High-speed train bearing cross-domain fault diagnosis method and system based on transfer learning
By employing transfer learning methods, adaptive denoising and multi-dimensional feature extraction, combined with dynamic weighted domain adversarial networks and pseudo-label generation, the problem of data distribution differences in cross-domain fault diagnosis of high-speed train bearings was solved, achieving high-precision and high-reliability fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-03-16
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies are insufficient to effectively address the lack of labeled data and cross-domain distribution differences in high-speed train bearings in actual operating scenarios, resulting in insufficient generalization ability of traditional machine learning models and failing to meet the requirements of high-precision and high-reliability fault diagnosis.
We employ a transfer learning-based approach, employing adaptive denoising, multi-dimensional feature extraction, and joint standardization to quantify distribution differences and construct a category-dynamically weighted domain adversarial network (CDW-DANN). By combining pseudo-label generation and dual-model probabilistic weighted fusion, we achieve feature distribution alignment and model optimization.
It significantly improves the accuracy and robustness of cross-domain fault diagnosis, effectively utilizes unlabeled data, and ensures the stability of transfer learning and the reliability of diagnostic results.
Smart Images
Figure CN122490296A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bearing fault diagnosis technology, specifically to a method and system for cross-domain fault diagnosis of high-speed train bearings based on transfer learning. Background Technology
[0002] As a core rotating component of the running system, the health status of high-speed train bearings directly affects the safety of train operation. Traditional bearing fault diagnosis methods rely on a large amount of labeled data to train models. However, bearing fault data in actual operation scenarios (target domain) often lacks labels and differs from test bench (source domain) data in terms of speed, noise, and operating conditions, resulting in insufficient model generalization ability.
[0003] In existing technologies, traditional machine learning models struggle to handle cross-domain distribution differences, and conventional transfer learning methods suffer from problems such as domain alignment and classification task imbalance, and the accumulation of pseudo-label errors, failing to meet the high precision and high reliability requirements of high-speed train bearing fault diagnosis. Therefore, there is an urgent need for an intelligent diagnostic method that can effectively eliminate cross-domain differences and fully utilize unlabeled data to achieve knowledge transfer from test bench data to actual operational data. Summary of the Invention
[0004] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by proposing a method and system for cross-domain fault diagnosis of high-speed train bearings based on transfer learning. Through multi-step data processing and model optimization, this invention appropriately solves the core challenges in cross-domain fault diagnosis. To address the above problem, this invention adopts the following technical solution:
[0005] First, this invention proposes a cross-domain fault diagnosis method for high-speed train bearings based on transfer learning, comprising the following steps:
[0006] Step 1: Preprocess the target domain data by adaptive denoising, multi-dimensional feature extraction, and joint standardization with the source domain to eliminate inter-domain scale differences.
[0007] Step 2: Quantify the distribution differences between the source domain and the target domain, and use distribution difference measurement combined with failure mechanism consistency verification;
[0008] Step 3: Construct the Class-Based and Dynamically Weighted Domain Adversarial Network CDW-DANN, introducing class-balanced weighted classification loss and dynamic domain adaptation intensity scheduling, and achieving feature distribution alignment through adversarial learning;
[0009] Step 4: Based on the fusion of dual-source information, generate pseudo-labels iteratively, fuse model prediction confidence and physical mechanism calibration, and iteratively optimize the pseudo-label set and model parameters;
[0010] Step 5: Use a dual-model probabilistic weighted fusion to identify faults in the target domain. Combine the prediction results of CDW-DANN and random forest to output the final diagnostic category with weights.
[0011] Preferably, the specific steps for target domain data preprocessing in step one are as follows:
[0012] (1) Adaptive VMD denoising technology is used to process the vibration signal in the target domain, and the number of decomposed modes is optimized with the goal of minimizing the sample entropy, while retaining the fault impact characteristics;
[0013] (2) Extract peak value, root mean square and kurtosis in time domain features, fault feature frequency amplitude ratio in frequency domain features, and wavelet packet energy entropy and other multi-dimensional features in time and frequency domain features to ensure consistency with source domain features.
[0014] (3) Jointly standardize the features of the source domain and the target domain to eliminate the scale difference between the two domains.
[0015] Preferably, in step two, the distribution difference between the source domain and the target domain is quantified, and the specific steps are as follows:
[0016] (1) Analyze the commonalities between the two domains, including fault mechanism, characteristic dimensions, signal type, and differences, including speed, noise, tag status, and operating conditions;
[0017] (2) The differences in feature distributions are quantified using the maximum mean difference (MMD) and the Wasserstein distance (WD).
[0018] (3) Ensure that the fault mechanisms of the two domains are consistent by verifying the consistency of fault frequency.
[0019] Preferably, in step three, the category-based and dynamically weighted domain adversarial network CDW-DANN is constructed, and the specific steps are as follows:
[0020] (1) Use class-balanced weighted classification loss to mitigate the impact of imbalanced class distribution in the source domain;
[0021] (2) Design a dynamic domain adaptation strength scheduling strategy to achieve a balance between stable learning of source domain knowledge in the early stage of training and strong domain adaptation in the later stage. In the early stage, focus on improving the classification accuracy of the source domain, and then gradually enhance the adversarial strength of the domain discriminator.
[0022] (3) By using the gradient reversal layer, the model can achieve adversarial learning between the feature extractor and the domain classifier, preventing excessive domain adaptation in the later stages of training and resulting in rapid forgetting of source domain knowledge, thus ensuring the stability of the transfer process.
[0023] Preferably, the iterative pseudo-label generation based on dual-source information fusion in step four includes the following specific steps:
[0024] (1) Based on the prediction confidence of the CDW-DANN model, select high-confidence initial pseudo-label samples;
[0025] (2) Combine the spectrum matching score to perform physical mechanism calibration, correct the model prediction bias, and enhance the physical reliability of the pseudo-label;
[0026] (3) Iteratively update the pseudo-label set and model parameters to improve the model's feature representation ability and the reliability of pseudo-labels.
[0027] Preferably, the specific steps of the dual-model probabilistic weighted fusion and target domain fault calibration in step five are as follows:
[0028] (1) Construct a robust random forest model, suppress overfitting through ensemble learning, enhance tolerance to data noise and local perturbations, and complement the CDW-DANN model;
[0029] (2) Based on the model validation error, the fusion weights are determined, and the prediction results of the two models are integrated using a probability-weighted fusion mechanism;
[0030] (3) Label the fault category of the target domain sample and output the diagnosis results and confidence level.
[0031] Meanwhile, this invention proposes a cross-domain fault diagnosis system for high-speed train bearings based on transfer learning, comprising:
[0032] The data preprocessing module is configured to perform the following process: preprocessing the target domain data by adaptive denoising, multi-dimensional feature extraction, and joint standardization with the source domain to eliminate inter-domain scale differences;
[0033] The domain difference measurement module is configured to perform the following process: quantify the distribution difference between the source domain and the target domain, use distribution difference measurement and combine it with failure mechanism consistency verification;
[0034] The Dynamically Weighted Adversarial Network module is configured to perform the following process: constructing a Class-Based and Dynamically Weighted Domain Adversarial Network (CDW-DANN), introducing a class-balanced weighted classification loss and dynamic domain adaptation intensity scheduling, and achieving feature distribution alignment through adversarial learning;
[0035] The dual-source pseudo-label generation module is configured to perform the following process: iteratively generating pseudo-labels based on dual-source information fusion, calibrating the model prediction confidence and physical mechanism fusion, and iteratively optimizing the pseudo-label set and model parameters.
[0036] The dual-model fusion diagnostic module is configured to perform the following process: target domain fault identification is performed using dual-model probabilistic weighted fusion, and the final diagnostic category is output by combining the prediction results of CDW-DANN and random forest.
[0037] Meanwhile, the present invention proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in the present invention.
[0038] Furthermore, the present invention proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method described in the present invention.
[0039] Finally, the present invention provides a computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described in the present invention.
[0040] The present invention adopts the above technical solution and has the following technical effects compared with the prior art:
[0041] (1) Adaptive VMD denoising, multi-dimensional feature extraction, and joint normalization are used to eliminate the scale difference between the source and target domains; MMD and Wasserstein distance are combined to quantize the distribution offset, and the consistency of fault mechanisms is ensured through fault frequency consistency verification. This technique provides a high-quality, alignable data foundation for subsequent migration diagnostics.
[0042] (2) A category-balanced weighted classification loss is introduced to alleviate the problem of imbalanced distribution of source domain categories; a dynamic domain adaptation intensity scheduling strategy is designed to achieve a smooth transition from stable learning of source domain knowledge in the early stage to enhanced domain alignment in the later stage; adversarial learning is combined with a gradient reversal layer to prevent over-adaptation from causing forgetting of source domain knowledge. This network structure significantly improves the accuracy of cross-domain fault diagnosis while ensuring transfer stability.
[0043] (3) Physical mechanism calibration is performed by combining model prediction confidence with spectrum matching score, and highly reliable pseudo-labels are generated iteratively to make full use of unlabeled target domain data; a complementary model of CDW-DANN and random forest is constructed, the fusion weight is determined based on the verification error, and the diagnostic results are output by probability weighted fusion. This mechanism effectively suppresses the accumulation of pseudo-label errors, reduces the risk of single model bias, and significantly improves the robustness and engineering applicability of the diagnostic results. Attached Figure Description
[0044] Figure 1 This is a statistical comparison chart of the distribution of core fault features in the source and target domains involved in the embodiment.
[0045] Figure 2 The examples involve the source domain and target domain feature correlation matrix and inter-domain difference matrix.
[0046] Figure 3 This is a radar chart showing the comprehensive performance evaluation of the transfer learning model and the traditional machine learning model involved in the example.
[0047] Figure 4 The bar chart shows the diagnostic results and confidence levels of each document involved in the example.
[0048] Figure 5 This is a histogram showing the distribution of diagnostic confidence levels in the example.
[0049] Figure 6 This is a flowchart illustrating the overall steps involved in the embodiment. Detailed Implementation
[0050] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
[0051] Example 1: This example proposes a cross-domain fault diagnosis method for high-speed train bearings based on transfer learning, such as... Figure 6 As shown, it includes the following steps:
[0052] Step 1: Preprocess the target domain data by adaptive denoising, multi-dimensional feature extraction, and joint standardization with the source domain to eliminate inter-domain scale differences.
[0053] Step 2: Quantify the distribution differences between the source domain and the target domain, and use distribution difference measurement combined with failure mechanism consistency verification;
[0054] Step 3: Construct the Class-Based and Dynamically Weighted Domain Adversarial Network CDW-DANN, introducing class-balanced weighted classification loss and dynamic domain adaptation intensity scheduling, and achieving feature distribution alignment through adversarial learning;
[0055] Step 4: Based on the fusion of dual-source information, generate pseudo-labels iteratively, fuse model prediction confidence and physical mechanism calibration, and iteratively optimize the pseudo-label set and model parameters;
[0056] Step 5: Use a dual-model probabilistic weighted fusion to identify faults in the target domain. Combine the prediction results of CDW-DANN and random forest to output the final diagnostic category with weights.
[0057] Furthermore, in step one, regarding the preprocessing of the target domain data, VMD denoising technology is first used:
[0058]
[0059] in, The preset total number of modal decompositions, For the first One modality, The modal center frequency, The original signal, This is the Dirac function.
[0060] The optimal number of decomposition modes is determined based on the principle of minimizing sample entropy. Subsequently, 12 multi-dimensional features consistent with the source domain are extracted, including peak value, root mean square, and kurtosis in the time domain, the frequency amplitude ratio of fault features in the frequency domain, and the energy entropy of 8 wavelet packet bands in the time and frequency domains. Finally, the scale difference between the source and target domains is eliminated through joint normalization.
[0061]
[0062] The joint mean is:
[0063]
[0064] The joint variance is:
[0065]
[0066] in, The total number of samples in the source domain. The total number of samples in the target domain. This represents the value of the i-th sample in the source domain on the d-th dimension feature; This represents the value of the j-th sample in the source domain on the d-th dimension feature. Joint standard deviation The square root of the variance .
[0067] Figure 1 The results demonstrate a highly consistent distribution pattern between the source and target domains across key feature dimensions: temporal peak and kurtosis effectively distinguish between fault and normal states, with a significantly higher numerical range for fault samples; the frequency domain fault frequency ratio matches theoretical values across different fault types, exhibiting high cross-domain overlap; and composite features such as peak-kurtosis ratio demonstrate stable discriminative power in both domains. These statistical results indicate that feature extraction from the source to the target domain maintains unified physical meaning and statistical regularity, satisfying the core requirement of feature consistency in transfer learning and providing a reliable data foundation for subsequent domain adaptation algorithms.
[0068] Furthermore, in step two, the commonalities and differences between the two domains are first analyzed.
[0069] Figure 2 The study demonstrates that comparative analysis of the feature correlation matrices of the source and target domains reveals a high degree of consistency in the association patterns among core features. The correlation coefficients of key feature pairs, such as time-domain peak value and kurtosis, and frequency-domain fault frequency ratio and confidence level, are largely consistent across both domains, with relatively small overall fluctuations in the difference matrix. Overall, the feature associations exhibit good cross-domain stability, validating the transferability of the feature system.
[0070] Then, using the maximum mean difference (MMD), let the source domain have... There are samples, and the target domain has... A sample, recorded , ,
[0071]
[0072] in:
[0073]
[0074] It is the square of the Euclidean distance between samples x and y.
[0075] The difference between the Wasserstein distance quantization distribution and the distribution is that P is the source domain feature distribution and Q is the target domain feature distribution, both defined in... The probability distribution on WD, where D is the feature dimension, is expressed mathematically as follows:
[0076]
[0077] in, Let represent the set of joint distributions of P and Q.
[0078] Finally, the consistency of fault frequency calculation logic is verified to ensure that the fault characteristic frequency calculation logic of the two domains is consistent.
[0079] Figure 3 The study presents radar charts across five dimensions—feature alignment, classification accuracy, and generalization performance—to comprehensively evaluate the performance advantages of the dual-model probabilistic weighted fusion model compared to traditional methods. It reveals that CDW-DANN and the dual-model fusion method demonstrate significant advantages over traditional machine learning approaches in multi-dimensional performance comparisons. This method significantly improves feature alignment and domain adaptability through domain adversarial training and dynamic adaptation mechanisms, thereby achieving higher classification accuracy and generalization performance on the target domain test set.
[0080] Furthermore, in step three, a category-based dynamic weighted domain adversarial network (CDW-DANN) is constructed.
[0081] Introducing a category-balanced weighted classification loss, assuming that there are a total of [number] categories in the source domain. The first category, the The number of samples in the class is Then the maximum number of samples in the largest category The basic weight ratio is Weight calculation formula:
[0082]
[0083] Weighted classification loss function:
[0084]
[0085] in, The total number of samples in the source domain. For the first The true label of each sample For the model to the first Predicted label for each sample, For the first Feature vectors of source domain samples This represents the weight corresponding to the label.
[0086] Class weights are assigned inversely based on the number of samples to mitigate the impact of source domain class imbalance.
[0087] Design a dynamic domain adaptive strength scheduling strategy, with the linear scheduling function as follows:
[0088]
[0089] in, The total number of training cycles. For the current cycle, As the initial weights, This is the final weight.
[0090] Perform derivative analysis on it:
[0091]
[0092] Achieve smooth training by "classifying first and then aligning".
[0093] An adversarial framework is constructed by combining a gradient inversion layer, and the domain classification loss uses binary cross-entropy.
[0094]
[0095] in For the output of the domain discriminator, The total number of samples in the source domain. The total number of samples in the target domain. For the first One source domain sample, For the first Samples from the target domain. After dynamic weight scheduling is introduced through CDW-DANN, the following is obtained:
[0096]
[0097] Cross-domain feature distribution alignment is achieved through adversarial training between the feature extractor and the domain classifier.
[0098] Furthermore, in step four, iterative pseudo-tag generation is based on dual-source information fusion.
[0099] Suppose the improved CDW-DANN model is applied to the target domain samples. The predicted output is:
[0100]
[0101] in, Let K be a logits vector, K be the number of classes, and the softmax function be defined as follows:
[0102]
[0103] Where T is the temperature parameter (typically, T=1). Filtering high-confidence initial pseudo-label samples:
[0104]
[0105] Physical mechanism calibration is performed by combining spectrum matching scores to quantify the degree of matching between the sample spectrum and the theoretical fault frequency, and to correct the model prediction bias.
[0106] High-confidence pseudo-label sample set The training set is merged into the source domain training set to form the expanded training set. This will be used to retrain the CDW-DANN model. The pseudo-label generation process can be constructed as a dynamic system.
[0107]
[0108] Among them, the fusion filter function
[0109]
[0110] Iteration can improve the model's ability to represent features.
[0111] Furthermore, in step five, the dual-model (CDW-DANN and Random Forest) probabilistic weighted fusion is used for target domain fault calibration.
[0112] Building a robust random forest model:
[0113]
[0114] in, The total number of decision trees in the random forest. This indicates that the true category of the sample is the first. kind, This is the prediction result of the t-th decision tree, which complements the CDW-DANN model. The fusion weights are determined based on the model validation error, and a probabilistic weighted fusion mechanism is used to integrate the prediction results of the two models for the target domain test samples. Obtain the CDW-DANN model respectively and Random Forest Model Predicted probability vector and The final fusion probability is:
[0115] Final diagnostic label of the sample It is determined by the probability distribution after fusion, that is:
[0116] Reduce the risk of error in a single model.
[0117] like Figure 4 As shown, the diagnostic confidence scores for most normal state files are 0.8 or higher, indicating that the model has high stability in identifying normal states. Although the confidence scores for some fault files are slightly lower, they can still effectively distinguish specific fault types. Overall, the model performs well in classifying different states with good confidence in fault diagnosis of target domain files. Figure 5 In the data, the diagnostic confidence scores were mainly concentrated in the range of 0.83-0.84, with 12 files falling within this range, representing a significant proportion. Overall, the average diagnostic confidence score was 0.822, indicating that the model performed well in terms of diagnostic confidence, and the diagnostic results for most files showed high reliability.
[0118] Example 2: This example proposes a cross-domain fault diagnosis system for high-speed train bearings based on transfer learning, including:
[0119] The data preprocessing module is configured to perform the following process: preprocessing the target domain data by adaptive denoising, multi-dimensional feature extraction, and joint standardization with the source domain to eliminate inter-domain scale differences;
[0120] The domain difference measurement module is configured to perform the following process: quantify the distribution difference between the source domain and the target domain, use distribution difference measurement and combine it with failure mechanism consistency verification;
[0121] The Dynamically Weighted Adversarial Network module is configured to perform the following process: constructing a Class-Based and Dynamically Weighted Domain Adversarial Network (CDW-DANN), introducing a class-balanced weighted classification loss and dynamic domain adaptation intensity scheduling, and achieving feature distribution alignment through adversarial learning;
[0122] The dual-source pseudo-label generation module is configured to perform the following process: iteratively generating pseudo-labels based on dual-source information fusion, calibrating the model prediction confidence and physical mechanism fusion, and iteratively optimizing the pseudo-label set and model parameters.
[0123] The dual-model fusion diagnostic module is configured to perform the following process: target domain fault identification is performed using dual-model probabilistic weighted fusion, and the final diagnostic category is output by combining the prediction results of CDW-DANN and random forest.
[0124] Example 3: This example proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method as described in this invention.
[0125] Example 4: This example proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed, it implements the steps of the method described in this invention.
[0126] Example 5: This example proposes a computer program product, including a computer program / instructions, which, when executed by a processor, implements the steps of the method described in this invention.
[0127] It should be noted that the processing flow of embodiments 2-5 corresponds to the specific steps of the method provided in embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in embodiment 1 of the present invention.
[0128] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0129] The specific implementation schemes described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.
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: Preprocess the target domain data by adaptive denoising, multi-dimensional feature extraction, and joint standardization with the source domain to eliminate inter-domain scale differences. Step 2: Quantify the distribution differences between the source domain and the target domain, and use distribution difference measurement combined with failure mechanism consistency verification; Step 3: Construct the Class-Based and Dynamically Weighted Domain Adversarial Network CDW-DANN, introducing class-balanced weighted classification loss and dynamic domain adaptation intensity scheduling, and achieving feature distribution alignment through adversarial learning; Step 4: Based on the fusion of dual-source information, generate pseudo-labels iteratively, fuse model prediction confidence and physical mechanism calibration, and iteratively optimize the pseudo-label set and model parameters; Step 5: Use a dual-model probabilistic weighted fusion to identify faults in the target domain. Combine the prediction results of CDW-DANN and random forest to output the final diagnostic category with weights.
2. The method according to claim 1, characterized in that, The specific steps for target domain data preprocessing in Step 1 are as follows: (1) Adaptive VMD denoising technology is used to process the vibration signal in the target domain, and the number of decomposed modes is optimized with the goal of minimizing the sample entropy, while retaining the fault impact characteristics; (2) Extract peak value, root mean square and kurtosis in time domain features, fault feature frequency amplitude ratio in frequency domain features, and wavelet packet energy entropy and other multi-dimensional features in time and frequency domain features to ensure consistency with source domain features. (3) Jointly standardize the features of the source domain and the target domain to eliminate the scale difference between the two domains.
3. The method according to claim 1, characterized in that, Step two involves quantifying the distribution differences between the source and target domains. The specific steps are as follows: (1) Analyze the commonalities between the two domains, including fault mechanism, characteristic dimensions, signal type, and differences, including speed, noise, tag status, and operating conditions; (2) The differences in feature distributions are quantified using the maximum mean difference (MMD) and the Wasserstein distance (WD). (3) Ensure that the fault mechanisms of the two domains are consistent by verifying the consistency of fault frequency.
4. The method according to claim 1, characterized in that, Step three involves constructing the category-based and dynamically weighted domain adversarial network CDW-DANN. The specific steps are as follows: (1) Use class-balanced weighted classification loss to mitigate the impact of imbalanced class distribution in the source domain; (2) Design a dynamic domain adaptation strength scheduling strategy to achieve a balance between stable learning of source domain knowledge in the early stage of training and strong domain adaptation in the later stage. In the early stage, focus on improving the classification accuracy of the source domain, and then gradually enhance the adversarial strength of the domain discriminator. (3) By using the gradient reversal layer, the model can achieve adversarial learning between the feature extractor and the domain classifier, preventing excessive domain adaptation in the later stages of training and resulting in rapid forgetting of source domain knowledge, thus ensuring the stability of the transfer process.
5. The method according to claim 1, characterized in that, Step four involves iterative pseudo-label generation based on dual-source information fusion, and the specific steps are as follows: (1) Based on the prediction confidence of the CDW-DANN model, select high-confidence initial pseudo-label samples; (2) Combine the spectrum matching score to perform physical mechanism calibration, correct the model prediction bias, and enhance the physical reliability of the pseudo-label; (3) Iteratively update the pseudo-label set and model parameters to improve the model's feature representation ability and the reliability of pseudo-labels.
6. The method according to claim 1, characterized in that, Step five involves the probabilistic weighted fusion of the two models and the fault calibration in the target domain. The specific steps are as follows: (1) Construct a robust random forest model, suppress overfitting through ensemble learning, enhance tolerance to data noise and local perturbations, and complement the CDW-DANN model; (2) Based on the model validation error, the fusion weights are determined, and the prediction results of the two models are integrated using a probability-weighted fusion mechanism; (3) Label the fault category of the target domain sample and output the diagnosis results and confidence level.
7. A high-speed train bearing cross-domain fault diagnosis system based on transfer learning, characterized in that, include: The data preprocessing module is configured to perform the following process: preprocessing the target domain data by adaptive denoising, multi-dimensional feature extraction, and joint standardization with the source domain to eliminate inter-domain scale differences; The domain difference measurement module is configured to perform the following process: quantify the distribution difference between the source domain and the target domain, use distribution difference measurement and combine it with failure mechanism consistency verification; The Dynamically Weighted Adversarial Network module is configured to perform the following process: constructing a Class-Based and Dynamically Weighted Domain Adversarial Network (CDW-DANN), introducing a class-balanced weighted classification loss and dynamic domain adaptation intensity scheduling, and achieving feature distribution alignment through adversarial learning; The dual-source pseudo-label generation module is configured to perform the following process: iteratively generating pseudo-labels based on dual-source information fusion, calibrating the model prediction confidence and physical mechanism fusion, and iteratively optimizing the pseudo-label set and model parameters. The dual-model fusion diagnostic module is configured to perform the following process: target domain fault identification is performed using dual-model probabilistic weighted fusion, and the final diagnostic category is output by combining the prediction results of CDW-DANN and random forest.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 6.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 6.
10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 6.