Bearing fault diagnosis and interpretable learning method, system, equipment and medium

Abstract

The invention discloses a bearing fault diagnosis and interpretable learning method, system, device and medium, and belongs to the technical field of bearing fault diagnosis and learning, and the method comprises the steps: obtaining and preprocessing source domain bearing data and target domain bearing data, and obtaining a feature data set; constructing diagnosis models according to the feature data set, and screening out an optimal diagnosis model from the diagnosis models; constructing a hierarchical diagnosis and verification framework, and diagnosing the target domain bearing data according to the optimal diagnosis model to obtain a diagnosis result; and observation learning is carried out on a diagnosis result through level analysis. The problems that misdiagnosis occurs in cross-domain diagnosis in the prior art, fault data and normal data are difficult to distinguish, and black box characteristics of deep learning are difficult to distinguish are solved.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis and learning technology, specifically to a bearing fault diagnosis and interpretable learning method, system, device and medium. Background Technology

[0002] Efficient fault detection of bearings, as crucial power components, has always been a challenge. Currently, vibration data of bearings on high-speed trains in transit is mainly obtained from experimental platforms. This data differs somewhat from actual bearing vibration data on high-speed trains, leading to misdiagnosis and incorrect diagnosis by the model during actual fault diagnosis. To avoid this, it is necessary to establish a relationship between the source domain data obtained from the experimental platform and the target domain data of the actual high-speed train.

[0003] High-speed trains, with their significant advantages of safety, efficiency, convenience, comfort, and low carbon emissions, undertake the main transportation tasks in China's passenger transport sector and have become the mainstream mode of transportation in the country. Bearings are crucial rotating components in the transmission system of high-speed trains. Because bearings are subjected to high speeds, alternating loads, impact vibrations, and other extreme and complex operating conditions for extended periods, they exhibit high failure rates and are prone to damage, thus becoming a major cause of equipment failures in the running gear system. Once a high-speed train bearing fails, it can not only lead to train delays but also potentially cause fatal accidents such as derailments, directly endangering the lives of passengers and staff. Therefore, accurate condition monitoring and fault diagnosis of bearings have significant practical importance and safety value.

[0004] As rail transit develops towards high density, strong coupling, and intelligence, traditional diagnostic methods—such as feature indicators built from expert experience or traditional signal processing techniques—face challenges in terms of accuracy, generalization, and real-time performance in complex scenarios. Benefiting from the rapid development of big data and artificial intelligence technologies, deep learning models driven by massive amounts of data have gradually become a research hotspot for intelligent fault diagnosis of train bearings. However, in practical applications, problems such as data noise interference and the scarcity of real fault data due to timely maintenance severely restrict the engineering application of deep learning models. In contrast, bench test data is abundant, fully labeled, and the fault mechanisms are similar to those in reality. Against this backdrop, this paper analyzes bench test data using transfer learning techniques to extract feature values ​​and build a model, which is then transferred to real-world data for application.

[0005] This study aims to construct an efficient and reliable bearing fault transfer diagnosis framework based on deep learning and transfer learning technologies, providing theoretical reference and technical support for its engineering application. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is: how to achieve bearing fault diagnosis and interpretable learning in the process of diagnosis through a bearing fault diagnosis and interpretable learning method, while solving the problems of low diagnostic accuracy, insufficient robustness and difficulty in tracing model decisions caused by the difference in data distribution between the source domain and the target domain.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a bearing fault diagnosis and interpretable learning method, comprising the following steps: acquiring source domain bearing data and target domain bearing data and preprocessing them to obtain a feature dataset; constructing a diagnostic model based on the feature dataset and selecting the optimal diagnostic model from the diagnostic models; constructing a hierarchical diagnosis and verification framework and diagnosing the target domain bearing data based on the optimal diagnostic model to obtain diagnostic results; and performing observational learning on the diagnostic results through hierarchical analysis.

[0009] As a preferred embodiment of the bearing fault diagnosis and interpretable learning method of the present invention, the step of obtaining the feature dataset includes: filtering the source domain bearing data and selecting data with a specific sampling frequency as a data subset; resampling the data subset; performing sliding window segmentation and stratified sampling on the resampled data subset to obtain a feature set; extracting features based on the feature set; and obtaining the feature dataset after standardization.

[0010] As a preferred embodiment of the bearing fault diagnosis and interpretable learning method described in this invention, the step of obtaining the optimal diagnostic model includes: constructing the diagnostic model based on the feature dataset; processing the feature dataset using a feature selector to obtain a model-fitting feature set; preprocessing the model-fitting feature set and inputting it into the diagnostic model to obtain an output result; scoring the output result using performance evaluation metrics, and selecting the diagnostic model with the highest score as the optimal diagnostic model. The beneficial effects of this preferred embodiment are: processing the feature dataset using a feature selector removes redundant and interfering information, improving the quality of the input data; the model-fitting feature set, after preprocessing, adapts to the input requirements of different diagnostic models, reducing the impact of outliers and missing values ​​on the diagnostic results; and the optimal model is selected through multi-dimensional performance evaluation metrics, reflecting the model's comprehensive diagnostic capability.

[0011] As a preferred embodiment of the bearing fault diagnosis and interpretable learning method described in this invention, the step of obtaining the diagnosis result includes: inputting the feature dataset into the preferred diagnostician, alternative diagnostician, and pattern validator in the hierarchical diagnosis and verification framework respectively; judging the output results of the preferred diagnostician, alternative diagnostician, and pattern validator to obtain the diagnosis result.

[0012] As a preferred embodiment of the bearing fault diagnosis and interpretable learning method described in this invention, the step of judging the output result and obtaining the diagnosis result includes: the preferred diagnostic tool calling the optimal diagnostic model and outputting a first diagnostic label and a first confidence level; the alternative diagnostic tool performing feature alignment processing on the feature dataset and inputting it into the optimal diagnostic model, outputting a second diagnostic label and a second confidence level; the pattern validator outputting a third diagnostic label and a third confidence level based on the received target domain feature data; when the first diagnostic label and the second diagnostic label are consistent, the first diagnostic label is taken as the final diagnostic label, and the larger value between the first confidence level and the second confidence level is taken as the final confidence level. Reliability; when the first diagnostic label is inconsistent with the second diagnostic label, the third diagnostic label is compared with the first and second diagnostic labels; when the third diagnostic label is consistent with the first diagnostic label, the first diagnostic label is used as the final diagnostic label; when the third diagnostic label is consistent with the second diagnostic label, the second diagnostic label is used as the final diagnostic label; when the third diagnostic label is inconsistent with both the first and second diagnostic labels, the third label is used as the final diagnostic label; the corresponding confidence level of the final diagnostic label is adjusted and used as the final confidence level; the final diagnostic label and the final confidence level are output as the diagnostic result.

[0013] As a preferred embodiment of the bearing fault diagnosis and interpretable learning method described in this invention, the step of observing the diagnosis results includes analyzing the diagnosis results at the pre-event, during-event, and post-event levels. The beneficial effects of this preferred embodiment are: by progressively analyzing at the pre-event, during-event, and post-event levels, it comprehensively solves the model black box problem; at the pre-event level, it establishes a correspondence between features and the physical mechanism of the fault, providing a theoretical basis for diagnosis; at the during-event level, it presents the cross-domain transfer process through dimensionality reduction visualization, quantifying the fidelity of feature transmission; and at the post-event level, it uses a dual-attribution model to quantify the contribution of single-sample features, ensuring the traceability of the diagnostic basis.

[0014] As a preferred embodiment of the bearing fault diagnosis and interpretable learning method described in this invention, the diagnostic model includes: Q algorithms based on traditional machine learning and a deep learning algorithm based on subsequent algorithms.

[0015] This invention provides a bearing fault diagnosis and interpretable learning system.

[0016] To address the aforementioned technical problems, the present invention further provides the following technical solution: a bearing fault diagnosis and interpretable learning system, comprising: a data acquisition module for acquiring source domain bearing data and target domain bearing data and performing preprocessing to obtain a feature dataset; a model selection module for constructing a diagnostic model based on the feature dataset and selecting the optimal diagnostic model from the diagnostic models; a reliability judgment module for constructing a hierarchical diagnosis and verification framework and diagnosing the target domain bearing data based on the optimal diagnostic model to obtain diagnostic results; and an interpretability analysis module for observing and learning the diagnostic results through layered analysis.

[0017] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the aforementioned bearing fault diagnosis and interpretable learning method.

[0018] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the aforementioned bearing fault diagnosis and interpretable learning method.

[0019] The beneficial effects of this invention are as follows: By processing the source domain bearing data files collected from the test bench, high-quality data from the drive end is selected. The sampling frequency is unified to the target domain through resampling, and the data imbalance problem is solved through sliding window segmentation and stratified sampling. Features are extracted from four dimensions: time domain, frequency domain, time-frequency domain, and envelope domain, constructing a high-quality feature dataset. Subsequently, an integrated diagnostic model set containing nine traditional machine learning and deep learning algorithms is constructed. Random forest is selected as the optimal diagnostic model through multi-dimensional performance index evaluation and cross-validation. Addressing the data distribution differences between the source and target domains, a hierarchical diagnostic and verification framework is proposed, integrating three core components: a preferred diagnostic tool, alternative diagnostic tools, and a pattern validator. Cross-domain fault diagnosis is achieved through consensus decision-making and disagreement arbitration, improving diagnostic confidence. An analysis system is established at the pre-event, during-event, and post-event levels, taking into account both signal physical mechanisms and data-driven modeling. This provides a reliable and interpretable diagnostic tool for on-line health monitoring and intelligent operation and maintenance of train bearings, ensuring the safe operation of high-speed trains. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1This is a general flowchart of a bearing fault diagnosis and interpretable learning method provided in one embodiment of the present invention.

[0022] Figure 2 This is a flowchart illustrating the acquisition of a feature dataset for a bearing fault diagnosis and interpretable learning method, as provided in one embodiment of the present invention.

[0023] Figure 3 This is a flowchart illustrating a model evaluation method for bearing fault diagnosis and interpretable learning, provided as an embodiment of the present invention.

[0024] Figure 4 This is a flowchart of a layered analysis method for bearing fault diagnosis and interpretable learning, provided as an embodiment of the present invention.

[0025] Figure 5 This is a feature importance analysis diagram at the pre-concept level of a bearing fault diagnosis and interpretable learning method provided in one embodiment of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a bearing fault diagnosis and interpretable learning method, including: S100: Obtain source domain bearing data and target domain bearing data and perform preprocessing to obtain feature dataset.

[0028] S200: Construct a diagnostic model based on the feature dataset, and select the optimal diagnostic model from the diagnostic models.

[0029] S300: Construct a hierarchical diagnostic and verification framework, and diagnose the target domain bearing data according to the optimal diagnostic model to obtain diagnostic results.

[0030] S400: Through layer analysis, the diagnostic results are observed and learned.

[0031] It should be noted that existing bearing fault diagnosis technologies often rely on source domain data from experimental platforms to train models. However, there are significant differences between the source domain and the actual train target domain data in terms of sampling rate, noise level, and operating conditions, which can easily lead to misdiagnosis and incorrect diagnosis when conducting cross-domain diagnosis. At the same time, the distribution of fault data and normal data is unbalanced, making it difficult to balance diagnostic accuracy and generalization. Furthermore, the black-box nature of deep learning models makes the diagnostic process lack transparency and makes it impossible to trace the decision-making basis, which seriously restricts the reliable application of bearing fault diagnosis technology in engineering practice.

[0032] Therefore, addressing the issues of misdiagnosis in cross-domain diagnosis, difficulty in distinguishing fault data from normal data, and the black-box nature of deep learning in the existing technologies, this application constructs a bearing fault diagnosis and interpretable learning method through steps S100-S400. First, source domain bearing data and target domain bearing data are acquired and preprocessed to obtain a feature dataset. Second, a diagnostic model is constructed based on the feature dataset, and the optimal diagnostic model is selected from the available models. Then, a hierarchical diagnosis and verification framework is constructed, and the target domain bearing data is diagnosed based on the optimal diagnostic model to obtain diagnostic results. Finally, the diagnostic results are observed and learned through hierarchical analysis.

[0033] Example 2, refer to Figures 1 to 5 This is the second embodiment of the present invention, which provides a bearing fault diagnosis and interpretable learning method.

[0034] In this embodiment of the invention, step S100 involves acquiring source domain bearing data and target domain bearing data and performing preprocessing to obtain a feature dataset, including the following steps A1~A3: A1: Filter the source domain bearing data and select data with a specific sampling frequency as a subset of the data.

[0035] Specifically, the Yuanyu bearing data uses vibration data collected from the test bench to simulate the bearing's entire life cycle operation. The data format is .mat file, covering four typical operating conditions: normal, outer ring failure, inner ring failure, and rolling element failure.

[0036] Furthermore, during data filtering, the 12kHz and 48kHz sampling frequency data collected by the driver-end sensor are selected as the data subset first, and abnormal data segments caused by sensor detachment or electromagnetic interference during the acquisition process are removed.

[0037] A2: Resample a subset of data, then perform sliding window segmentation and stratified sampling on the resampled subset to obtain a feature set.

[0038] Specifically, resampling takes the following forms: ; In the formula, L is the upsampling factor, M is the downsampling factor, h(t) is the filter, and x[n] is the original signal point. This is the output signal.

[0039] It should be noted that the purpose of resampling is to standardize the data subsets with the selected 12kHz and 48kHz sampling frequencies to the target sampling frequency of 32kHz, thereby eliminating the dimensional differences between cross-frequency data.

[0040] The upsampling factor L and downsampling factor M are adaptively determined based on the ratio of the original sampling frequency to the target sampling frequency. When the original sampling frequency is 12kHz, a factor combination of L=8 and M=3 is used, i.e., 12kHz×8 / 3=32kHz; when the original sampling frequency is 48kHz, a factor combination of L=2 and M=3 is used, i.e., 48kHz×2 / 3=32kHz. The filter h(t) is a low-pass filter designed with a Hanning window, and the cutoff frequency is set to 16kHz.

[0041] In one possible implementation, resampling can also be replaced by linear interpolation, which selects two adjacent sampling points in the original signal and calculates the interpolated data at the target sampling frequency based on the linear relationship between the two points; specifically, the interpolation position is determined and the output signal value is calculated based on the ratio of the original sampling time interval to the target sampling time interval.

[0042] In another possible implementation, resampling can also be replaced by sinc interpolation, which is based on the time-domain response function of an ideal low-pass filter. It calculates the output signal with the target sampling frequency by weighted summation of the original signal sampling points using the sinc function.

[0043] Furthermore, the specific manifestations of sliding window segmentation and stratified sampling are as follows: ; In the formula, For the original signal After processing, at discrete time... The signal value, For the discrete-time index currently being processed, For branch index, This is the step size parameter.

[0044] In one possible implementation, sliding window segmentation and hierarchical sampling can also be replaced by fixed-length framing and oversampling. Fixed-length framing uses the same data point window length as sliding window segmentation and directly extracts signal segments in a non-overlapping manner. Oversampling is used for fault samples and increases the number of samples by random repeated sampling or synthesizing minority class samples, so that the proportion of the four types of fault samples tends to be balanced.

[0045] In another possible implementation, sliding window segmentation and stratified sampling can be replaced by adaptive window length segmentation and cluster sampling. Adaptive window length segmentation automatically adjusts the window length according to the fault impact period of the signal and determines the segment boundary by detecting the signal amplitude change point, ensuring that each sub-segment contains complete fault features. The cluster sampling method first performs K-means clustering on samples of each fault type, and then randomly selects the same number of samples from each cluster.

[0046] A3: Extract features from the feature set and obtain the feature dataset after standardization.

[0047] Specifically, the feature extraction adopts a multi-domain collaborative strategy, extracting bearing fault features from four aspects: time domain features, frequency domain features, time-frequency domain features, and envelope domain features.

[0048] Furthermore, the extracted bearing fault features are normalized to obtain a feature dataset.

[0049] In this embodiment of the invention, step S200, which involves constructing a diagnostic model based on a feature dataset and selecting the optimal diagnostic model from the diagnostic models, includes the following steps B1 to B3: B1: Build a diagnostic model based on the feature dataset.

[0050] Specifically, there are Q diagnostic models, and the value of Q is 9.

[0051] Among them, the nine diagnostic models include seven traditional machine learning algorithms and two deep learning algorithms.

[0052] It should be noted that traditional machine learning algorithms include random forest, gradient boosting, support vector machine, extreme random tree, K-nearest neighbor, logistic regression, and Naive Bayes; deep learning algorithms include multilayer perceptron and attention-based neural network algorithms.

[0053] B2: Use a feature selector to process the feature dataset and obtain a feature set that fits the model.

[0054] Specifically, four complementary feature selection methods were selected: principal component analysis (PCA), recursive feature elimination (RFE), univariate feature selection based on the F-statistic (F-Score), and mutual information (MI) to form a feature selection matrix.

[0055] Furthermore, for traditional machine learning algorithms, the RFE+F-Score combination method is adopted to prioritize the retention of features that are strongly correlated with fault labels and fit the model's decision-making logic; for deep learning algorithms, the PCA+MI combination method is adopted to retain the non-linear correlation information between features and labels while reducing dimensionality.

[0056] B3: Preprocess the feature set adapted to the model and input it into the diagnostic model to obtain the output results.

[0057] Specifically, outliers in the feature set are removed using box plots. Then, for features with missing values, K-nearest neighbor interpolation is used to fill in the missing values. Finally, the feature set is standardized according to the input requirements of different diagnostic models. Traditional machine learning models use a two-dimensional feature matrix format, while deep learning models reshape the feature matrix into a three-dimensional tensor.

[0058] Furthermore, the feature set is divided into a training set and a test set according to a fixed ratio, and each algorithm model is trained and tested to obtain the output results.

[0059] B4: The output results are scored using performance evaluation metrics, and the diagnostic model with the highest score is selected as the optimal diagnostic model.

[0060] Specifically, five metrics are selected: accuracy, precision, recall, F1 score, and YIJIAUG score. The output results are scored using an equal-weighted average method, and the algorithm model with the highest score is selected as the optimal diagnostic model.

[0061] It should be noted that this technical solution does not specify a particular optimal diagnostic model.

[0062] In this embodiment of the invention, step S300 involves constructing a hierarchical diagnosis and verification framework, and diagnosing the target domain bearing data according to the optimal diagnosis model to obtain the diagnosis results.

[0063] Specifically, the feature dataset is input into the preferred diagnosticr, alternative diagnosticr, and pattern validator in the hierarchical diagnostic and validation framework.

[0064] Furthermore, the preferred diagnostic tool calls the optimal diagnostic model and outputs the first diagnostic label and the first confidence score; the alternative diagnostic tool performs feature alignment processing on the feature dataset and inputs it into the optimal diagnostic model, outputting the second diagnostic label and the second confidence score. Furthermore, the pattern validator outputs a third diagnostic label and a third confidence score based on the received target domain feature data; Furthermore, when the first diagnostic label and the second diagnostic label are consistent, the first diagnostic label is taken as the final diagnostic label, and the larger value between the first confidence level and the second confidence level is taken as the final confidence level. Furthermore, when the first diagnostic label is inconsistent with the second diagnostic label, the third diagnostic label is compared with the first and second diagnostic labels. Furthermore, when the third diagnostic label matches the first diagnostic label, the first diagnostic label shall be used as the final diagnostic label; Furthermore, when the third diagnostic label matches the second diagnostic label, the second diagnostic label shall be used as the final diagnostic label; Furthermore, when the third diagnostic label is inconsistent with both the first and second diagnostic labels, the third label shall be used as the final diagnostic label. Furthermore, the confidence level corresponding to the final diagnostic label is adjusted and used as the final confidence level; the final diagnostic label and the final confidence level are output as the diagnostic result.

[0065] It should be noted that the first diagnostic label points to the fault category classification, which is used to distinguish whether the equipment has a fault.

[0066] The second diagnostic label points to the detailed fault category, which is used to locate the location of the fault.

[0067] The third diagnostic label indicates the severity of the fault and is used to quantify the stage of fault development.

[0068] Furthermore, the first confidence level is the predicted probability of the diagnostic model outputting the first diagnostic label.

[0069] The second confidence level is the predicted probability of the second diagnostic label output by the diagnostic model. It only takes effect when the first diagnostic label is a fault state, and reflects the reliability of the model's judgment on the location of the fault.

[0070] The third confidence level is the predicted probability of the third diagnostic label output by the diagnostic model. It only takes effect after the second diagnostic label clarifies the fault sub-category and reflects the reliability of the model's judgment on the degree of fault damage.

[0071] In this embodiment of the invention, step S400 involves observation and learning of the diagnostic results through layer analysis.

[0072] Specifically, the diagnostic results are analyzed at the pre-event, during-event, and post-event levels.

[0073] It should be noted that the pre-failure level focuses on the correlation between characteristics and mechanisms. By analyzing the correspondence between characteristics and the physical mechanisms of bearing failure, the fault characterization significance of core characteristics is clarified.

[0074] The specific manifestations of the importance of features analyzed at the pre-hoc level are as follows: ; In the formula, For the first Using features in a tree The set of nodes to be split. To reach the node The sample proportion Features At the node The amount of reduction in impurity of the Gini at the location.

[0075] Specifically, the pre-analysis of Gini impurity manifests in the following ways: ; In the formula, This represents the total number of fault categories. For nodes No. The probability of a sample appearing.

[0076] Furthermore, the specific manifestation of ranking importance is as follows: ; In the formula, For the first 3D features The importance score of the arrangement, The first to be evaluated Dimensional features, To assess the diagnostic performance of the original model on the target samples, Features The total number of times random shuffling is performed. For the k-th shuffle feature Then, the model's diagnostic performance on the target sample.

[0077] Among them, the in-process level focuses on visualizing the cross-domain migration process, quantifying the fidelity of feature transfer between the source and target domains, presenting the impact path of inter-domain differences on diagnostic results, and verifying the cross-domain adaptation effectiveness of the hierarchical diagnostic framework.

[0078] In one possible implementation, in-process analysis can be replaced by feature distribution similarity analysis. By calculating the Bach distance between the feature datasets of the source domain and the target domain, the similarity of the feature distributions of the two domains is quantified. The smaller the Bach distance value, the smaller the distribution difference and the better the cross-domain transfer effect. At the same time, cosine similarity is combined to evaluate the directional consistency of feature vectors and judge the effectiveness of feature transfer between domains.

[0079] In another possible implementation, in-process analysis can be replaced by dynamic monitoring of migration loss. By constructing a domain adaptive loss function, the difference in classification loss between the model in the source domain and the target domain is calculated. When the difference exceeds a set threshold, it indicates that the inter-domain difference has an impact on diagnostic performance. At the same time, feature alignment loss is combined to monitor the degree of model adaptation to features in the two domains.

[0080] In particular, the post-hoc level focuses on the attribution of single-sample diagnoses, identifying the key features and contributions that affect the diagnostic results of individual samples, tracing the core basis of model decisions, and ensuring the interpretability and credibility of diagnostic results.

[0081] In one possible implementation, post-hoc analysis can be replaced by decision path backtracking. For the optimal diagnostic model, this implementation takes the random forest algorithm as an example. It backtracks the splitting path of the target sample in each decision tree, records the key features, splitting threshold and decision direction corresponding to each splitting node, and clarifies the step-by-step reasoning process in which the sample is finally judged as a certain type of fault. At the same time, it counts the frequency of occurrence of key features on each path and quantifies their dominant role in the diagnostic results.

[0082] In another possible implementation, post-hoc analysis can be replaced by feature sensitivity analysis. By keeping other features constant, the features of each dimension of the target sample are perturbed, and the magnitude of change in the model's diagnostic results is recorded. The greater the magnitude of change, the stronger the sensitivity of the feature to the diagnostic results of the sample, which is the influencing feature. By quantifying the correlation between feature perturbation and result change, the degree of contribution of each feature can be clarified.

[0083] In summary, this invention processes source domain bearing data files collected from the test bench, filters high-quality data from the drive end, unifies the sampling frequency to the target domain through resampling, and then solves the data imbalance problem through sliding window segmentation and stratified sampling. Features are extracted from four dimensions: time domain, frequency domain, time-frequency domain, and envelope domain, constructing a high-quality feature dataset. Subsequently, an integrated diagnostic model set containing nine traditional machine learning and deep learning algorithms is constructed. Random forest is selected as the optimal diagnostic model through multi-dimensional performance index evaluation and cross-validation. Addressing the data distribution differences between the source and target domains, a hierarchical diagnostic and verification framework is proposed, integrating three core components: a preferred diagnostic tool, alternative diagnostic tools, and a pattern validator. Cross-domain fault diagnosis is achieved through consensus decision-making and disagreement arbitration, improving diagnostic confidence. An analysis system is established at the pre-event, during-event, and post-event levels, taking into account both signal physical mechanisms and data-driven modeling. This provides a reliable and interpretable diagnostic tool for on-line health monitoring and intelligent operation and maintenance of train bearings, ensuring the safe operation of high-speed trains.

[0084] Example 3 is the third embodiment of the present invention. This embodiment provides a bearing fault diagnosis and interpretable learning system, including...

[0085] The data acquisition module acquires source domain bearing data and target domain bearing data, performs preprocessing, and obtains a feature dataset. The model selection module constructs diagnostic models based on the feature dataset and selects the optimal diagnostic model from the diagnostic models. The reliability assessment module constructs a hierarchical diagnosis and verification framework, and diagnoses the target domain bearing data according to the optimal diagnosis model to obtain the diagnosis results. The interpretability analysis module uses layer analysis to observe and learn from diagnostic results.

[0086] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: if the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0088] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0089] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination of all three. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0090] Example 5, refer to Figures 1 to 5 This is the fifth embodiment of the present invention, which provides a bearing fault diagnosis and interpretable learning method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0091] Resampling section: The data subset consists of 12kHz and 48kHz data collected by the drive unit (DE). The 12kHz and 48kHz data exhibit different characteristic distributions under the same fault condition, forming a natural data augmentation mechanism that makes the diagnostic model more robust and avoids overfitting caused by a single data set. However, the sampling rate of the actual train data in the target domain is 32kHz. This difference in sampling rates between the two sets of data leads to a significant inconsistency in the time scale, impacting subsequent research.

[0092] Inconsistent sampling rates can lead to differences in the comparability of signals in the time and frequency domains. Directly concatenating or comparing source and target domain data will inevitably introduce noise and bias caused by the sampling rate mismatch, further obscuring the signal characteristics for fault diagnosis. This not only weakens the robustness of transfer learning models in extracting universal features but also hinders the effective achievement of the core objective of "reducing inter-domain differences."

[0093] Therefore, to build a fair, reliable, and highly explanatory bearing fault diagnosis model within the transfer learning framework, the sampling rates of the source and target domains must first be made consistent. This is the primary prerequisite for the successful application of transfer learning in the model, and also the reasonable and effective foundation for subsequent feature alignment, fault diagnosis, and transfer learning.

[0094] The core of resampling is to convert a signal from one sampling frequency to another target sampling frequency. The comprehensive resampling model is as follows: ; In the formula, The processed output signal, For the original input signal, L Increase sampling factor, To reduce the sampling factor, For filters.

[0095] For the upsampling factor L and the downsampling factor M, their relationship during the resampling process is as follows: ; In the formula, The sampling rate conversion factor is L>1 and M=1. When L=1 and M>1, simple upsampling increases the sampling rate; when L=1 and M>1, simple downsampling decreases the sampling rate; when both are greater than 1, upsampling and downsampling are combined, representing complex sampling rate conversion.

[0096] For the filter h(t), the function of the filter is to avoid frequency aliasing. During the upsampling and downsampling process, new frequency components may introduce spurious signals, i.e., high-frequency noise or spectral overlap, requiring the use of a low-pass filter for signal smoothing and frequency processing.

[0097] The integrated model is implemented in two steps. First, signal interpolation, or upsampling, is performed. The sampling point density is increased by a factor of L, and each interpolation interval is generated through filtering. Then, signal downsampling is performed. The upsampled signal is sampled every M points, thus achieving a change in the sampling frequency.

[0098] After resampling the driver-end data from a sampling rate of 12kHz to 32kHz, the amplitude of the driver-end signal ranged from -6 to 6, a relatively large range. This amplitude distribution characteristic reflects the presence of a significant impulse component in the signal, indicating that the signal has high complexity. Comparing the signals in the above datasets, this dynamic characteristic is even more pronounced in the B028 data, which not only contains multiple clearly identifiable impulse peaks but also exhibits complex temporal structure features.

[0099] After resampling the fan-end data from a 12kHz sampling rate to 32kHz, it was found that the variation range of this sampling rate was relatively small, ranging from -0.4 to 0.4, reflecting the stability and overall smoothness of the signal. Furthermore, the fan-end signal lacks the high-frequency components of the driver-end signal compared to the driver-end signal, resulting in less pronounced dynamic changes at the fan end.

[0100] This signal characteristic may be due to the fact that only the driver end was selected as the sensor in the data screening, and the transmission path from the fan end to the driver end is relatively long, causing the signal energy to attenuate during the transmission path, or the relevant fault modes are not as obvious as those at the driver end. This also indirectly proves that the data from the DE driver end is the optimal choice of source domain data.

[0101] After resampling the 48kHz driver data to 32kHz, the signal amplitude ranges from -1 to 1, and the amplitude distribution and dynamic characteristics are between the 12kHz resampled data and the fan data.

[0102] The anti-aliasing filter during resampling suppresses high-frequency noise and extreme amplitudes, making the signal energy more concentrated. Compared to the wide fluctuation of -6 to 6 in 12kHz_DE_data, the original 48kHz data itself contains more high-frequency details. After the filtering effectively removes high-frequency noise and abnormal peaks, the fault-related impact characteristics become more prominent.

[0103] After resampling the normal state data at a sampling rate of 48kHz to 32kHz, the signal amplitude ranged from -0.2 to 0.15, representing the smallest amplitude variation among all resampled signals. This amplitude variation indicates that the signal is very stable with minimal fluctuations. Furthermore, the signal is very smooth, appearing as random noise, without any obvious impulse signals or frequency domain phenomena.

[0104] This signal reflects the normal vibration characteristics of the equipment under healthy operating conditions, and this characteristic is further enhanced by anti-aliasing filtering and interpolation algorithms.

[0105] Signal segmentation section: Because the source domain dataset exhibits extremely uneven characteristics, with the number of normal bearings being far less than the number of faulty bearings, this can lead to a severe bias in subsequent fault diagnosis models. Specifically, the machine learning model will predict the more frequently occurring category to minimize the overall loss. This extreme imbalance can cause the model to predict most or all samples as faulty. Therefore, signal segmentation is performed, with sufficient segmentation of the smaller number of normal bearings to maintain a numerical balance with other faulty bearings.

[0106] Sliding window segmentation is a commonly used signal processing and data analysis technique. Its core idea is to extract continuous segments from a time series signal through a fixed window and then slide the window sequentially to cover the complete signal sequence.

[0107] The sliding window segmentation process can be described by the following mathematical model: ; In the formula, the sliding window signal This represents the signal sequence captured within the current k-th window, where w is the data length within each sliding window; a larger w indicates that each window contains more data.

[0108] After sliding window partitioning, the source domain data subset yielded approximately 30,129 windows, which were used to obtain training samples of normal data. However, due to the imbalance between normal and faulty data, stratified sampling was still required for data with different fault types.

[0109] Before balancing, there were only four files of normal data, while the fault data was unevenly distributed, exhibiting a severe class imbalance problem. By using sliding window segmentation and stratified sampling, relative balance of the four classes of data was achieved (Normal: 996, B: 994, IR: 994, OR: 802), providing balanced training data for the subsequent fault diagnosis model.

[0110] Stratified sampling can effectively achieve data balance, avoid the shortcomings of simple random sampling, provide data quality assurance for subsequent model training, and ensure that the model will not be biased due to data imbalance.

[0111] Interpretability at the in-process level: To explain why local decision-making was not used in the model, we first construct an attribution model based on LIME. The idea is that any local region of a given sample can be approximated by a simple, interpretable model. LIME aims to find an interpretable proxy model. ,in It is a collection of interpretable models. The explanation is determined by the following optimization objective function: ; In the formula, This is a fidelity function used to measure the fidelity of samples. Within the neighborhood, the proxy model For the original model The degree of approximation of the prediction results. This function is usually defined as the approximation of the neighboring kernel on a neighborhood perturbation sample set. Weighted squared loss: ; In the formula, Through the The neighborhood sample set generated by perturbation. It is a neighbor kernel function used to assign a value to the nearest neighbor in a neighborhood. More recent samples have higher weights.

[0112] in, This is a regularization term used to penalize the proxy model. The complexity is minimized to ensure good interpretability. For linear models... , The weight vector of the linear model is usually related to the number of non-zero weights and is obtained by solving the equation. This constitutes a sample Attribution explanation for the predictions. Each weighted component. Quantified the first The direction and intensity of each feature's contribution to the prediction.

[0113] Model building section: In the domain-adaptive fault diagnosis task, there is a source domain dataset with sufficient labeled samples and a target domain dataset with only unlabeled feature data. The domain offset problem refers to the inconsistency between the probability distribution of the source domain data and the target domain data. In order to quantify the degree of domain offset, this paper adopts the t-SNE dimensionality reduction visualization method for qualitative observation and introduces the following two statistical indicators for quantitative analysis.

[0114] Mean shift describes the degree of change in the center of a feature distribution. The calculation formula is as follows: ; In the formula, For the first The difference between the mean values ​​of the dimensional features in the target domain and the source domain. The total number of samples in the target domain. Let be the k-th eigenvalue of the j-th sample in the target domain.

[0115] Variance ratio, the variance ratio refers to the change in the variance of the target domain relative to the variance of the source domain. The calculation formula is as follows: ; In the formula, Let be the ratio of the variance of the k-th feature in the target domain to that in the source domain. Let k be the feature value of the j-th sample in the target domain. This represents the total number of samples in the source domain.

[0116] Model solution section: 7.3.1 Detailed Description of the HDVF Hierarchical Decision Algorithm To enable automatic diagnosis of samples in target domains of different levels, we designed a hierarchical decision-making algorithm as shown in Algorithm 1. The entire algorithm integrates the outputs of the three core components into a dynamic and intelligent diagnostic process, guided by the principle of "first seeking similarities, then seeking differences." We will now describe the entire algorithm process in detail.

[0117] Level 1: Concurrent Diagnosis Layer This layer acts as the front-end diagnostic engine of the framework. Its main task is to perform parallel diagnostics on samples in the target domain and extract feature information. The specific steps are as follows.

[0118] Preferred Diagnostic Execution: The algorithm first calls the preferred diagnostic, which feeds the target domain sample features, which have not been modified, into the model trained in the source domain to obtain the predicted labels and confidence scores. This path represents the most direct and effective knowledge transfer.

[0119] Alternate diagnostics execution: At the same time, the algorithm starts the alternative diagnostics. It first transforms the features of the target domain samples using feature alignment to make them more closely aligned with the features of the source domain. It then inputs these features into the same source domain model to obtain the predicted labels and confidence scores. This is the more "conservative" prediction after statistical calibration of the transfer.

[0120] Through this parallelism, the algorithm makes different judgments about the target of the sample at the initial moment, which also lays the foundation for subsequent stream splitting.

[0121] Decision-making path one: consensus decision-making.

[0122] After completing the first layer of diagnosis, the algorithm moves on to the first detection point, which involves processing "simple samples" to quickly complete the algorithm.

[0123] Triggering condition: If the predicted labels of the preferred diagnostic tool and the alternative diagnostic tool are completely identical. Decision logic: The high degree of agreement indicates that the target object's features in the sample are either weakened by domain offset or completely distinguished by the feature alignment strategy. Both approaches yield the same result, so there's no need to trigger more complex arbitration. The final diagnostic label y directly uses the common prediction, and the larger of the two predictions is chosen for the final confidence level. This is because for every two independent inferences, the system's overall confidence in the result should increase, and the higher confidence level should be trusted.

[0124] This approach design enables the rapid and effective processing of a large number of highly certain and diagnostic sample frames, thereby greatly improving overall diagnostic efficiency.

[0125] Decision-making path two: Dispute arbitration.

[0126] Triggering condition: When the two diagnostics in the main diagnostic layer give inconsistent predictions, it indicates that the sample is a difficult sample, that is, the sample is near the boundary or the main attributes of the sample are greatly affected by the neighborhood offset. In this case, the divergence arbitration path is automatically triggered. At this time, the sample is a difficult sample and the algorithm will automatically enter the divergence arbitration path.

[0127] Decision-making logic: The logical approach to decision-making is to introduce a third-party pattern validator to independently arbitrate the decision.

[0128] The initiation mode validator is an algorithm call that computes the predicted label yv and confidence score based on the source domain data's own structure. Independent of the discriminant model's decision boundary, it relies on similarity to the source domain failure mode prototype, providing new support for arbitration—new evidence of the data's own topological structure.

[0129] Enforcing the arbitration rule: The algorithm executes a relatively complete arbitration rule based on the relationship between yv, yp, and ys.

[0130] Scenario 1: When yv = yp. If the pattern validator validates the preferred diagnostic, then the final label is yp. In other words, although feature alignment changes the prediction, from a data structure perspective, the original prediction is correct.

[0131] Case 2: When yv = ys. If the pattern validator detects alternative diagnostics, the final label will be determined. That is, feature alignment is valid and necessary, and its result is evidence of the inherent pattern in the data.

[0132] Scenario 3: Three-way divergence. In rare cases, yp, ys, and yv diverge, resulting in the strongest system uncertainty. The algorithm chooses a predefined "safe" strategy: accepting the result of the alternative diagnostic tool fs. This strategy is based on the fact that, given a clear domain shift and divergent diagnostic results, a deterministic model fs with a domain adaptation strategy is more reliable than a model fp without any domain adaptation strategy.

[0133] A confidence penalty is introduced, meaning that regardless of the arbitration outcome, as long as the path is entered, the final confidence calculation will incorporate a divergence penalty α∈(0,1), resulting in the final confidence score: ; In the formula, It represents the confidence level of the winning party, either fp or fs. This design is crucial in both its practical implications and its implications: it informs users that the diagnosis was obtained through a complex arbitration process, inherently carrying greater uncertainty than a consensus result, and therefore the confidence level should be reduced accordingly.

[0134] After arbitration of disagreements, the HDVF framework enables more in-depth analysis and reliable processing of difficult samples, enhancing the reliability and robustness of the entire diagnostic system.

[0135] Model validation section: To further verify the reliability of the model's predictions, confidence levels were calculated for samples within the target domain using LIME, SHAP, and ensemble methods respectively. Individual samples were compared, and statistical tests were performed on all samples. The evaluation model proposed using the ensemble method demonstrated higher confidence levels for the prediction results compared to LIME and SHAP, exhibiting not only higher confidence but, more importantly, less volatility. This provides a reliability guarantee for the subsequent application of this intelligent diagnostic model in industrial practice, facilitating maintenance personnel to determine the value of diagnostic results based on a more reliable indicator.

[0136] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A bearing fault diagnosis and interpretable learning method, characterized in that: include, Acquire source domain bearing data and target domain bearing data and perform preprocessing to obtain a feature dataset; A diagnostic model is constructed based on the feature dataset, and the optimal diagnostic model is selected from the diagnostic models. A hierarchical diagnostic and verification framework is constructed, and the target domain bearing data is diagnosed based on the optimal diagnostic model to obtain diagnostic results; The diagnostic results are observed and learned through layer analysis.

2. The bearing fault diagnosis and interpretable learning method as described in claim 1, characterized in that: The steps to obtain the feature dataset include: The source domain bearing data is filtered, and data with a specific sampling frequency is selected as a subset of the data; The data subset is resampled, and the resampled data subset is then segmented by a sliding window and subjected to hierarchical sampling to obtain a feature set; Feature extraction is performed based on the feature set, and the feature dataset is obtained after standardization.

3. The bearing fault diagnosis and interpretable learning method as described in claim 2, characterized in that: The steps to obtain the optimal diagnostic model include: The diagnostic model is constructed based on the feature dataset; A feature selector is used to process the feature dataset to obtain a model-fit feature set; All model-adaptive feature sets are preprocessed and input into the diagnostic model to obtain the output results; The output results are scored using performance evaluation metrics, and the diagnostic model with the highest score is selected as the optimal diagnostic model.

4. The bearing fault diagnosis and interpretable learning method as described in claim 3, characterized in that: The steps to obtain diagnostic results include: The feature dataset is input into the preferred diagnostic tool, alternative diagnostic tool, and pattern validator in the hierarchical diagnostic and verification framework, respectively. The output results of the preferred diagnostic tool, alternative diagnostic tool, and pattern validator are evaluated to obtain the diagnostic results.

5. The bearing fault diagnosis and interpretable learning method as described in claim 4, characterized in that: The steps of judging the output results and obtaining the diagnostic results include: The preferred diagnostic tool calls the optimal diagnostic model and outputs a first diagnostic label and a first confidence level. The alternative diagnostic tool performs feature alignment processing on the feature dataset and then inputs it into the optimal diagnostic model, outputting a second diagnostic label and a second confidence level. The pattern validator outputs a third diagnostic label and a third confidence level based on the received target domain feature data; When the first diagnostic label is consistent with the second diagnostic label, the first diagnostic label is taken as the final diagnostic label, and the larger value between the first confidence level and the second confidence level is taken as the final confidence level. When the first diagnostic label is inconsistent with the second diagnostic label, the third diagnostic label is compared with the first diagnostic label and the second diagnostic label. When the third diagnostic label is consistent with the first diagnostic label, the first diagnostic label shall be used as the final diagnostic label; When the third diagnostic label matches the second diagnostic label, the second diagnostic label shall be used as the final diagnostic label; When the third diagnostic label is inconsistent with both the first and second diagnostic labels, the third label shall be used as the final diagnostic label. The final confidence level is obtained by adjusting the corresponding confidence level of the final diagnostic label. The final diagnostic label and the final confidence level are output as the diagnostic result.

6. The bearing fault diagnosis and interpretable learning method as described in claim 5, characterized in that: The steps for observing the diagnostic results include: The diagnostic results are analyzed at the pre-event, during-event, and post-event levels.

7. The bearing fault diagnosis and interpretable learning method as described in claim 6, characterized in that: The diagnostic model includes: Q algorithms based on traditional machine learning and Q algorithms based on later-stage deep learning.

8. A bearing fault diagnosis and interpretable learning system, employing the bearing fault diagnosis and interpretable learning method as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module acquires source domain bearing data and target domain bearing data, performs preprocessing, and obtains a feature dataset. The model selection module constructs a diagnostic model based on the feature dataset and selects the optimal diagnostic model from the diagnostic models. The reliability assessment module constructs a hierarchical diagnosis and verification framework, and diagnoses the target domain bearing data according to the optimal diagnosis model to obtain the diagnosis results. The interpretability analysis module performs observational learning on the diagnostic results through layer analysis.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the bearing fault diagnosis and interpretable learning method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the bearing fault diagnosis and interpretable learning method according to any one of claims 1 to 7.

Images