods, this invention does not require the deployment of a large number of high-cost sensors in the target domain, nor does it rely on complex adversarial training or explicit distributed matching structures. While reducing the difficulty of engineering implementation and hardware investment, it significantly improves the accuracy and robustness of cross-domain fault diagnosis. Attached Figure Description

[0036] Figure 1 The flowchart illustrates the cross-domain locomotive bearing fault diagnosis method based on twin networks according to the present invention.

[0037] Figure 2 This is a structural diagram of the method according to an embodiment of the present invention;

[0038] Figure 3 This is a time-domain waveform diagram of the normal vibration signal collected during the implementation of this invention;

[0039] Figure 4 The time-domain waveform of the inner ring fault vibration signal collected during the implementation of this invention;

[0040] Figure 5 This is a time-domain waveform diagram of the outer ring fault vibration signal collected during the implementation of this invention;

[0041] Figure 6 This is a time-domain waveform of the ball bearing fault vibration signal collected during the implementation of this invention;

[0042] Figure 7 This is a loss curve diagram in step S2 of the present invention;

[0043] Figure 8 This is a loss curve diagram for multi-task classification in step S3 of the present invention;

[0044] Figure 9 This is a diagram showing the confusion matrix results of dataset 1 in an embodiment of the present invention;

[0045] Figure 10 This is a diagram showing the confusion matrix results of dataset 2 in this embodiment of the invention;

[0046] Figure 11 This is a diagram showing the confusion matrix results of dataset 3 in an embodiment of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0048] like Figure 1 and Figure 2 As shown, the present invention provides a cross-domain locomotive bearing fault diagnosis method based on twin networks, including the following steps.

[0049] S1. Collect vibration signals from locomotive bearings and current signals from traction motors, preprocess the vibration signals and current signals, and obtain a spectrum diagram.

[0050] Multi-source bearing operating signals, including vibration and current signals, are collected under locomotive bearing test benches or actual locomotive operating conditions. The vibration signals are measured by an accelerometer installed near the bearing, and the current signals are the current signals of each phase of the traction motor. The operating conditions include locomotive bearing operation at different speeds, under fault conditions, and other working conditions.

[0051] The vibration signal and current signal are cleaned and normalized. Variational mode decomposition (VMD) is used to decompose the original time-domain signal into several intrinsic mode components. Fast Fourier Transform (FFT) is performed on each mode component to obtain the corresponding spectrum or time-frequency graph. The obtained spectrum is cropped and interpolated to a preset size to serve as the input sample for the twin network.

[0052] VMD will handle complex signals It is decomposed into K intrinsic modal components (IMFs), each mode having a specific frequency component.

[0053] ;

[0054] This can be transformed into an optimization problem, with the goal of minimizing the following objective function:

[0055] ;

[0056] By alternately optimizing, each mode is iteratively updated. and its corresponding center frequency Minimize the objective function until the convergence condition is met. The convergence condition is set according to the actual data. For unit impact function, For Lagrange multipliers, The original signal to be decomposed. As a penalty factor, For time, It is the imaginary unit.

[0057] The Fast Fourier Transform (FFT) is used to compute the Discrete Fourier Transform (DFT), converting a time-domain signal into a frequency-domain signal.

[0058] S2. Construct a Siamese network with weight-sharing branches, input the vibration signal spectrum into the Siamese network, and perform self-supervised training on the encoder based on the contrastive learning loss function to obtain the vibration domain feature encoder.

[0059] The Siamese neural network consists of two subnetworks sharing weights, with inputs processed by each subnetwork. Cosine similarity is used to calculate the similarity between the inputs. A Siamese network with two weight-sharing branches is constructed, each branch including an encoder and a projector head. The encoder employs a two-dimensional convolutional neural network structure incorporating channel and spatial attention mechanisms. The feature encoder utilizes these mechanisms to weight the channel and spatial dimensions of the spectrogram.

[0060] Two different data augmentation views of the same vibration signal sample are input into two branches of the Siamese network, respectively. The corresponding feature vectors are obtained through the encoder and the projector. Positive and negative sample pairs are constructed using the contrastive learning loss function, and the feature vectors are subjected to self-supervised representation learning using the NT-XENT loss.

[0061] The NT-XENT loss trains the model by maximizing the similarity between sample pairs. Its formula is as follows:

[0062] ;

[0063] , It is the feature output by the network, that is, the feature vector extracted by the encoder; It's a temperature parameter, and its function is to control the scale of similarity values. This refers to the number of samples in each batch during training; Sim is... , Cosine similarity between them.

[0064] ;

[0065] The final losses are as follows:

[0066] .

[0067] Indicates to The gradient backpropagation of the feature vectors is interrupted. During the contrastive learning process, a gradient interruption strategy is set for the gradient propagation path. The gradient backpropagation of the feature vectors of one branch is cut off, and backpropagation is only performed on the other branch and the shared encoder. This improves the stability of self-supervised training and avoids network collapse.

[0068] The encoder employs a 2D CNN with a channel and spatial attention mechanism (CBAM). This mechanism combines channel and spatial attention, weighting the channel and spatial dimensions of the spectrogram separately to enhance the network's focus on key features and improve performance in subsequent fine-tuning stages.

[0069] S3. Based on the vibration domain feature encoder, a classifier for fault type classification and speed prediction is constructed. The classifier is then fine-tuned using labeled vibration signal samples to obtain a vibration domain bearing fault diagnosis model.

[0070] Based on the obtained feature encoder, a classifier for vibration signals is constructed. The classifier adopts a multi-task output head, which includes a classification branch for outputting fault types and a regression branch for outputting operating parameters such as rotational speed.

[0071] Labeled vibration signal samples are used to perform supervised fine-tuning training on the feature encoder and multi-task output head. The training loss is a weighted sum of the cross-entropy loss of fault types and the regression loss of speed prediction, where the weight coefficients of each loss term are preset constants. The labeled vibration signal samples contain vibration signal data and their corresponding speeds and fault types.

[0072] In S2, the network has already learned unsupervised feature representations. Then, labeled data is used to fine-tune the encoder weights and train a classifier to classify fault categories. The training process can involve unfreezing the last one or the last two layers of the encoder for training.

[0073] Preferably, during the fine-tuning training process, the parameters of the shallow convolutional layers of the feature encoder are frozen, and the parameters of the high-level convolutional layers are unfrozen and trained, so as to balance feature stability and adaptability to specific tasks, and obtain a fault diagnosis model trained in the vibration domain.

[0074] In the classification experiment of target domain current signals, it was found that the current signal is highly sensitive to changes in rotational speed. Therefore, during training, the classifier's task was to simultaneously distinguish between "rotational speed category" and "fault category". The loss function is as follows:

[0075] ;

[0076] ;

[0077] in Loss for classifying fault types, SupCon is the loss for classifying rotational speed types, and SupCon is the supervised comparison loss. The loss weights are for classifying rotational speed types. For supervised comparison of loss weights, It's a temperature parameter. For cosine similarity, It is a collection of samples of the same category. It is the temperature coefficient hyperparameter.

[0078] The rotational speed characteristics in vibration signals are not obvious. Overemphasizing the rotational speed branch during fine-tuning classification can interfere with the model's learning of fault features. Therefore, the proportion of rotational speed loss is reduced, allowing the model to focus on the representation of fault features, while retaining a small amount of speed information as an auxiliary factor. A step is adopted: first train the bottleneck layer and classifier, then unfreeze the encoder midway, with α set to 0.3. This completes the fine-tuning process.

[0079] S4. Transfer the vibration domain feature encoder to the current signal domain, and train the vibration domain feature encoder and current domain classifier transferred to the current domain based on real labeled current data to obtain the current domain bearing fault diagnosis model.

[0080] The vibration domain feature encoder trained in step S3 is transferred to the current signal domain. The spectrum of the current signal is input to the feature encoder to obtain the representation of the current signal in the feature space. The real labeled current data includes current data and corresponding rotational speed and fault information, which has the following two advantages: The construction method of the current domain classifier is the same as that of the vibration domain classifier.

[0081] Using real-labeled current data is superior to using pseudo-labeled current data: Supervised learning using labeled current domain data avoids the propagation of pseudo-label noise, resulting in more accurate classification boundaries, more stable convergence, and overall better accuracy and reliability than pseudo-label methods.

[0082] Fault and speed dual-task enhancement generalization: Joint training of fault identification and speed discrimination enables the model to learn a more robust representation of changes in operating conditions, resulting in more stable performance and stronger generalization in the target domain (current signal).

[0083] S5. Input the current signal collected under the working condition to be diagnosed into the current domain bearing fault diagnosis model, and output the corresponding bearing fault type and speed information.

[0084] The current signal acquired under the condition to be diagnosed is preprocessed according to step S1 and then input into the current and bearing fault diagnosis model obtained by the transferred feature encoder and current domain classifier (S4). The model outputs the corresponding fault type and speed information. The obtained fault diagnosis results are then visualized to realize the fault diagnosis of the locomotive bearing condition.

[0085] The method will be further explained below using the HXD1 locomotive motor bearing dataset from CRRC Zhuzhou Institute.

[0086] Taking the drive-end bearing of a locomotive traction motor as the research object, the cross-domain locomotive bearing fault diagnosis method based on Siamese networks described in this invention is verified. Figures 3-11 As shown, the test platform is equipped with vibration acceleration sensors on the outside of the bearing housing to collect bearing vibration signals. Simultaneously, three-phase stator current signals are collected from the stator side of the traction motor. The sampling period is uniformly set to a fixed time step, with each sample corresponding to a continuous period of locomotive operation data. After removing obviously erroneous records and severely missing operating conditions, several vibration signal segments are selected as the source domain dataset. Multiple sets of current signals are collected under different load and speed combinations to form the target domain dataset. The source domain dataset includes various typical fault types such as normal operation, inner race fault, outer race fault, and rolling element fault.

[0087] At each sampling time, based on the original time-domain sequence of the vibration or current signal, denoising, detrending, and normalization are first performed. Then, variational mode decomposition is used to decompose the signal into several modal components. Fast Fourier Transform is performed on the modes containing the main fault information to obtain a fixed-size spectrum as the network input feature. The source domain vibration data, after the above preprocessing, constitutes the source domain sample sequence, and the target domain current data, after the same preprocessing process, constitutes the target domain sample sequence.

[0088] In terms of model construction, the Siamese network structure with an attention mechanism proposed in this invention is adopted. The single-branch encoder consists of a cascade of multi-layer two-dimensional convolutional, batch normalization, and pooling layers, with channel attention and spatial attention units connected in series after several convolutional modules to enhance attention to key frequency bands and regions. The encoder output is mapped to the contrastive learning space via a fully connected prediction head. During training, a self-supervised contrastive learning task is constructed on source domain vibration data. For the same sample, two views are generated through data augmentation such as random cropping and perturbation, and input into the two branches of the Siamese network respectively. The NT-XENT contrastive loss is used for representation learning. At the same time, an alternating gradient interruption mechanism is introduced on one side of the prediction head to avoid feature collapse caused by the backpropagation of erroneous gradients.

[0089] After completing self-supervised pre-training, fault labels from the source domain vibration data are introduced, and a classifier is added after the shared encoder for supervised fine-tuning training of fault types. Based on the available rotational speed information in the samples, a rotational speed output branch can also be added simultaneously to form a multi-task loss function. A shallow freeze and deep thaw training strategy is adopted to improve the adaptability of high-level features to specific tasks while maintaining the stability of the underlying general features, resulting in a source domain fault diagnosis model and a shared feature encoder. Subsequently, the shared encoder is transferred to the target domain current data, and features are extracted from the current spectrum. A current domain classifier is trained based on a small number of labeled current samples or labels generated through prototype clustering and pseudo-labels, establishing a cross-domain bearing fault diagnosis model that relies solely on current signals. The overall structure of the method of this invention is illustrated below. Figure 1 As shown.

[0090] Several representative target domain datasets were selected, namely Dataset 1, Dataset 2, and Dataset 3. Each dataset corresponds to data from different bearings, and includes three speeds: 600 r / min, 1500 r / min, and 2500 r / min. Each speed corresponds to four types of faults: no fault, outer race fault, inner race fault, and ball bearing fault. The fault identification performance of the method of this invention was systematically evaluated using the three datasets under different speed fault combinations. Table 1 shows the fault identification results on each typical target domain dataset: On datasets where the current data feature distribution is relatively close to the source domain, the method of this invention maintains a high level in terms of accuracy, recall, and F1-score. On datasets where the current signal distribution is significantly drifted and the operating conditions fluctuate greatly, the identification accuracy and macro-average F1-score of each fault category remain stable overall, without significant degradation, although affected by changes in operating conditions. The comparison results of multiple datasets show that the method of this invention can maintain consistent and reliable diagnostic performance under different target domain distributions and complex operating conditions, and has good robustness to differences in source-target domain distributions, making it suitable for widespread application in multi-condition locomotive operating environments.

[0091] Table 1 Comparison of evaluation metrics for each dataset

[0092] Data Precision Recall F1-Score Dataset 1 0.9409 0.9333 0.9341 Dataset 2 0.9355 0.9333 0.9334 Dataset 3 0.9402 0.9333 0.9346

[0093] Figures 9 to 11 The classification confusion matrix results for each dataset show that the classification model maintains high accuracy and stability across different current datasets, demonstrating good diagnostic performance across domains. Based on pre-trained representations in the vibration domain, the model can accurately identify faults using only a few samples in the current domain, exhibiting outstanding generalization ability.

[0094] The present invention also provides a cross-domain locomotive bearing fault diagnosis device based on twin networks, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described above. The specific method steps will not be described in detail.

[0095] The above embodiments are used to explain the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A twin network based cross-domain locomotive bearing fault diagnosis method, characterized by The method includes: S1. Collect vibration signals from locomotive bearings and current signals from traction motors, preprocess the vibration signals and current signals, and obtain a spectrum diagram; S2. Construct a Siamese network containing a weight-sharing branch. Input the vibration signal spectrum into the Siamese network and perform self-supervised training on the encoder based on the contrastive learning loss function to obtain the vibration domain feature encoder. In step S2, two different data augmentation views of the same vibration signal sample are input into the two branches of the Siamese network respectively. The corresponding feature vectors are obtained through the encoder and the projection head. Positive and negative sample pairs are constructed using the contrastive learning loss function, and the feature vectors are subjected to self-supervised representation learning using NT-XENT loss to obtain the feature encoder. The feature encoder uses the introduction of channel and spatial attention mechanisms to weight the channel and spatial dimensions of the spectrum respectively. S3. Based on the vibration domain feature encoder, a classifier for fault type classification and speed prediction is constructed. Supervised fine-tuning is performed using labeled vibration signal samples to obtain a vibration domain bearing fault diagnosis model. Based on the feature encoder, a vibration signal-oriented classifier is constructed, including a classification branch for outputting fault types and a regression branch for outputting speed parameters. Supervised fine-tuning training is performed on the feature encoder and classifier using labeled vibration signal samples to classify fault categories. During the supervised fine-tuning training process, the parameters of the shallow convolutional layers of the feature encoder are frozen, and the parameters of the higher-level convolutional layers are unfrozen and trained. S4. Transfer the vibration domain feature encoder to the current signal domain, and train the vibration domain feature encoder and current domain classifier transferred to the current domain based on real labeled current data to obtain the current domain bearing fault diagnosis model. S5. Input the current signal collected under the working condition to be diagnosed into the current domain bearing fault diagnosis model, and output the corresponding bearing fault type and speed information.

2. The twin network based cross-domain locomotive bearing fault diagnosis method according to claim 1, characterized in that The spectrum is generated using variational mode decomposition and fast Fourier transform, and the methods include: Variational mode decomposition is used to decompose the original time-domain signal into several intrinsic mode components by minimizing the following objective function: ； By alternately optimizing, each mode is iteratively updated. and its corresponding center frequency Minimize the objective function until the convergence condition is met, where For unit impact function, For Lagrange multipliers, The original signal to be decomposed. As a penalty factor, For time, It is the imaginary unit.

3. The twin network based cross-domain locomotive bearing fault diagnosis method of claim 2, wherein The model is trained by maximizing the similarity between sample pairs, as follows: ； , These are features output through a twin network; It's a temperature parameter. This refers to the number of samples per batch during training; Sim is... , Cosine similarity between them: ； The final losses are as follows: ； Indicates to The backpropagation of the path feature gradient is interrupted.

4. The twin network based cross-domain locomotive bearing fault diagnosis method of claim 1, wherein During the self-supervised representation learning process of the feature vector using NT-XENT loss, the gradient backpropagation of one branch of the feature vector is cut off.

5. The twin network based cross-domain locomotive bearing fault diagnosis method of claim 1, wherein The training loss is a weighted sum of the cross-entropy loss of the fault type and the regression loss of the rotational speed prediction.

6. The twin network based cross-domain locomotive bearing fault diagnosis method of claim 5, wherein The classifier distinguishes between rotational speed categories and fault categories, and the training loss used is as follows: ； ； in Loss for classifying fault types, SupCon is the loss for classifying rotational speed types, and SupCon is the supervised comparison loss. The loss weights are for classifying rotational speed types. For supervised comparison of loss weights, It's a temperature parameter. For cosine similarity, It is a collection of samples of the same category. It is the temperature coefficient hyperparameter.

7. A cross-domain locomotive bearing fault diagnosis device based on twin networks, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-6.

Images