A bearing fault diagnosis method based on similarity clustering region migration network
By employing a similarity clustering region migration network method, source domain samples are screened and time-frequency domain features are extracted and dynamic clustering migration is performed. This solves the stability problem of bearing fault diagnosis under multiple working conditions and noisy environments, and achieves fault identification with high reliability and high adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING TECH UNIV
- Filing Date
- 2026-02-24
- Publication Date
- 2026-06-02
AI Technical Summary
Existing bearing fault diagnosis methods are difficult to achieve stable optimization under multiple operating conditions and noise environments. They suffer from problems such as additional bias introduced by noise samples and insufficient category perception, resulting in poor model stability in environments with missing labels and difficulty in adapting to variable speed, variable load and noise scenarios.
A similarity-based clustering region transfer network approach is adopted. By selecting source domain samples, extracting time-frequency domain features, and using a dynamic clustering transfer module, category-aware cross-domain alignment is achieved, baseline offset and noise interference are reduced, and the model's sensitivity to fault features and cross-domain generalization ability are improved.
It significantly improves the accuracy and stable operation of bearing fault diagnosis across various working conditions and noise interference environments, reduces reliance on expert experience, and enhances the robustness of the model under varying working conditions and the adaptability of the diagnostic system.
Smart Images
Figure CN122132835A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a bearing fault diagnosis method based on similarity clustering region migration network. Background Technology
[0002] With the rapid development of industry and intelligent manufacturing, the reliability of high-end equipment such as rotating machinery, wind turbines, aero engines, and CNC machine tools has become a core concern. Bearings, as core transmission components of these devices, directly affect the overall system's operational safety. Traditional fault diagnosis methods rely on signal processing and expert experience, but are ill-suited to the multi-condition, high-noise industrial environments. Therefore, developing an intelligent diagnostic model that can adapt to various operating conditions and noise environments is crucial for improving the predictive maintenance level of industrial equipment. Currently, while transfer learning-based fault diagnosis methods have alleviated the inter-domain discrepancy problem to some extent, significant shortcomings remain. First, existing methods do not perform refined screening and optimization of source and target domain data, leading to the inclusion of noisy samples or distribution anomalies in the training, introducing additional bias and weakening the model's ability to perceive real fault characteristics. Second, existing alignment strategies largely rely on global distribution matching, lacking category-aware local structure alignment. This easily leads to "negative transfer," where target domain samples are incorrectly mapped to non-corresponding category regions in the source domain, causing diagnostic confusion and resulting in poor model stability in label-deficient environments, making stable optimization difficult. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention provides a bearing fault diagnosis method based on a similarity clustering region transfer network. This method filters low-value source domain samples, performs statistical projection correction optimization on target domain samples based on similarity assessment, effectively captures the time-frequency dynamic characteristics of fault signals through a time-frequency domain feature extraction network, and achieves category-aware cross-domain alignment based on a dynamic clustering transfer module. This invention aims to achieve accurate identification and stable operation of bearing fault diagnosis under various working conditions and noise interference environments. It effectively solves the performance degradation problem in variable speed, variable load, and noise scenarios through a knowledge transfer mechanism that does not rely on a large amount of target domain labeled data. Simultaneously, based on adaptive learning capabilities, it deeply mines the deep fault features in vibration signals, significantly improving the diagnostic system's adaptability to unknown working conditions and reducing reliance on expert experience and manual intervention, ultimately achieving a highly reliable mapping from vibration signals to fault types.
[0004] To achieve the above-mentioned technical objectives, the present invention proposes the following technical solution: A bearing fault diagnosis method based on similarity clustering region migration network specifically includes the following steps: S1. Collect vibration signals of bearings under known operating conditions to obtain source domain dataset; collect vibration signals of bearings to be tested to obtain target domain dataset; the vibration signals in the source domain dataset and the target domain dataset are stored in the form of samples. A source domain sample / a target domain sample consists of N vibration signals belonging to the same bearing. S2. Score each source domain sample in the source domain dataset and select the best samples to form the source domain sample set. Based on the distribution pattern of the data in the source domain sample set, perform similarity evaluation and statistical projection correction optimization on the target domain dataset in turn, and use the optimized target domain dataset as the target domain sample set. S3. Input the source domain sample set and the target domain sample set into the time-frequency domain feature extraction network, and perform time-frequency domain feature extraction and fusion on the source domain samples and the target domain samples to obtain the source domain sample fusion feature and the target domain sample fusion feature. S4. Construct a dynamic clustering transfer module, taking source domain sample fusion features and target domain sample fusion features as inputs and fault identification results as output targets; based on the fault identification results, calculate the composite loss function by adjusting dynamic entropy weights, using Gaussian mixture model clustering and Hungarian algorithm to assign pseudo-labels; iteratively update the parameters of the time-frequency domain feature extraction network and the dynamic clustering transfer module based on the composite loss function until training converges; S5. After training, online diagnosis is performed. Based on the trained time-frequency domain feature extraction network and dynamic clustering transfer module, the fault diagnosis of the bearing under test is realized.
[0005] Furthermore, the specific steps of scoring each source domain sample in the source domain dataset and selecting the best samples to form the source domain sample set are as follows: For any source domain sample, calculate its information entropy score and anomaly score respectively; then calculate the total score of the source domain sample based on the information entropy score and anomaly score; sort all source domain samples in the source domain dataset from high to low according to the total score, select the samples with high total scores as excellent samples, and form the source domain sample set.
[0006] Furthermore, the step of performing similarity evaluation and statistical projection correction optimization on the target domain dataset based on the distribution pattern of the data in the source domain sample set, and using the optimized target domain dataset as the target domain sample set, specifically involves: Calculate the arithmetic mean of the means of each source domain sample in the source domain sample set. Arithmetic mean of the standard deviations of each source domain sample And statistically analyze the range of values for the source domain sample mean. The range of values for the standard deviation of the source domain samples The mean of the source domain samples is the mean of N vibration signals in a source domain sample, and the standard deviation of the source domain samples is the standard deviation of the vibration signals in a source domain sample. Calculate any sample in the target domain sample set Mean value of vibration signal and standard deviation ,like or Based on and right The statistical projection correction optimization is expressed by the following formula: ; in, For the mapping function of statistical projection; The target domain sample set is obtained by integrating all target domain samples that meet the range of mean / standard deviation values and all optimized target domain samples.
[0007] Furthermore, the step of inputting the source domain sample set and the target domain sample set into the time-frequency domain feature extraction network, and performing time-frequency domain feature extraction and fusion on the source domain samples and the target domain samples to obtain the source domain sample fusion features and the target domain sample fusion features specifically involves: The time-frequency domain feature extraction network adopts the time-frequency domain U-Net architecture, which includes four convolutional layers, five Hilbert modules, and four linear modules. The inputs and outputs of the four convolutional layers are connected sequentially, and the outputs of the first, second, third, and fourth convolutional layers are connected to the inputs of the second, third, fourth, and fifth Hilbert modules, respectively. The outputs of the five Hilbert modules are also connected sequentially. Input Sample The inputs are fed into the first convolutional layer and the first Hilbert module, respectively; input samples The input data is extracted layer by layer through four convolutional layers, and the downsampled features output by each convolutional layer are used as the input data of the Hilbert module connected to it. Each Hilbert module performs a Hilbert transform on the input data to obtain the frequency domain features of the input data. The frequency domain features and the input data form a dual-channel feature, which is integrated through a built-in gated convolutional layer to obtain the time-frequency domain features of each Hilbert module. The inputs and outputs of the four linear modules are connected sequentially. Each linear module contains two parallel linear layers. The first linear module takes the time-frequency domain features output by the fourth and fifth Hilbert modules as input, processes them through the two parallel linear layers, and then superimposes them in the spatial dimension to form the integrated feature output by the first linear module. The second linear module takes the integrated feature output by the first linear module and the time-frequency domain features output by the third Hilbert module as input, the third linear module takes the integrated feature output by the second linear module and the time-frequency domain features output by the second Hilbert module as input, and the fourth linear module takes the integrated feature output by the third linear module and the time-frequency domain features output by the first Hilbert module as input, and undergoes the same processing as the first linear module. Finally, the integrated features output by the fourth linear module are the final outputs of the entire time-frequency domain feature extraction network. When the input sample is a source domain sample, the final output is the source domain sample fusion feature, and when the input sample is a target domain sample, the final output is the target domain sample fusion feature.
[0008] Furthermore, the construction of the dynamic clustering transfer module, taking source domain sample fusion features and target domain sample fusion features as inputs and fault identification results as the output target, specifically includes: fusing source domain samples into feature sets The input is processed through a fully connected layer in the built-in classifier of the dynamic clustering transfer module to obtain the deep feature set of the source domain. ; After passing through another fully connected layer, the source domain fault identification result set is obtained. ; Target domain sample fusion feature set After the same processing, the deep feature set of the target domain is obtained sequentially. and target domain fault identification result set .
[0009] Furthermore, the calculation of the composite loss function based on the fault identification results, through dynamic entropy weight adjustment, Gaussian mixture model clustering, and Hungarian algorithm for pseudo-label allocation, is specifically as follows: Based on the target domain fault identification result set Calculate the dynamic entropy weights that reflect the uncertainty of the target domain prediction. Its formula is expressed as: ; in, Indicates the total number of samples in the target domain. This represents the total number of real tag categories in the source domain. Indicates Logarithmic function with base 0. This indicates that the target domain classifier predicts the first... Target domain samples Belongs to the label The probability of; Clustering algorithm based on Gaussian mixture model for deep feature sets of the target domain Perform cluster analysis to obtain The clustering results are stored for each deep feature cluster in the target domain; a cost matrix is defined to reflect the degree of matching between the deep feature clusters in the target domain and the true fault categories in the source domain. The number of rows in the cost matrix CM is equal to the number of deep feature clusters in the target domain. The number of columns represents the total number of real tag categories in the source domain. The first in the matrix Line number Column elements Indicates the first All deep features of the target domain in each feature cluster do not belong to the label. The arithmetic average probability; the optimal assignment mapping that minimizes the sum of costs is solved using the Hungarian algorithm, and different target domain pseudo-labels are assigned to each deep feature cluster of the target domain; Define global distribution difference loss , for , Wasserstein distance between them; Define prediction loss , For based on and source domain real tags The multi-class cross-entropy loss function; Define clustering loss , This is the sum of Wasserstein distances between all deep feature clusters in the target domain and their cluster centers; Define the feature region distribution loss , The sum of Wasserstein distances between all deep feature clusters in the target domain and deep feature clusters in the source domain with the same label; the same label means that the pseudo-label of the deep feature cluster in the target domain is equal to the true label of the deep feature cluster in the source domain; the deep feature clusters in the source domain are divided based on the true labels of the source domain. Each source domain deep feature cluster; based on , , , and dynamic entropy weight Calculate the composite loss function Its formula is expressed as: .
[0010] Furthermore, this application also discloses a bearing fault diagnosis system based on a similarity clustering region migration network, which specifically includes: The multi-source vibration data acquisition module is used to collect vibration signals of bearings under known operating conditions to construct a source domain dataset, and at the same time collect vibration signals of bearings to be tested to construct a target domain dataset. The source domain sample screening and target domain distribution correction module is used to score the quality of source domain samples and screen out excellent samples. At the same time, based on the distribution pattern of source domain samples, it performs similarity assessment and statistical projection correction on target domain samples to align the data distribution differences across working conditions. The time-frequency domain feature extraction and fusion module is used to extract and fuse time-frequency domain features of source domain and target domain samples through a time-frequency domain feature extraction network, generating source domain sample fusion features and target domain sample fusion features; The dynamic clustering transfer training and pseudo-label optimization module is used to construct a dynamic clustering transfer module, take fused features as input, and assign pseudo-labels through dynamic entropy weight adjustment, Gaussian mixture model clustering and Hungarian algorithm, calculate the composite loss function, and iteratively update the network parameters until training converges. The online fault diagnosis reasoning and result output module is used to load the trained model, perform feature extraction and transfer reasoning on the real-time vibration signal of the bearing to be tested, and output the fault type diagnosis result.
[0011] An electronic device is also disclosed, comprising a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; The processor is configured to execute, while running the computer program, a bearing fault diagnosis method based on a similarity clustering region migration network as described above.
[0012] A computer-readable storage medium is also disclosed, which stores computer instructions for causing a processor to execute and implement the bearing fault diagnosis method based on a similarity clustering region migration network as described above.
[0013] Based on the above technical solution, the present invention has at least the following beneficial effects: This method can filter low-value source domain samples by calculating source domain sample scores and optimize target domain samples by statistical projection correction based on similarity assessment, thereby reducing baseline offset and noise interference, significantly enhancing data purity and inter-domain consistency, and improving the model's sensitivity to fault characteristics. This method incorporates a time-frequency domain feature extraction network based on the Hilbert module, which can effectively capture the time-frequency dynamic characteristics of fault signals, enhance the robustness of the model under varying operating conditions, improve the discriminativeness of domain-invariant features, and avoid feature loss or confusion caused by a single network structure. This invention also designs a dynamic clustering transfer module to achieve category-aware cross-domain alignment, ensuring that the training process takes into account both global stability and local accuracy, significantly improving diagnostic convergence speed and cross-domain generalization ability, and reducing the risk of mismatch.
[0014] Ultimately, this invention achieves accurate identification and stable operation of bearing fault diagnosis under various working conditions and noise interference environments by designing a similarity clustering region transfer network that includes a source domain sample screening and target domain distribution correction module, a time-frequency domain feature extraction network, and a dynamic clustering transfer module. Through a knowledge transfer mechanism that does not rely on a large amount of target domain labeled data, it effectively solves the performance degradation problem in variable speed, variable load, and noise scenarios. At the same time, based on adaptive learning capabilities, it deeply mines the deep fault features in vibration signals, significantly improving the diagnostic system's adaptability to unknown working conditions and reducing its dependence on expert experience and manual intervention, ultimately achieving a highly reliable mapping from vibration signals to fault types. Attached Figure Description
[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is an overall flowchart of a bearing fault diagnosis method based on similarity clustering region migration network proposed in this invention; Figure 2 This is a schematic diagram of the time-frequency domain feature extraction network in this embodiment; Figure 3 This is a schematic diagram of the dynamic clustering migration module in this embodiment; Figure 4 This is a graph showing the accuracy of bearing fault diagnosis in this embodiment; Figure 5 This is a graph showing the prediction loss versus training loss of the overall prediction model in this embodiment. Figure 6 This is a schematic diagram of the confusion matrix of bearing fault diagnosis results in this embodiment. Detailed Implementation
[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the following description is provided in conjunction with... Figures 1-6 The present invention will be further described in detail below with reference to specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.
[0017] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0018] like Figure 1 As shown in the figure, this embodiment proposes a bearing fault diagnosis method based on a similarity clustering region migration network, which specifically includes the following steps: S1. Collect vibration signals of bearings under known operating conditions to obtain source domain dataset; collect vibration signals of bearings to be tested to obtain target domain dataset; the vibration signals in the source domain dataset and the target domain dataset are stored in the form of samples. A source domain sample / a target domain sample consists of N vibration signals belonging to the same bearing. In this embodiment, the vibration acceleration value of the bearing at a specific time point or time period is selected as the vibration signal reflecting the bearing's operating state. Specifically, in the preferred implementation, step S1 is as follows: Vibration acceleration signals of bearings under known operating conditions in the laboratory were collected by installing vibration signal sensors on the bearings, forming a source domain dataset. Collect vibration acceleration signals of the bearing to be tested to form a target domain dataset. .
[0019] S2. Score each source domain sample in the source domain dataset and select the best samples to form the source domain sample set. In this embodiment, the i-th sample in the source domain dataset is denoted as... ,in accordance with The range of vibration signals is divided into 10 evenly spaced intervals. Then, based on the magnitude of the acceleration, each vibration signal is assigned to a corresponding interval. After dividing the intervals, calculations are performed separately. Information entropy score and anomaly score; Signal entropy score The formula is expressed as: ; ; in, For the sample The probability vector, Indicates the first The number of vibration signals contained in each interval. Let N be a logarithmic function to the base 10, and let N be a logarithmic function. The total number of vibration signals in the middle, Probability vector The first in One element; Information entropy is used to measure the uncertainty of a vibration signal in a source domain sample. The higher the information entropy score, the stronger the certainty and the more significant the features of the vibration signal in the sample. The lower the information entropy score, the weaker the certainty and the less obvious the features of the vibration signal in the sample. Abnormal scores The formula is expressed as: ; ; in, This is the mapping function for the DeepSVDD deep anomaly single-class classification detection model. It is a normalization function. This indicates that DeepSVDD is applied to the source domain dataset. and target domain dataset The center of the hypersphere in the feature space to be divided; express The square of the distance to the center of the hypersphere in the characteristic space; The larger the value, the more likely it is to indicate The less similar the distribution of samples to other source domains is, the less valuable it is for transfer learning; Then, the total score of the source domain sample is calculated based on the information entropy score and the anomaly score. The formula is expressed as: A higher total score indicates that the source domain samples balance data certainty and similarity with other samples, making them excellent samples. All source domain samples in the source domain dataset are sorted from highest to lowest total score, and samples with higher total scores (in this embodiment, the top 50% of samples sorted from largest to smallest total score are selected; this ratio can be adjusted according to actual needs) are chosen as the excellent samples to form the source domain sample set. This completes the selection of source domain samples.
[0020] Next, based on the distribution patterns of the data in the source domain sample set, this application sequentially performs similarity assessment and statistical projection correction optimization on the target domain dataset, and uses the optimized target domain dataset as the target domain sample set; specifically: Calculate the arithmetic mean of the means of each source domain sample in the source domain sample set. Arithmetic mean of the standard deviations of each source domain sample And statistically analyze the range of values for the source domain sample mean. The range of values for the standard deviation of the source domain samples The mean of the source domain samples is the mean of N vibration signals in a source domain sample, and the standard deviation of the source domain samples is the standard deviation of the vibration signals in a source domain sample. Calculate any sample in the target domain sample set Mean value of vibration signal and standard deviation ,like or Based on and right The statistical projection correction optimization is expressed by the following formula: ; in, For the mapping function of statistical projection; The target domain sample set is obtained by integrating all target domain samples that meet the mean / standard deviation range with all optimized target domain samples. After optimization, the target domain samples in the target domain sample set are mathematically aligned with the source domain samples, reducing the impact of baseline offset differences on subsequent transfer learning.
[0021] It should be noted that this invention does not impose strict requirements on the number of source domain sample sets and target domain sample sets; their numbers can be the same or different. Generally speaking, the number depends on the difficulty of signal acquisition in practice: the greater the acquisition difficulty, the fewer the number of usable samples; at the same time, the more samples, the higher the final accuracy tends to be. In this embodiment, the number of source domain sample sets and target domain sample sets are different.
[0022] S3. Input the source domain sample set and the target domain sample set into the time-frequency domain feature extraction network, and perform time-frequency domain feature extraction and fusion on the source domain samples and the target domain samples to obtain the source domain sample fusion feature and the target domain sample fusion feature. In a preferred embodiment, step S3 specifically comprises: like Figure 2 As shown, the time-frequency domain feature extraction network adopts the time-frequency domain U-Net architecture, which includes four convolutional layers, five Hilbert modules, and four linear modules. The inputs and outputs of the four convolutional layers are connected in sequence, and the outputs of the first, second, third, and fourth convolutional layers are connected to the inputs of the second, third, fourth, and fifth Hilbert modules, respectively. The outputs of the five Hilbert modules are also connected in sequence. Input Sample The inputs are fed into the first convolutional layer and the first Hilbert module, respectively; input samples The input data is downsampled through four convolutional layers, and the formula is as follows: ; ; ; ; in, This represents the mapping function of the convolutional layer. , , , These are the downsampled features output by the first, second, third, and fourth convolutional layers, respectively. The downsampled features output by each convolutional layer are then used as input data for the Hilbert modules connected to it. Each Hilbert module performs a Hilbert transform on the input data to obtain the frequency domain features of the input data. The frequency domain features and the input data form a dual-channel feature, which is integrated through a built-in gated convolutional layer to obtain the time-frequency domain features of each Hilbert module. The formula for extracting time-frequency domain features using the first Hilbert module is expressed as follows: ; ; ; in, The mapping function representing the gated convolutional layer. This represents the frequency domain characteristics obtained in the first Hilbert module. The mapping function representing the Hilbert transform, This represents a mapping function that is superimposed along the channel dimension. This is an intermediate feature of the first Hilbert module. This represents the time-frequency domain features output by the first Hilbert module; the processing of the remaining four Hilbert modules is the same as that of the first Hilbert module, each outputting its own time-frequency domain features, denoted as "output time-frequency domain features". , , , ; The inputs and outputs of the four linear modules are connected sequentially. Each linear module contains two parallel linear layers. The first linear module takes the time-frequency domain features output from the fourth and fifth Hilbert modules as input. After processing by the two parallel linear layers, these features are superimposed in the spatial dimension to form the integrated feature output from the first linear module. The formula is as follows: ; ; ; in, This indicates superposition in the spatial dimension. and These represent the output features of the two linear layers, and These represent the first and second linear modules, respectively. The integrated features output by the first linear module; The second linear module takes the integrated features output by the first linear module and the time-frequency domain features output by the third Hilbert module as input. The third linear module takes the integrated features output by the second linear module and the time-frequency domain features output by the second Hilbert module as input. The fourth linear module takes the integrated features output by the third linear module and the time-frequency domain features output by the first Hilbert module as input. After the same processing as the first linear module, they output their respective integrated features. , , ; The final integrated feature output by the fourth linear module That is, the final output of the entire time-frequency domain feature extraction network. When the input sample is a source domain sample, the final output is the source domain sample fusion feature; when the input sample is a target domain sample, the final output is the target domain sample fusion feature.
[0023] In this embodiment, the designed time-frequency domain feature extraction network is first based on a multi-layer convolutional layer structure, which takes into account both feature extraction and progressive downsampling, reducing the number of parameters and computational cost while effectively controlling the risk of overfitting. The introduction of a Hilbert module enhances the signal representation capability by constructing a dual-channel input containing both the original signal and its transformed features, providing a more comprehensive time-frequency domain information foundation for subsequent analysis. The linear module achieves effective feature complementarity and global enhancement in a unified spatial dimension. The extracted and integrated time-frequency domain features of the source / target domain samples thus lay a solid data foundation for improving the accuracy of bearing fault diagnosis.
[0024] S4, construct as follows Figure 3 The dynamic clustering transfer module shown takes source domain sample fusion features and target domain sample fusion features as inputs, and the fault identification result as the output target, specifically:
[0025] fusing source domain samples into feature sets The input is processed through a fully connected layer in the built-in classifier of the dynamic clustering transfer module to obtain the deep feature set of the source domain. ; After passing through another fully connected layer, the source domain fault identification result set is obtained. ; Target domain sample fusion feature set After the same processing, the deep feature set of the target domain is obtained sequentially. and target domain fault identification result set .
[0026] Based on the fault identification results, a composite loss function is calculated by adjusting the dynamic entropy weights, using Gaussian mixture model clustering and Hungarian algorithm to assign pseudo-labels. The parameters of the time-frequency domain feature extraction network and the dynamic clustering transfer module are iteratively updated based on the composite loss function until training converges. In this preferred embodiment, the specific process for calculating the composite loss function is as follows: Based on the target domain fault identification result set Calculate the dynamic entropy weights that reflect the uncertainty of the target domain prediction. Its formula is expressed as: ; in, Indicates the total number of samples in the target domain. This represents the total number of real tag categories in the source domain. Indicates Logarithmic function with base 0. This indicates that the target domain classifier predicts the first... Target domain samples Belongs to the label The probability of; Clustering algorithm based on Gaussian mixture model for deep feature sets of the target domain Perform cluster analysis to obtain The clustering results are stored for each deep feature cluster in the target domain; a cost matrix is defined to reflect the degree of matching between the deep feature clusters in the target domain and the true fault categories in the source domain. The number of rows in the cost matrix CM is equal to the number of deep feature clusters in the target domain. The number of columns represents the total number of real tag categories in the source domain. The first in the matrix Line number Column elements Indicates the first All deep features of the target domain in each feature cluster do not belong to the label. The arithmetic mean probability; the formula for the cost matrix CM is expressed as: ; in, For the first The total number of samples contained in a cluster. For the first No. 1 in the cluster Each sample belongs to the label The predicted probability; the optimal assignment mapping that minimizes the sum of costs is solved using the Hungarian algorithm, and different target domain pseudo-labels are assigned to each deep feature cluster of the target domain; in this embodiment, the formula for finding the minimum sum of costs is expressed as: ; in, It is an injective mapping function. It is the cost matrix No. Line number Column elements.
[0027] In this embodiment, calculating the dynamic entropy weights and assigning target domain pseudo-labels are both for calculating the composite loss function used to train the time-frequency domain feature extraction network and the dynamic clustering transfer module. This function consists of the global distribution difference loss, prediction loss, clustering loss, and feature region distribution loss, specifically: Define global distribution difference loss , for , Wasserstein distance between them; It should be noted here that the formula for the Wasserstein distance is as follows: ; in, and Represents any two sets of data, express and Wasserstein distance between them Represents data group The total number of samples, Represents data group The total number of samples, Represents data group The One sample, Represents data group The One sample; , Substituting it as the independent variable, we get = ; Define prediction loss , For based on and source domain real tags The multi-class cross-entropy loss function is expressed as follows: ; in, Indicates the batch size. Represents the logarithmic function with base 2. Indicates the prediction of the first Each sample belongs to the label The probability, Indicates the first Labels of each sample The numerical value; Define clustering loss , The sum of Wasserstein distances between all deep feature clusters in the target domain and their cluster centers is expressed by the formula: ; in, This represents the set of cluster center locations for each target domain's deep feature clusters. Indicates the first A target domain deep feature cluster, Indicates the first The cluster center location of each target domain deep feature cluster; Define the feature region distribution loss , The sum of Wasserstein distances between all deep feature clusters in the target domain and deep feature clusters in the source domain with the same label; the same label means that the pseudo-label of the deep feature cluster in the target domain is equal to the true label of the deep feature cluster in the source domain; the deep feature clusters in the source domain are divided based on the true labels of the source domain. Each source domain deep feature cluster; The formula is expressed as: ; in, Indicates the first Each source domain deep feature cluster, its true label and The pseudo-tags are the same; based on , , , and dynamic entropy weight Calculate the composite loss function Its formula is expressed as: .
[0028] In this embodiment, dynamic entropy weight It is based on the determination of the target domain classification results: when the target domain classifier's prediction of the target domain is full of uncertainty, Smaller; when the target domain classifier has high certainty in predicting the target domain, Larger.
[0029] Among the loss functions mentioned above, the global distribution difference loss is... This is used to characterize the similarity between the source domain fault identification results and the target domain fault identification results, aiming to minimize the overall difference in the output distribution between the two domains; feature region distribution loss. The difference between pseudo-labeled target domain samples with the same label and real-labeled source domain samples is calculated to achieve conditional distribution alignment guided by pseudo-labels, ensuring the accuracy of category transfer across the source domain to the target domain; clustering loss. It operates within deep feature clusters of each target domain, aiming to improve the intra-class compactness and structural clarity of the target domain features by minimizing the sum of feature distances between samples within the cluster; prediction loss. The aim is to utilize the supervisory information from the real labels in the source domain to optimize the basic discriminative capabilities of the time-frequency domain feature extraction network and the source domain classifier.
[0030] In the early stages of training, because the model is not yet familiar with the data in the target domain, its predictions are often vague and uncertain. On the one hand, vague predictions can lead to unreliable pseudo-labels that may differ significantly from the true target domain labels. On the other hand, vague predictions result in smaller dynamic entropy weights, which reduces the proportion of feature region distribution loss based on pseudo-labels in the composite loss function. This lowers the requirement for fine-grained category alignment and instead emphasizes that the overall data distribution of the two domains should roughly converge first, preventing feature region distribution loss from misleading the model in the early stages of training.
[0031] As training progresses, the model's predictions become more confident, leading to a gradual increase in the dynamic entropy weight and a reduction in the proportion of global distribution difference loss. While global distribution difference loss can reduce the overall difference between datasets, it may cause deep features of the target domain for a certain fault type to be incorrectly aligned with deep features of the source domain for different fault types. Therefore, in the later stages of training, feature region distribution loss should take precedence.
[0032] The changes in the proportions of various losses during training are not linear, but rather based on the changes in the certainty of the classification results caused by changes in accuracy. This application uses dynamic entropy weights, a quantity that measures the magnitude of certainty, to automatically adjust the proportions of feature region distribution loss and global distribution difference loss in the composite loss function.
[0033] S5. After training, online diagnosis is performed. Fault diagnosis of the bearing under test is achieved based on the trained time-frequency domain feature extraction network and dynamic clustering transfer module. In this embodiment, for the vibration signal samples of the bearing under test... After passing through a time-frequency domain feature extraction network, the following is obtained: Sample fusion features Then The pre-trained dynamic clustering transfer module is input, and its built-in target domain classifier outputs samples. The probability of different fault types is considered, and the fault type with the highest probability is selected as the fault diagnosis result for the bearing.
[0034] This concludes the description of the entire process of the method proposed in this invention.
[0035] Furthermore, this application also discloses a bearing fault diagnosis system based on a similarity clustering region migration network, which specifically includes: The multi-source vibration data acquisition module is used to collect vibration signals of bearings under known operating conditions to construct a source domain dataset, and at the same time collect vibration signals of bearings to be tested to construct a target domain dataset. The source domain sample screening and target domain distribution correction module is used to score the quality of source domain samples and screen out excellent samples. At the same time, based on the distribution pattern of source domain samples, it performs similarity assessment and statistical projection correction on target domain samples to align the data distribution differences across working conditions. The time-frequency domain feature extraction and fusion module is used to extract and fuse time-frequency domain features of source domain and target domain samples through a time-frequency domain feature extraction network, generating source domain sample fusion features and target domain sample fusion features; The dynamic clustering transfer training and pseudo-label optimization module is used to construct a dynamic clustering transfer module, take fused features as input, and assign pseudo-labels through dynamic entropy weight adjustment, Gaussian mixture model clustering and Hungarian algorithm, calculate the composite loss function, and iteratively update the network parameters until training converges. The online fault diagnosis reasoning and result output module is used to load the trained model, perform feature extraction and transfer reasoning on the real-time vibration signal of the bearing to be tested, and output the fault type diagnosis result.
[0036] To verify the performance of the method proposed in this invention, the following specific examples are also provided in this embodiment: This method was implemented on the publicly available bearing failure dataset provided by CWRU and the PT bearing failure dataset collected in the laboratory. The CWRU dataset contains bearing vibration acceleration signals acquired by placing an accelerometer above the bearing housing at the motor drive end under a sampling frequency of 12 kHz. All failed bearings in the CWRU dataset were subjected to single-point damage machining using electrical discharge machining. The PT dataset contains bearing vibration acceleration signals acquired by placing an accelerometer above the bearing housing at the motor drive end of a PT500mini mechanical failure simulation test bench under a sampling frequency of 48 kHz. All failed bearings in the PT dataset were subjected to damage machining using manual cutting. The bearing types collected in both the CWRU and PT datasets are three types of failed bearings and one type of normal bearing. The failure locations of the failed bearings are the outer ring, rolling elements, and inner ring, with failure diameters of 0.18, 0.36, and 0.54 mm. All data files were sampled, with each sample having a signal length of 2048. Specifically, Table 1 below shows all bearing failure types and their corresponding labels: Table 1. Fault types and labels in the CWRU and PT datasets
[0037] Table 2 below shows the working conditions and corresponding working condition labels for the CWRU dataset; Table 2. Operating conditions and labels of the CWRU dataset
[0038] Table 3 below shows the operating conditions and corresponding labels for the PT dataset: Table 3 PT Dataset Operating Conditions and Operating Condition Labels
[0039] For example, the operating condition label C0 represents the CWRU dataset collected under the operating condition of 0 hp load and 1797 r / min speed. The operating condition label N0 represents the PT dataset collected under the operating condition of 0 hp load and 500 r / min speed.
[0040] In the experiment, this invention uses the CWRU public bearing fault dataset as the training set and the PT bearing fault dataset as the validation and test sets. The training and validation sets are processed using a similarity processing module; the training and validation sets are mapped using a time-frequency domain U-Net; the time-frequency domain U-Net, source domain classifier, and target domain classifier are trained using the mapped training and validation sets; finally, the trained time-frequency domain U-Net is mapped to the test set, and the fault diagnosis accuracy is verified on the trained target domain classifier using the mapped test set.
[0041] This invention sets the training iterations to 100, the batch size to 64, and the learning rate to an initial Adam algorithm of 0.001. The mapped bearing samples output by the trained time-frequency domain U-Net are used as input to the target domain classifier. The output of the target domain classifier is the probability that the input sample belongs to different labels. Table 1 shows the bearing fault type corresponding to the label with the highest probability, i.e., the bearing fault identification result. To demonstrate the robustness and generality of the model, Gaussian noise with a signal-to-noise ratio of 0 dB was artificially added to the training, validation, and test sets in the experiment to simulate a real industrial environment. Specifically, the experimental results are shown in Table 4 below: Table 4. Comparison of accuracy of different models across different operating conditions and noisy environments.
[0042] In Table 4, C0→N0 indicates that the training set uses the CWRU dataset collected under operating conditions of 0 hp load and 1797 r / min speed, while the validation and test sets use the PT dataset collected under operating conditions of 0 hp load and 500 r / min speed. As shown in Table 4, the accuracy of the model in this invention is higher than other intelligent models (CNN, BDTN, DANN, DDAN) across different operating conditions and in noisy environments.
[0043] As an explanation, Figure 4 , Figure 5 and Figure 6 This demonstrates the changes in accuracy, loss, and confusion matrix of the similarity clustering region transfer network during training under the C2→N2 condition. Figure 4 It can be seen that the training accuracy of the model of this invention can be higher than 96%; Figure 5 It can be seen that the training loss of the model in this invention can be lower than 0.1; from Figure 6 As can be seen, the model of this invention has a high diagnostic accuracy for all types of faults. Therefore, the bearing fault method proposed in this invention is reliable.
[0044] In summary, the method proposed in this invention solves the problems of insufficient extraction of fault features when there is a lack of bearing fault data, poor versatility of traditional bearing fault diagnosis methods, and low accuracy of bearing fault diagnosis. It can output more accurate bearing fault diagnosis results and can apply the fault diagnosis model trained based on known bearing data to the fault diagnosis of bearings with unknown data, thereby improving the versatility and accuracy of the bearing fault diagnosis model.
[0045] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.
[0046] 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-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).
[0047] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A bearing fault diagnosis method based on similarity clustering region migration network, characterized in that, Specifically, the following steps are included: S1. Collect vibration signals of bearings under known operating conditions to obtain source domain dataset; collect vibration signals of bearings to be tested to obtain target domain dataset; the vibration signals in the source domain dataset and the target domain dataset are stored in the form of samples. A source domain sample / a target domain sample consists of N vibration signals belonging to the same bearing. S2. Score each source domain sample in the source domain dataset and select the best samples to form the source domain sample set. Based on the distribution pattern of the data in the source domain sample set, the target domain dataset is sequentially evaluated for similarity and optimized by statistical projection correction, and the optimized target domain dataset is used as the target domain sample set. S3. Input the source domain sample set and the target domain sample set into the time-frequency domain feature extraction network, and perform time-frequency domain feature extraction and fusion on the source domain samples and the target domain samples to obtain the source domain sample fusion feature and the target domain sample fusion feature. S4. Construct a dynamic clustering transfer module, taking source domain sample fusion features and target domain sample fusion features as inputs and fault identification results as output targets; based on the fault identification results, calculate the composite loss function by adjusting dynamic entropy weights, using Gaussian mixture model clustering and Hungarian algorithm to assign pseudo-labels; iteratively update the parameters of the time-frequency domain feature extraction network and the dynamic clustering transfer module based on the composite loss function until training converges; S5. After training, online diagnosis is performed. Based on the trained time-frequency domain feature extraction network and dynamic clustering transfer module, the fault diagnosis of the bearing under test is realized.
2. The bearing fault diagnosis method based on similarity clustering region migration network according to claim 1, characterized in that, The specific steps for scoring each source domain sample in the source domain dataset and selecting the best samples to form the source domain sample set are as follows: For any source domain sample, calculate its information entropy score and anomaly score respectively; then calculate the total score of the source domain sample based on the information entropy score and anomaly score; sort all source domain samples in the source domain dataset from high to low according to the total score, select the samples with high total scores as excellent samples, and form the source domain sample set.
3. The bearing fault diagnosis method based on similarity clustering region migration network according to claim 1, characterized in that, Based on the distribution patterns of data in the source domain sample set, the target domain dataset is sequentially subjected to similarity evaluation and statistical projection correction optimization, and the optimized target domain dataset is used as the target domain sample set. Calculate the arithmetic mean of the means of each source domain sample in the source domain sample set. Arithmetic mean of the standard deviations of each source domain sample And statistically analyze the range of values for the source domain sample mean. The range of values for the standard deviation of the source domain samples The mean of the source domain samples is the mean of N vibration signals in a source domain sample, and the standard deviation of the source domain samples is the standard deviation of the vibration signals in a source domain sample. Calculate any sample in the target domain sample set Mean value of vibration signal and standard deviation ,like or Based on and right The statistical projection correction optimization is expressed by the following formula: ； in, For the mapping function of statistical projection; The target domain sample set is obtained by integrating all target domain samples that meet the range of mean / standard deviation values and all optimized target domain samples.
4. The bearing fault diagnosis method based on similarity clustering region migration network according to claim 1, characterized in that, The step of inputting the source domain sample set and the target domain sample set into the time-frequency domain feature extraction network, and performing time-frequency domain feature extraction and fusion on the source domain samples and the target domain samples to obtain the source domain sample fusion feature and the target domain sample fusion feature is as follows: The time-frequency domain feature extraction network adopts the time-frequency domain U-Net architecture, which includes four convolutional layers, five Hilbert modules, and four linear modules. The inputs and outputs of the four convolutional layers are connected sequentially, and the outputs of the first, second, third, and fourth convolutional layers are connected to the inputs of the second, third, fourth, and fifth Hilbert modules, respectively. The outputs of the five Hilbert modules are also connected sequentially. Input Sample The inputs are fed into the first convolutional layer and the first Hilbert module, respectively; input samples The input data is extracted layer by layer through four convolutional layers, and the downsampled features output by each convolutional layer are used as the input data of the Hilbert module connected to it. Each Hilbert module performs a Hilbert transform on the input data to obtain the frequency domain features of the input data. The frequency domain features and the input data form a dual-channel feature, which is integrated through a built-in gated convolutional layer to obtain the time-frequency domain features of each Hilbert module. The inputs and outputs of the four linear modules are connected in sequence. Each linear module contains two parallel linear layers. The first linear module takes the time-frequency domain features output by the fourth and fifth Hilbert modules as input. After being processed by the two parallel linear layers, they are superimposed in the spatial dimension as the integrated features output by the first linear module. The second linear module takes the integrated features output by the first linear module and the time-frequency domain features output by the third Hilbert module as input, the third linear module takes the integrated features output by the second linear module and the time-frequency domain features output by the second Hilbert module as input, and the fourth linear module takes the integrated features output by the third linear module and the time-frequency domain features output by the first Hilbert module as input, and undergoes the same processing as the first linear module. Finally, the integrated features output by the fourth linear module are the final outputs of the entire time-frequency domain feature extraction network. When the input sample is a source domain sample, the final output is the source domain sample fusion feature, and when the input sample is a target domain sample, the final output is the target domain sample fusion feature.
5. The bearing fault diagnosis method based on similarity clustering region migration network according to claim 1, characterized in that, The construction of the dynamic clustering transfer module, taking source domain sample fusion features and target domain sample fusion features as inputs and fault identification results as the output target, specifically involves: fusing source domain samples into feature sets The input is processed through a fully connected layer in the built-in classifier of the dynamic clustering transfer module to obtain the deep feature set of the source domain. ; After passing through another fully connected layer, the source domain fault identification result set is obtained. ; Target domain sample fusion feature set After the same processing, the deep feature set of the target domain is obtained sequentially. and target domain fault identification result set .
6. The bearing fault diagnosis method based on similarity clustering region migration network according to claim 5, characterized in that, The calculation of the composite loss function based on the fault identification results, through dynamic entropy weight adjustment, Gaussian mixture model clustering, and Hungarian algorithm for pseudo-label allocation, is as follows: Based on the target domain fault identification result set Calculate the dynamic entropy weights that reflect the uncertainty of the target domain prediction. Its formula is expressed as: ； in, Indicates the total number of samples in the target domain. This represents the total number of real tag categories in the source domain. Indicates Logarithmic function with base 0. This indicates that the target domain classifier predicts the first... Target domain samples Belongs to the label The probability of; Clustering algorithm based on Gaussian mixture model for deep feature sets of the target domain Perform cluster analysis to obtain The clustering results are stored for each deep feature cluster in the target domain; a cost matrix is defined to reflect the degree of matching between the deep feature clusters in the target domain and the true fault categories in the source domain. The number of rows in the cost matrix CM is equal to the number of deep feature clusters in the target domain. The number of columns represents the total number of real tag categories in the source domain. The first in the matrix Line number Column elements Indicates the first All deep features of the target domain in each feature cluster do not belong to the label. The arithmetic average probability; the optimal assignment mapping that minimizes the sum of costs is solved using the Hungarian algorithm, and different target domain pseudo-labels are assigned to each deep feature cluster of the target domain; Define global distribution difference loss , for , Wasserstein distance between them; Define prediction loss , For based on and source domain real tags The multi-class cross-entropy loss function; Define clustering loss , This is the sum of Wasserstein distances between all deep feature clusters in the target domain and their cluster centers; Define the feature region distribution loss , The sum of Wasserstein distances between all deep feature clusters in the target domain and deep feature clusters in the source domain with the same label; the same label means that the pseudo-label of the deep feature cluster in the target domain is equal to the true label of the deep feature cluster in the source domain; the deep feature clusters in the source domain are divided based on the true labels of the source domain. Each source domain deep feature cluster; based on , , , and dynamic entropy weight Calculate the composite loss function Its formula is expressed as: The training objective is set to minimize the composite loss function. .
7. A bearing fault diagnosis system based on similarity clustering region migration network, characterized in that, Specifically, it includes: The multi-source vibration data acquisition module is used to collect vibration signals of bearings under known operating conditions to construct a source domain dataset, and at the same time collect vibration signals of bearings to be tested to construct a target domain dataset. The source domain sample screening and target domain distribution correction module is used to score the quality of source domain samples and screen out excellent samples. At the same time, based on the distribution pattern of source domain samples, it performs similarity assessment and statistical projection correction on target domain samples to align the data distribution differences across working conditions. The time-frequency domain feature extraction and fusion module is used to extract and fuse time-frequency domain features of source domain and target domain samples through a time-frequency domain feature extraction network, generating source domain sample fusion features and target domain sample fusion features; The dynamic clustering transfer training and pseudo-label optimization module is used to construct a dynamic clustering transfer module, take fused features as input, and assign pseudo-labels through dynamic entropy weight adjustment, Gaussian mixture model clustering and Hungarian algorithm, calculate the composite loss function, and iteratively update the network parameters until training converges. The online fault diagnosis reasoning and result output module is used to load the trained model, perform feature extraction and transfer reasoning on the real-time vibration signal of the bearing to be tested, and output the fault type diagnosis result.
8. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; A processor is configured to, when running the computer program, execute a bearing fault diagnosis method based on a similarity clustering region migration network as described in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute a bearing fault diagnosis method based on a similarity clustering region migration network as described in any one of claims 1-6.