Bearing fault diagnosis method and system based on transfer learning and interpretable analysis

Abstract

The invention belongs to the related technical field of bearing fault diagnosis, and provides a bearing fault diagnosis method and system based on transfer learning and interpretable analysis, which introduces transfer learning to maintain high diagnosis accuracy under the condition that target domain data lacks labels, introduces an interpretable algorithm to analyze a transfer learning process and decision logic, and improves the diagnosis accuracy. Key problems in the migration process can be recognized, the fault classification accuracy of a target domain is improved, the requirement of bearing fault diagnosis for interpretability is met, and the method is particularly suitable for the field related to equipment safety; according to bearing core fault features and classification accuracy, a fitness function is set, core requirements of bearing fault diagnosis are clarified, blind search of a particle swarm in a meaningless feature space is avoided, algorithm convergence time is shortened, the particle swarm can be guided to preferentially search a feature space strongly related to bearing faults, irrelevant feature interference is avoided, and bearing fault diagnosis accuracy is improved. And the fault feature extraction precision is improved.

Description

Technical Field

[0001] This invention belongs to the technical field of bearing fault diagnosis, and particularly relates to a bearing fault diagnosis method and system based on transfer learning and interpretable analysis. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As the most critical rotating component in the running gear system, the bearings of high-speed trains directly affect the train's operational safety and reliability. Due to their long-term exposure to high speeds, alternating loads, and complex vibrations, they are highly susceptible to damage to the outer rings, inner rings, or rolling elements. Traditional diagnostic methods based on expert experience or signal analysis are prone to failure under varying noise levels and operating conditions.

[0004] While current deep learning models can achieve high-precision recognition on test bench data, the significant differences in signal distribution and operating conditions between real-vehicle operating environments and test bench environments often lead to a sharp drop in accuracy when directly transferring the model for application, resulting in the so-called "inter-domain generalization failure." Furthermore, existing transfer learning methods mostly employ single-domain adaptation strategies, lacking adaptability to bearing failure mechanisms and exhibiting limited cross-domain alignment accuracy. Moreover, black-box models are difficult to interpret in terms of their judgment criteria, hindering their application in safety-critical scenarios. In addition, existing schemes that use optimization algorithms to optimize model parameters generally suffer from fixed inertial weights and a single fitness function, making them prone to getting trapped in local optima, and resulting in insufficient efficiency and specificity in hyperparameter optimization.

[0005] Therefore, there is an urgent need for a bearing fault diagnosis method that can solve the above problems in order to improve the accuracy of bearing fault diagnosis. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a bearing fault diagnosis method and system based on transfer learning and interpretable analysis, introducing an interpretable algorithm to analyze the transfer learning process and decision logic, thereby improving the accuracy of fault classification in the target domain.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a bearing fault diagnosis method based on transfer learning and interpretable analysis, comprising: Source domain test bench data and target domain real vehicle data are acquired, and a source domain fault diagnosis model is trained using the source domain test bench data. In the training of the source domain fault diagnosis model, an improved particle swarm optimization algorithm is used for hyperparameter optimization. The improved particle swarm optimization algorithm is as follows: a fitness function is set according to the core fault mechanism characteristics of the bearing and the classification accuracy, and the inertia weight of the corresponding iteration number is dynamically adjusted according to the deviation between the individual particle fitness and the optimal fitness of the group. Align the feature distribution of the source domain bench data and the target domain real vehicle data, and transfer the source domain fault diagnosis model to the target domain to obtain the target domain fault diagnosis model. The SHAP algorithm is used to perform an interpretability analysis of the migration process from the source domain fault diagnosis model to the target domain fault diagnosis model.

[0008] Secondly, the present invention provides a bearing fault diagnosis system based on transfer learning and interpretable analysis, comprising: The training module is configured to: acquire source domain bench data and target domain real vehicle data, and train a source domain fault diagnosis model using the source domain bench data; in the training of the source domain fault diagnosis model, an improved particle swarm optimization algorithm is used for hyperparameter optimization, and the improved particle swarm optimization algorithm is: setting a fitness function according to the core fault mechanism characteristics of the bearing and the classification accuracy, and dynamically adjusting the inertia weight of the corresponding iteration number according to the deviation between the individual particle fitness and the optimal fitness of the group; The migration module is configured to: align the feature distribution of source domain bench data and target domain real vehicle data, migrate the source domain fault diagnosis model to the target domain, and obtain the target domain fault diagnosis model; The interpretation and analysis module is configured to perform interpretability analysis of the migration process from the source domain fault diagnosis model to the target domain fault diagnosis model using the SHAP algorithm.

[0009] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0010] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0011] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0012] The above one or more technical solutions have the following beneficial effects: In this invention, transfer learning is introduced to maintain high diagnostic accuracy even when target domain data lacks annotation. An interpretable algorithm is introduced to analyze the transfer learning process and decision logic, which can identify key issues in the transfer process, improve the accuracy of fault classification in the target domain, and meet the interpretability requirements of bearing fault diagnosis, making it particularly suitable for fields related to equipment safety. The fitness function is set according to the core fault characteristics of the bearing and the classification accuracy, clarifying the core requirements of bearing fault diagnosis, avoiding blind searching of the particle swarm in meaningless feature spaces, shortening the algorithm convergence time, and guiding the particle swarm to prioritize searching the feature space strongly correlated with bearing faults, avoiding interference from irrelevant features, and improving the extraction accuracy of fault features.

[0013] This invention achieves transfer learning optimization through an adaptive domain adaptation method. Compared with existing single-domain adaptation strategies, it incorporates bearing fault mechanism features for strategy selection and weight allocation, which significantly improves the accuracy and stability of cross-domain feature alignment. It solves the core problem of "inter-domain generalization failure" in traditional transfer learning. The target domain diagnosis accuracy is 16% higher than that of direct transfer and more than 10% higher than that of single-domain adaptation methods.

[0014] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0015] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0016] Figure 1 This is a flowchart of a bearing fault diagnosis method based on transfer learning and interpretable analysis according to Embodiment 1 of the present invention; Figure 2 The multi-class ROC curve and confusion heatmap are shown in Embodiment 1 of the present invention. Figure 3 This is a flowchart of the domain adaptation migration and interpretability verification in Embodiment 1 of the present invention. Detailed Implementation

[0017] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0019] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0020] Example 1 This embodiment discloses a bearing fault diagnosis method based on transfer learning and interpretable analysis, including: Source domain test bench data and target domain real vehicle data are acquired, and a source domain fault diagnosis model is trained using the source domain test bench data. In the training of the source domain fault diagnosis model, an improved particle swarm optimization algorithm is used for hyperparameter optimization. The improved particle swarm optimization algorithm is as follows: a fitness function is set according to the core fault mechanism characteristics of the bearing and the classification accuracy, and the inertia weight of the corresponding iteration number is dynamically adjusted according to the deviation between the individual particle fitness and the optimal fitness of the group. Align the feature distribution of the source domain bench data and the target domain real vehicle data, and transfer the source domain fault diagnosis model to the target domain to obtain the target domain fault diagnosis model. The SHAP algorithm is used to perform an interpretability analysis of the migration process from the source domain fault diagnosis model to the target domain fault diagnosis model.

[0021] In this embodiment, a fitness function is set according to the core fault characteristics of the bearing and the classification accuracy. This clarifies the core requirements for bearing fault diagnosis, avoids the particle swarm blindly searching in a meaningless feature space, shortens the algorithm convergence time, and guides the particle swarm to prioritize searching the feature space that is strongly correlated with the bearing fault, avoiding interference from irrelevant features and improving the extraction accuracy of fault features.

[0022] The following is combined Figure 1 The bearing fault diagnosis method based on transfer learning and interpretable analysis proposed in this embodiment is described in detail: Step S1: Obtain source domain test bench data and target domain vehicle data, and preprocess the source domain test bench data; the preprocessing includes outlier cleaning and feature extraction.

[0023] In step S1, source domain bench data and target domain vehicle data are acquired, and the source domain bench data is preprocessed.

[0024] Outlier handling is crucial during the data preprocessing stage because outliers are observations in the dataset that deviate significantly from other data points, and they can negatively impact data analysis results and the performance of machine learning models.

[0025] In this embodiment, the isolated forest algorithm is used to clean the source domain vibration signal, and the maximum outlier removal rate is strictly limited to 5%.

[0026] Subsequently, feature extraction was carried out in conjunction with the bearing failure mechanism; finally, 17 features were extracted in the time domain, frequency domain, and time-frequency domain. After heatmap redundancy removal, 12 core features were retained. The 12 core features include 5 time domain features: root mean square, peak value, peak-to-peak value, kurtosis, and impulse factor; 4 frequency domain features: spectral centroid, spectral variance, outer ring fault frequency amplitude, and inner ring fault frequency amplitude; and 3 time-frequency domain features: wavelet energy, wavelet entropy, and envelope kurtosis.

[0027] Step S2: Construct a source domain fault diagnosis model based on the preprocessed source domain test bench data, and use an improved particle swarm optimization algorithm to optimize the hyperparameters of the source domain fault diagnosis model.

[0028] In step S2, a source domain fault diagnosis model is constructed based on the preprocessed source domain test bench data, and the hyperparameters of the source domain fault diagnosis model are optimized using the particle swarm optimization algorithm.

[0029] In selecting the fault diagnosis model, after comparing six algorithms including Support Vector Machines and Naive Bayes, Random Forest was chosen as the base classifier. The training set was then re-partitioned, and an improved PSO (Particle Swarm Optimization) algorithm was used to optimize key hyperparameters. After multiple iterations, the optimal hyperparameters were determined. The fault diagnosis model was then repeatedly trained and optimized, resulting in an accuracy improvement of up to 4%. The optimized fault diagnosis model achieved an accuracy of 91% on the source domain test set, with F1-scores ≥ 0.85 for each class and AUC values ​​exceeding 0.95, significantly outperforming comparative models such as SVM and MLP.

[0030] The improved particle swarm optimization algorithm includes a fitness function that adapts dynamic inertia weight adjustment and fault feature correlation.

[0031] The fitness function integrates the classification accuracy of the model with the importance weights of core fault mechanism features (i.e., kurtosis, wavelet energy, and time-frequency entropy), and the formula is as follows:

[0032] in, =0.7 is the weight coefficient, and Acc is the classification accuracy of the source domain fault diagnosis model on the validation set. For the first The global contribution metric of the core failure mechanism characteristics.

[0033] Selection criteria for core fault mechanism features: The three core fault mechanism features (kurtosis, wavelet energy, and time-frequency entropy) in the above formula are selected as the top three key features through global contribution ranking of SHAP values ​​and fault mechanism verification. Analysis shows that these three features contribute more than 60% to the decision-making of bearing fault diagnosis and are the most sensitive indicators for distinguishing between inner ring, outer ring, and rolling element faults. Prioritizing the integration of their importance can improve the targeting of optimization.

[0034] The determination of the core failure mechanism characteristics is as follows: Step 1: Based on the trained source domain fault diagnosis model (such as random forest), calculate the mean absolute value of the SHAP value of each core feature.

[0035]

[0036] in, For the first The first sample The SHAP value of the feature; This represents the number of samples.

[0037] Step 2: Normalize the absolute mean of the SHAP values, mapping the range to [0,1], to obtain the final value. This ensures that feature importance is on the same order of magnitude when participating in fitness calculation.

[0038] The 12 core features were sorted in descending order based on the mean absolute value of the SHAP value of each core feature, and the top 5 candidate features with the highest contribution were initially selected. In this embodiment, the top 5 candidate features are kurtosis, wavelet energy, time-frequency entropy, envelope kurtosis, and outer ring fault frequency amplitude. Then, combined with the physical mechanism of bearing failure, the fault sensitivity and discriminative power of the 5 candidate features were verified one by one: From the perspective of the fault mechanism, failures of the inner ring, outer ring, and rolling elements of the bearing can all lead to the generation of impact components in the vibration signal. Kurtosis, as a core indicator of time-domain impact characteristics, can effectively quantify the non-Gaussianity of the signal. When a fault occurs, the kurtosis value will increase significantly, which is a direct characterization of impact-type faults. The energy of the fault signal will be concentrated in a specific frequency band. Wavelet energy decomposes the signal to different scales through wavelet transform, which can accurately capture the energy distribution distortion caused by the fault. Significant differences exist in the energy concentration frequency bands corresponding to different fault types, demonstrating fault differentiation capabilities. Time-frequency entropy reflects the complexity and disorder of the signal's time-frequency distribution. Normal bearings exhibit a relatively uniform time-frequency distribution (high entropy value), while in fault conditions, the time-frequency distribution tends to concentrate due to the intervention of impact signals (decreasing entropy value). Furthermore, the magnitude of entropy changes caused by different fault types shows clear differences. While envelope kurtosis and outer ring fault frequency amplitude contribute to the diagnosis, the sensitivity of envelope kurtosis depends on the setting of envelope demodulation parameters, resulting in insufficient stability. The outer ring fault frequency amplitude is only sensitive to a single fault type (outer ring fault), exhibiting poor generalization. Neither of these meet the core requirement of high sensitivity to multiple fault types. Considering the global contribution of SHAP values ​​(top 3) and the fault mechanism verification results (sensitive to multiple fault types, clear physical meaning, and strong stability), kurtosis, wavelet energy, and time-frequency entropy are identified as the top 3 key features. Analysis shows that these three features contribute over 60% to the bearing fault diagnosis decision, making them the most sensitive indicators for distinguishing between inner ring, outer ring, and rolling element faults. Prioritizing their integration can improve the targeting of optimization efforts.

[0039] Particle velocity updated to:

[0040] in: No. The particle in the first The velocity vector of the next iteration has the same dimension as the number of hyperparameters. Dynamic inertia weighting improves the optimization targeting and achieves search balance by integrating the fitness function of diagnostic accuracy and the importance of core fault mechanism features; , The learning factor (usually 2.0) represents the following weight of a particle for its own historical best and the global best of the group, respectively. , Indicates in Using random numbers within a certain range increases the randomness of the search and avoids getting trapped in local optima; No. The historical best position of a particle (i.e., the hyperparameter combination that corresponds to the best model performance during the iteration process of that particle). For the first The particle in the first The current position of the next iteration (current hyperparameter combination); The global optimal position of the entire particle swarm (the hyperparameter combination that corresponds to the optimal model performance during all particle iterations).

[0041] The particle adjusts its position based on the updated velocity, which is an iterative update of the hyperparameter combination, as shown in the formula:

[0042] in: No. The particle in the first The new position (new hyperparameter combination) in the next iteration.

[0043] The original linear dynamic inertia weights relied solely on the number of iterations, lacking adaptive feedback on the particle search state, which easily led to over-searching in local areas or under-exploration globally in later stages. The improved version introduces a particle fitness adaptive factor, dynamically adjusting the weights based on iteration progress and the particle's current performance, balancing global exploration with local development capabilities.

[0044] The improved particle swarm optimization algorithm employs an adaptive nonlinear adjustment strategy for its dynamic inertia weights, incorporating feedback on the deviation between individual particle fitness and the swarm's optimal fitness into the iteration count. The optimized formula is as follows:

[0045] in, =0.9、 =0.4 is the upper and lower limits of the inertia weight. This represents the current iteration number. The maximum number of iterations, For the first The particle in the first The fitness value of the next iteration, i.e., the model performance corresponding to the current combination of hyperparameters. For the first The optimal fitness value of the population in the next iteration. To avoid the minimum value where the denominator is 0; adaptive factor Function: When the particle fitness is close to the population optimum When the factor approaches 1, the weights are adjusted according to the original iteration rhythm to strengthen the local search; when the particle fitness deviates significantly from the swarm optimum... When the factor is greater than 1, the weight increases, which helps particles escape local optima and enhances global exploration.

[0046] Step S3: Use the domain adaptation method to align the feature distributions of the source domain and the target domain, and transfer the optimized source domain fault diagnosis model to the target domain to obtain the target domain fault diagnosis model.

[0047] In step S3, addressing the cross-domain diagnostic challenge of "unlabeled and highly differentiated" real-vehicle data in the target domain, this embodiment employs an adaptive CORAL domain adaptation method to achieve transfer learning optimization, rather than a single CORAL algorithm: the bearing fault mechanism features are expanded based on 12 core features obtained from previous screening, with kurtosis, wavelet energy, and time-frequency entropy being the top 3 key fault mechanism features among these 12 core features; by incorporating these 3 key fault mechanism features, a domain adaptation strategy is selected and weights are assigned to the 12 core features: higher alignment weights are given to feature dimensions strongly correlated with the fault mechanism among the 12 core features, with weight coefficients determined by quantifying the contribution of fault features; low weights are given to noise interference features among the 12 features, significantly improving the accuracy and stability of cross-domain feature alignment and solving the problem of insufficient adaptability of traditional single-domain adaptation strategies.

[0048] Comparative validation of CORAL and TCA domain adaptation methods: CORAL improves the overlap between the target domain samples and the source domain by 40% in the t-SNE low-dimensional space by aligning second-order statistics, and the model achieves an accuracy of 81% in the target domain, which is 16% higher than the 65% of direct transfer. Although TCA achieves feature projection, the scarcity of target domain samples leads to a single predicted category, thus validating the applicability advantage of adaptive CORAL.

[0049] Let the source domain characteristic matrix be , The number of samples in the source domain. The feature dimension is ; the target domain feature matrix is , Let the number of samples in the target domain be denoted as , then the covariance matrices of the two domains are as follows:

[0050]

[0051] in: The source domain feature mean vector; The target domain feature mean vector; For the source domain Feature vectors of each sample; For the target domain The feature vector of each sample.

[0052] By performing a weighted linear transformation on the source domain features, with weights assigned based on the importance of fault mechanism features, the second-order statistic is aligned with the target domain. The aligned source domain feature matrix is ​​as follows:

[0053] in: The source domain feature matrix after CORAL alignment; , Let be the inverse square root and square root of the covariance matrix. The fault mechanism feature weight matrix ( ), that is, the first The global contribution of the core feature, with a value range of [0,1], and ).

[0054] Using the aligned source domain feature matrix Based on the original source domain labels, the optimized source domain fault diagnosis model is fine-tuned a second time. The core structure of the source domain fault diagnosis is fixed, and only the feature mapping layer parameters are updated to adapt the source domain fault diagnosis to the feature distribution of the target domain. The feature matrix of the target domain real vehicle data is then used. By inputting the finely tuned source domain fault diagnosis model, and utilizing the cross-domain fault identification capability learned by the source domain fault diagnosis model, fault classification of unlabeled target domain samples can be achieved.

[0055] In this embodiment, a pseudo-label iterative optimization strategy is used to further train the source domain fault diagnosis model. Specifically, target domain samples with a prediction confidence of ≥0.9 from the source domain fault diagnosis model are selected, pseudo-labels are assigned, and the samples are added to the training set. The source domain fault diagnosis model is then fine-tuned again, and the final stable target domain fault diagnosis model is obtained after three iterations.

[0056] Step S4: Verify the interpretability of the target domain fault diagnosis model and migration process using visualization and interpretability analysis tools, and generate a diagnosis report.

[0057] Based on the characteristics of bearing failures, a multi-dimensional analysis is conducted from pre-failure to migration process to post-failure: pre-failure, core features are selected based on the failure mechanism; during the migration process, visualization tools such as t-SNE and U-MAP are used to track the cross-domain feature alignment effect; post-failure, the SHAP tool is used to analyze the decision logic of the target domain model, while covering the validity verification of the migration process and the reliability interpretation of the target domain model.

[0058] The role of the global baseline: The target domain fault diagnosis model applies this baseline to all samples (including those aligned with the source domain). With target domain samples The average predicted probability of a feature is used as a benchmark for decision decomposition to quantify the degree of contribution of a single feature to the deviation of the prediction result.

[0059] in: This is the global baseline value predicted by the fault diagnosis model in the target domain; The total number of samples ( ); For the target domain fault diagnosis model, the first Sample The prediction result is the probability of a certain fault category in the classification task.

[0060] No. In the nth sample SHAP value of each feature This indicates the degree to which the feature contributes to the deviation of the prediction result from the baseline value. For tree models, its calculation is based on the node split contribution of the decision tree:

[0061] in: This is the set of all decision trees in the random forest; The number of decision trees; For the tree In the middle, features For the sample The predicted contribution; Features In the tree The average contribution of the sample to all samples.

[0062] The global contribution of a feature to the decision-making of the fault diagnosis model in the target domain is measured by the mean of the absolute values ​​of the feature SHAP values. The importance formula for core discriminative features (such as kurtosis and wavelet energy) is as follows:

[0063] t-SNE maps high-dimensional features to 2D / 3D space, intuitively showing the overlap of feature distributions between the source and target domains. Its core components include high-dimensional space probability, low-dimensional space probability, and KL divergence objective function.

[0064] High-dimensional features specifically refer to the 12 core features extracted in bearing fault diagnosis (5 time-domain features + 4 frequency-domain features + 3 time-frequency-domain features). These are the original feature dimensions of the source domain samples and the target domain samples, and belong to the core analysis object of cross-domain diagnosis.

[0065] In high-dimensional space, define the target domain sample. With source domain samples The conditional probability of considering a sample as a neighbor characterizes the similarity between samples:

[0066] in: For the sample Gaussian bandwidth; , It represents the feature vector of the sample in a high-dimensional space.

[0067] In low-dimensional space, the t-distribution is used to define the similarity between samples, avoiding the "crowding problem" in high-dimensional space:

[0068] in, The feature vector of the target domain sample in the low-dimensional space after high-dimensional feature mapping. The source domain sample feature vector is generated in the low-dimensional space after high-dimensional feature mapping. After t-SNE mapping, the high-dimensional features in the low-dimensional space are... The feature vector of each sample.

[0069] The optimization objective of t-SNE is to minimize the probability distribution in the high-dimensional space. With low-dimensional space probability distribution The KL divergence makes the low-dimensional distribution approximate the high-dimensional distribution as closely as possible:

[0070] Minimize using gradient descent Update low-dimensional features :

[0071] In the t-SNE feature visualization stage of this embodiment, the core purpose of KL divergence is as an optimization objective function, namely, to optimize the dimensionality reduction process by minimizing the KL divergence between the high-dimensional distribution P and the low-dimensional distribution Q. Its specific function is as follows: Quantitative matching degree: By calculating the KL divergence between the similarity distribution P of high-dimensional features in the source domain and the low-dimensional distribution Q of low-dimensional features in the target domain, the matching degree between high-dimensional feature relationships and low-dimensional visual relationships is accurately measured. Ensuring distribution consistency: By minimizing the KL divergence through gradient descent, the low-dimensional distribution Q is forced to approximate the high-dimensional distribution P as closely as possible. This ensures that the clustering relationships and similarity structures of samples in the low-dimensional space are consistent with those in the high-dimensional space. The core purpose of this consistency is to avoid distortion of the intrinsic correlation of features caused by the dimensionality reduction process, so that the low-dimensional visualization results can truly map the essential laws of high-dimensional features, providing an intuitive and reliable support for subsequent interpretability analysis. From the perspective of the migration process: if the KL divergence is significantly reduced after dimensionality reduction, and the similar fault samples in the source and target domains in the low-dimensional space change from scattered to clustered, it directly proves that cross-domain feature alignment is effective, indicating that the model has captured the core fault features that are not affected by operating conditions and noise. From the perspective of model decision-making: different fault types form independent and clear clusters in low-dimensional space, which can intuitively explain the classification logic of the model, that is, the model achieves category distinction based on the inherent differences in fault characteristics, rather than relying on domain-specific noise; Anomaly analysis explains: If a certain type of sample deviates from its corresponding cluster in a low-dimensional space, the anomaly can be quickly located, providing a visual explanation for the reasons for model misjudgment.

[0072] This process complements the SHAP algorithm: KL divergence and low-dimensional visualization explain "whether the transfer is effective and whether the classification criteria are reasonable", while SHAP value analysis explains "which features the model specifically relies on for decision-making". Together, they constitute the full-link interpretability of "transfer process - model decision-making", meeting the verification requirements of model reliability in safety-critical scenarios.

[0073] This embodiment introduces a multimodal interaction mechanism that supports visualization tools and interpretable analysis tools, enabling the analysis of transfer learning processes and decision-making logic. It achieves semantic-level human-computer interaction and flexible task configuration, which can give the fault diagnosis system a better level of intelligence and human-computer collaboration capabilities, thereby adapting to the needs of various flexible production and service scenarios.

[0074] Example 2 The purpose of this embodiment is to provide a bearing fault diagnosis system based on transfer learning and interpretable analysis, including: The training module is configured to: acquire source domain bench data and target domain real vehicle data, and train a source domain fault diagnosis model using the source domain bench data; in the training of the source domain fault diagnosis model, an improved particle swarm optimization algorithm is used for hyperparameter optimization, and the improved particle swarm optimization algorithm is: setting a fitness function according to the core fault mechanism characteristics of the bearing and the classification accuracy, and dynamically adjusting the inertia weight of the corresponding iteration number according to the deviation between the individual particle fitness and the optimal fitness of the group; The migration module is configured to: align the feature distribution of source domain bench data and target domain real vehicle data, migrate the source domain fault diagnosis model to the target domain, and obtain the target domain fault diagnosis model; The interpretation and analysis module is configured to perform interpretability analysis of the migration process from the source domain fault diagnosis model to the target domain fault diagnosis model using the SHAP algorithm.

[0075] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0076] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0077] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0078] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0079] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0080] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0081] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0082] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0083] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0084] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A bearing fault diagnosis method based on transfer learning and explainable analysis, characterized in that, include: Source domain test bench data and target domain real vehicle data are acquired, and a source domain fault diagnosis model is trained using the source domain test bench data. In the training of the source domain fault diagnosis model, an improved particle swarm optimization algorithm is used for hyperparameter optimization. The improved particle swarm optimization algorithm is as follows: a fitness function is set according to the core fault mechanism characteristics of the bearing and the classification accuracy, and the inertia weight of the corresponding iteration number is dynamically adjusted according to the deviation between the individual particle fitness and the optimal fitness of the group. Align the feature distribution of the source domain bench data and the target domain real vehicle data, and transfer the source domain fault diagnosis model to the target domain to obtain the target domain fault diagnosis model. The SHAP algorithm is used to perform an interpretability analysis of the migration process from the source domain fault diagnosis model to the target domain fault diagnosis model.

2. The bearing fault diagnosis method based on transfer learning and interpretable analysis as described in claim 1, characterized in that, The determination of the core failure mechanism characteristics of the bearing is specifically as follows: Based on the trained source domain fault diagnosis model, the mean absolute value of the SHAP value of each core feature is calculated. The core features are sorted according to the mean absolute value of the calculated SHAP values, and candidate features are selected. The fault sensitivity and discriminative power of the candidate feature data were verified one by one, and the core fault mechanism features of the bearing were obtained by screening.

3. The bearing fault diagnosis method based on transfer learning and interpretable analysis as described in claim 1, characterized in that, The covariance alignment method is used to align the feature distributions of the source domain bench data and the target domain real vehicle data.

4. The bearing fault diagnosis method based on transfer learning and interpretable analysis as described in claim 1, characterized in that, The inertia weight for the corresponding iteration number is dynamically adjusted based on the deviation between the individual particle fitness and the optimal fitness of the population, specifically as follows: ； t= ； in, , These are the upper and lower limits of the inertia weight. This represents the current iteration number. The maximum number of iterations, For the first The particle in the first The fitness value of the next iteration; For the first The optimal fitness value of the population in the next iteration; It is the minimum value; Here, represents the weighting coefficients; Acc represents the classification accuracy of the source domain fault diagnosis model on the validation set. For the first The global contribution metric of the core failure mechanism characteristics.

5. The bearing fault diagnosis method based on transfer learning and interpretable analysis as described in claim 1, characterized in that, Also includes: Based on t-SNE, the effect of cross-domain feature alignment is tracked as follows: In high-dimensional space, the similarity between samples is characterized by the conditional probability that the target domain sample and the source domain sample are regarded as neighbors; In low-dimensional space, the t-distribution is used to define the similarity between samples; By calculating the KL divergence between the similarity distribution of high-dimensional features in the source and target domains and the low-dimensional distribution, the degree of matching between the high-dimensional feature relationships and the low-dimensional visual relationships is measured.

6. The bearing fault diagnosis method based on transfer learning and interpretable analysis as described in claim 1, characterized in that, The optimized source domain fault diagnosis model is fine-tuned by using the aligned source domain feature matrix and the original source domain labels. The core structure of source domain fault diagnosis is fixed, and only the feature mapping layer parameters are updated, so that the source domain fault diagnosis can be adapted to the target domain feature distribution. The feature matrix of real vehicle data in the target domain is input into the fine-tuned source domain fault diagnosis model. The cross-domain fault identification capability learned by the source domain fault diagnosis model is used to classify faults in unlabeled target domain samples.

7. A bearing fault diagnosis system based on transfer learning and interpretable analysis, characterized in that, include: The training module is configured to: acquire source domain bench data and target domain real vehicle data, and train a source domain fault diagnosis model using the source domain bench data; in the training of the source domain fault diagnosis model, an improved particle swarm optimization algorithm is used for hyperparameter optimization, and the improved particle swarm optimization algorithm is: setting a fitness function according to the core fault mechanism characteristics of the bearing and the classification accuracy, and dynamically adjusting the inertia weight of the corresponding iteration number according to the deviation between the individual particle fitness and the optimal fitness of the group; The migration module is configured to: align the feature distribution of source domain bench data and target domain real vehicle data, migrate the source domain fault diagnosis model to the target domain, and obtain the target domain fault diagnosis model; The interpretation and analysis module is configured to perform interpretability analysis of the migration process from the source domain fault diagnosis model to the target domain fault diagnosis model using the SHAP algorithm.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images