A method and system for small-sample bearing fault mode recognition based on multi-source data integration

Abstract

A method and system for identifying small-sample bearing fault modes based on multi-source data integration, relating to the field of intelligent operation and maintenance and health management of mechanical equipment. It solves the problems of difficulty in identification and inaccurate identification due to the limited number of samples in existing bearing fault modes. The method includes: constructing multi-source data samples based on a public dataset; extracting features from the multi-source data samples to construct source domain training sets and target domain training sets; combining the target domain training sets and source domain training sets to obtain a combined source domain training set; training a base classifier based on the target domain test set and the combined multi-source domain sample set; calculating inter-domain distribution metrics and sample similarity; constructing a weight matrix; classifying the target domain test set using the base classifier to obtain class probabilities; and weighting and integrating the class probabilities according to the weight matrix to obtain the classification results of the target domain test set, thus completing fault mode identification. This invention is applied in the field of fault identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and health management of mechanical equipment, and in particular to a small-sample bearing fault mode recognition method based on multi-source data integration. Background Technology

[0002] In the fault diagnosis and pattern recognition of rotating components in mechanical equipment, historical fault data for the monitored components is usually limited due to the complexity of the component structure and the constraints of the working environment. Bearing fault pattern recognition typically requires sufficient training data to build an accurate model. However, the small sample size problem means that only a limited number of fault samples are available for training, and the number of samples for each fault pattern may be unbalanced. This poses a challenge to training the model, as it may not be able to adequately learn and represent different types of faults. Bearing fault monitoring usually involves vibration signal data collected by sensors, which may be affected by environmental noise and interference. In the case of a small sample size, due to the limited number of data points, it may be difficult to accurately distinguish fault signals from noise signals, thus affecting the fault pattern recognition performance. Furthermore, effective feature selection and extraction become more difficult in the case of a small sample size. Conventional feature selection and extraction methods may fail to capture the effective information hidden in the small sample size. Due to the limited number of samples, the extracted features may not be able to fully reflect the feature differences of the fault patterns, thus affecting the model's performance. In the case of a small sample size, the number of samples for different fault types may be unbalanced, which may lead to poor model recognition ability for a few categories of fault patterns. Due to the limited number of samples, there may be insufficient samples to adequately learn and distinguish rare fault types. Small sample sizes can limit a model's ability to generalize to other bearing systems or real-world applications. This is because bearings in real-world applications may have various models and operating conditions, and the model may not maintain good performance when migrated from a small sample dataset to another. Summary of the Invention

[0003] This invention addresses the problems of limited sample size leading to difficulty and inaccuracy in existing bearing fault mode identification methods. It proposes a method for identifying limited sample bearing fault modes based on multi-source data integration. The specific solution is as follows:

[0004] A method for identifying small-sample bearing fault modes based on multi-source data integration, the method comprising:

[0005] S1: Construct multi-source data samples based on publicly available datasets;

[0006] S2: Extract features from multi-source data samples, and construct n source domain training sets and 1 target domain training set based on the features;

[0007] S3: Combine the one target domain training set and n or more source domain training sets to obtain the combined source domain training set;

[0008] S4: Train a base classifier based on the target domain training set and the combined multi-source domain sample set;

[0009] S5: Calculate the inter-domain distribution metric and sample similarity based on the target domain test set and the combined multi-source domain sample set;

[0010] S6: Construct a weight matrix based on the inter-domain distribution metric and sample similarity;

[0011] S7: Classify and identify the target domain test set to be identified according to the base classifier to obtain the class probability of the test set samples;

[0012] S8: Based on the weight matrix, perform weighted integration of the class probabilities of the test set samples to obtain the classification results of the target domain test set, and complete the fault mode identification.

[0013] Furthermore, a preferred embodiment is provided, wherein the extraction of multi-source data sample features in step S2 includes:

[0014] Extract time-domain features, frequency-domain features, and time-frequency-domain features from multi-source data samples;

[0015] The frequency domain features include the average frequency, root mean square frequency, frequency center, and root variance frequency.

[0016] Furthermore, a preferred embodiment is provided in which the time-frequency domain features are obtained by an empirical mode decomposition algorithm, and the original vibration signal is decomposed into a set of IMF components using the empirical mode decomposition algorithm.

[0017] Furthermore, a preferred embodiment is also provided, wherein step S3 includes:

[0018]

[0019] in, This represents the initial source domain training set. D represents the labeled target domain training set. Si This represents the recombined source domain training set, where n is the number of source domain training sets.

[0020] Furthermore, a preferred embodiment is provided in which the extreme learning machine is used to train the base classifier in step S4.

[0021] Furthermore, a preferred embodiment is also provided, wherein the inter-domain distribution metric and sample similarity in step S5 are specifically as follows:

[0022] The inter-domain distribution metric is:

[0023]

[0024] Among them, w j MMD is a measure of inter-domain distribution, where X and Y represent different sample sets, and MMD represents the maximum mean difference.

[0025] The sample similarity is:

[0026]

[0027] Among them, v j Sample similarity n s x is the number of samples in the source domain. t ∈D T Given a target domain sample to be classified, Cosine() represents the cosine similarity.

[0028] Furthermore, a preferred embodiment is also provided, wherein step S7 specifically comprises:

[0029]

[0030] in, For the classification results, k represents the category label. This represents the classification results of different base classifiers.

[0031] Based on the same inventive concept, this invention also provides a small-sample bearing fault mode recognition system based on multi-source data integration, the system comprising:

[0032] Sample building unit: used to build multi-source data samples based on public datasets;

[0033] Feature extraction unit: used to extract features from multi-source data samples, and construct n source domain training sets and 1 target domain training set based on the features;

[0034] Sample recombination unit: used to combine the 1 target domain training set and n source domain training sets to obtain the combined source domain training set;

[0035] Training unit: used to train a base classifier based on the target domain test set and the combined multi-source domain sample set;

[0036] Calculation unit: used to calculate inter-domain distribution metrics and sample similarity based on the target domain test set and the combined multi-source domain sample set;

[0037] Weight matrix construction unit: used to construct a weight matrix based on the inter-domain distribution metric and sample similarity;

[0038] Multi-source integration unit: used to classify and identify the target domain test set to be identified according to the base classifier, and obtain the class probability of the test set samples;

[0039] Identification unit: used to perform weighted integration of the class probabilities of the test set samples according to the weight matrix, obtain the classification results of the target domain test set, and complete the fault mode identification.

[0040] Based on the same inventive concept, the present invention also provides a computer-readable storage medium for storing a computer program that executes the small-sample bearing fault mode recognition method based on multi-source data integration described in any of the above claims.

[0041] Based on the same inventive concept, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the small sample bearing fault mode recognition method based on multi-source data integration as described above.

[0042] The advantages of this invention are:

[0043] This invention solves the problems of difficulty in identification and inaccurate identification caused by the small number of samples in existing bearing fault modes.

[0044] The small-sample bearing fault mode recognition method based on multi-source data integration proposed in this invention has low dependence on historical fault data of the parts to be tested. When there is very little available historical fault data, it combines relevant data of parts in the same field or of the same type with simulation data to achieve bearing fault mode recognition, thus having high practical application value.

[0045] The proposed method for small-sample bearing fault mode recognition based on multi-source data integration fully calculates the combined weights of base classifiers trained on multi-source domain datasets. It considers not only the inter-domain distribution measure between datasets but also the influence of the similarity of individual samples in different datasets. By introducing a sample similarity matrix, the recognition results of the multi-classification model are effectively integrated, thereby improving the accuracy of target domain sample classification.

[0046] This invention proposes a small-sample bearing fault mode recognition method based on multi-source data integration. By constructing multi-source data samples using publicly available datasets, the historical fault data of the bearings to be monitored can be expanded. Introducing multi-source data, including relevant data from other fields or types, as well as simulation data, provides more samples for model training, increasing the number of samples in the target domain training set, thereby helping to improve the model's accuracy and generalization ability. This implementation uses multi-source data samples to extract features and constructs n source domain training sets and one target domain training set. By combining multiple source domain training sets with the target domain training set to obtain a combined source domain training set, the heterogeneity of data sources can be better considered. This helps improve the model's adaptability to target domain data and its recognition accuracy.

[0047] This invention proposes a small-sample bearing fault mode recognition method based on multi-source data integration. It extracts features from multi-source data samples and constructs source domain and target domain training sets. The feature selection and extraction process utilizes appropriate feature engineering methods to choose features with discriminative and representative characteristics. This reduces the interference of noise and redundant information on model training, thereby improving the model's accuracy and robustness.

[0048] This invention provides a small-sample bearing fault mode recognition method based on multi-source data integration. It combines the target domain test set with the combined multi-source domain sample set, using the same labeling and tagging system for model training. This solves the problem of inconsistent labels between different data sources and data types, ensuring consistency in model training and evaluation.

[0049] This invention is applied to the field of fault identification. Attached Figure Description

[0050] Figure 1 This is a flowchart of the small-sample bearing fault mode recognition method based on multi-source data integration described in Implementation Method 1;

[0051] Figure 2 The average classification accuracy results of various algorithms under different target domain training set sizes as described in Implementation Method Eleven;

[0052] Figure 3 The effect of the number of neurons described in Implementation Method Eleven on the average classification accuracy of the proposed algorithm. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0054] Implementation Method 1: A small-sample bearing fault mode recognition method based on multi-source data integration, as described in this implementation method, includes:

[0055] S1: Construct multi-source data samples based on publicly available datasets;

[0056] S2: Extract features from multi-source data samples, and construct n source domain training sets and 1 target domain training set based on the features;

[0057] S3: Combine the one target domain training set and n or more source domain training sets to obtain the combined source domain training set;

[0058] S3: Combine the target domain training set and n source domain training sets to obtain the combined source domain training set;

[0059] S4: Train a base classifier based on the target domain training set and the combined multi-source domain sample set;

[0060] S5: Calculate the inter-domain distribution metric and sample similarity based on the target domain test set and the combined multi-source domain sample set;

[0061] S6: Construct a weight matrix based on the inter-domain distribution metric and sample similarity;

[0062] S7: Classify and identify the target domain test set to be identified according to the base classifier to obtain the class probability of the test set samples;

[0063] S8: Based on the weight matrix, perform weighted integration of the class probabilities of the test set samples to obtain the classification results of the target domain test set, and complete the fault mode identification.

[0064] The method described in this embodiment expands the historical fault data of the bearing to be monitored by constructing multi-source data samples using publicly available datasets. Introducing multi-source data, including relevant data from other domains or types, as well as simulation data, provides more samples for model training, increasing the number of samples in the target domain training set, thereby improving the model's accuracy and generalization ability. This embodiment uses multi-source data samples to extract features and constructs n source domain training sets and one target domain training set. By combining multiple source domain training sets with the target domain training set to obtain a combined source domain training set, the heterogeneity of data sources can be better considered. This helps improve the model's adaptability to target domain data and its recognition accuracy.

[0065] This implementation extracts features from multi-source data samples and constructs source domain training sets and target domain training sets. The feature selection and extraction process utilizes appropriate feature engineering methods to choose features with discriminative and representative characteristics. This reduces the interference of noise and redundant information on model training, thereby improving the model's accuracy and robustness.

[0066] This implementation combines the target domain test set and the combined multi-source domain sample set, using the same labeling and tagging system for model training. This solves the problem of inconsistent labels between different data sources and data types, ensuring consistency in model training and evaluation.

[0067] This implementation method utilizes publicly available datasets to acquire more data samples and expand historical fault data. This helps increase the number of training samples for the model, improving its accuracy and generalization ability. By extracting features from multi-source data samples, key information characterizing bearing failure modes can be extracted. Constructing source domain training sets and target domain training sets from the features of multi-source data samples separately helps to consider the characteristics of different data sources and the differences in failure modes. By combining the target domain training set with multiple source domain training sets, the information from multi-source data can be better utilized, improving the model's learning ability and smoothing the data distribution, thus increasing the model's adaptability to target domain data. By using the combined multi-source domain sample set and target domain test set, training the base classifier can reduce its dependence on target domain samples, improving the model's generalization ability on unseen target domain samples. By calculating the inter-domain distribution metric and sample similarity between the target domain test set and the multi-source domain sample set, the degree of association between different samples can be obtained. Based on these metrics, a weight matrix can be constructed to adjust the influence of samples in the weighted ensemble according to their importance, improving the model's stability and accuracy. By classifying the target domain test set using a base classifier and obtaining the probability of each category, we can provide a confidence estimate of the sample classification and gain a more comprehensive understanding of the model's discriminative ability in the target domain data. By weighted ensemble of the category probabilities of the test set samples based on a weight matrix, we can comprehensively consider the importance of different source domains and samples to obtain the final classification result of the target domain test set. This can improve the robustness and recognition accuracy of the model, thereby achieving effective identification of bearing failure modes.

[0068] Implementation Method Two: This implementation method further defines the small-sample bearing fault mode recognition method based on multi-source data integration described in Implementation Method One. Step S2, extracting multi-source data sample features, includes:

[0069] Extract time-domain features, frequency-domain features, and time-frequency-domain features from multi-source data samples;

[0070] The frequency domain features include the average frequency, root mean square frequency, frequency center, and root variance frequency.

[0071] In practical applications, the time-domain features are shown in Table 1:

[0072] Table 1 Temporal Characteristics

[0073]

[0074] Where, x i Let i = 1, 2, ..., n represent the time series, where n is the number of data points, and x max =max|x i |and This represents the absolute mean.

[0075] The frequency domain features are shown in Table 1:

[0076] Table 2 Frequency Domain Characteristics

[0077]

[0078] Where, p i x represents i The power spectrum of i = 1, 2, L, N, where N represents the number of spectral lines. i F represents the amplitude of the i-th spectral line. mf It can represent the vibrational energy in the frequency domain, F rmsf and F fc Describes the location of the main frequency. F rvf It can characterize the degree of concentration or dispersion of spectral power energy.

[0079] Implementation Method 3: This implementation method further defines the small-sample bearing fault mode recognition method based on multi-source data integration described in Implementation Method 2. The time-frequency domain features are obtained by the empirical mode decomposition algorithm, which decomposes the original vibration signal into a set of IMF components.

[0080] Specifically, taking the energy of the first six IMF components as time-frequency domain features, the amplitude energy of the IMF is calculated as follows:

[0081]

[0082] Among them, E j Let N be the amplitude energy of the IMF, N represent the length of the j-th IMF component data, and H[g] represent the Hilbert transform.

[0083] This implementation decomposes the original vibration signal into IMF components, allowing for the extraction of vibration characteristics across different frequency ranges. Each IMF component can be viewed as a vibration mode with different frequencies and amplitudes, facilitating a more comprehensive description of the vibration signal's time-frequency domain characteristics. EMD can separate noise and interference in the signal into the high-frequency IMF components, while retaining fault characteristic information in the low-frequency IMF components. This helps improve the accuracy and reliability of fault mode identification. The decomposed IMF components exhibit good interpretability, with each component associated with a specific vibration mode or fault mode. This makes the characteristic analysis of the vibration signal more intuitive and understandable, aiding in fault diagnosis and fault mode interpretation.

[0084] This implementation utilizes the empirical mode decomposition algorithm to decompose the original vibration signal into a set of IMF components, which can extract rich time-frequency domain feature information, enhance the accuracy of fault mode identification, and provide independent analysis capabilities for vibration characteristics in different frequency bands.

[0085] This embodiment, described in conjunction with Embodiment 2, addresses the challenge that when a fault is relatively weak, relying solely on the statistical characteristics of time-domain and frequency-domain signals is insufficient to fully characterize the health status of components. Therefore, in this embodiment, EMD (Electronic Dynamics Analysis) is employed to extract more characteristic information about the bearing. EMD can reveal vibration signals from the perspective of time-frequency amplitude distribution. The EMD algorithm decomposes the original vibration signal into a set of IMFs (Integrated Motion Frames):

[0086]

[0087] Among them, c i (t) represents the j-th IMF component of x(t), which represents different frequency bands of signal from high frequency to low frequency. n (t) represents a residual signal with a monotonic trend. The amplitude energy of the IMF is calculated as follows:

[0088]

[0089] Typically, the first six IMFs can contain almost all the information of the vibration signal. For vibration signals under different fault modes, their intrinsic IMF components contain different information. Therefore, in this implementation, the amplitude energy of the first six IMFs is extracted as the time-frequency domain feature of the vibration signal. Implementations two and three together construct a feature set consisting of twelve time-domain features, four frequency-domain features, and six time-frequency-domain features.

[0090] Implementation Method Four: This implementation method further defines the small-sample bearing fault mode recognition method based on multi-source data integration described in Implementation Method One. Step S3 includes:

[0091]

[0092] in, This represents the initial source domain training set. D represents the labeled target domain training set. Si This represents the recombined source domain training set, where n is the number of source domain training sets.

[0093] In fault mode recognition, a base classifier needs to be trained first. However, in real-world scenarios, due to the difficulty of data acquisition and the high cost of labeling target domain data, the amount of labeled sample data in the target domain is usually limited. Training the model using only a small amount of labeled data from the same working conditions results in low accuracy and unsatisfactory classification results. Utilizing datasets from the same domain can effectively increase the number of training samples for the model, and the model can effectively leverage knowledge from multiple source domain samples. Combining labeled target domain data with data from multiple source domains can improve the classification accuracy of the trained model.

[0094] Implementation Method 5: This implementation method further defines the small-sample bearing fault mode recognition method based on multi-source data integration described in Implementation Method 1. In step S4, an extreme learning machine is used to train the base classifier.

[0095] In fault pattern recognition, real-time performance is a crucial consideration. The training process of ELM (Elastic Mode Model) is highly efficient, requiring no iterative optimization. Instead, it directly solves for the output weights by randomly initializing weights and biases. This training method makes ELM extremely fast, suitable for large-scale fault pattern recognition problems and real-time applications. Fault patterns often involve large amounts of sensor data and complex signal information, leading to the need for processing high-dimensional and complex data. ELM effectively handles high-dimensional and complex data and improves fault pattern recognition performance by mapping data to a high-dimensional feature space and utilizing neurons in random hidden layers to perform nonlinear processing. Fault patterns are often affected by data noise, incomplete information, and imbalanced samples. ELM exhibits strong robustness during training and is relatively insensitive to noise and incomplete information. Furthermore, ELM achieves a nonlinear classification decision boundary in a high-dimensional feature space, possessing good generalization ability and effectively extending to unseen fault patterns, improving the accuracy and robustness of fault pattern recognition. Based on a single-iteration weight adjustment process, ELM has good online learning capabilities. In practical fault mode identification (FMD) scenarios, there is often a need to address changes in data flow and the requirement for real-time updates. ELM can quickly adapt to new data samples, enabling dynamic model updates and online learning.

[0096] This implementation utilizes Extreme Learning Machine (ELM) to train the base classifier, which offers advantages in fault mode recognition such as high training speed, scalability, ability to handle high-dimensional and complex data, robustness, generalization ability, and the ability to meet online learning requirements. These advantages make ELM an effective method for building fault mode recognition systems and achieving accurate, fast, and robust fault diagnosis and monitoring capabilities.

[0097] Implementation Method Six: This implementation method further defines the small-sample bearing fault mode recognition method based on multi-source data integration described in Implementation Method One. Specifically, the inter-domain distribution metric and sample similarity in step S5 are as follows:

[0098] The inter-domain distribution metric is:

[0099]

[0100] Among them, w j MMD is a measure of inter-domain distribution, where X and Y represent different sample sets, and MMD represents the maximum mean difference.

[0101] The sample similarity is:

[0102]

[0103] Among them, v j Sample similarity n s x is the number of samples in the source domain. t ∈D T Given a target domain sample to be classified, Cosine() represents the cosine similarity.

[0104] Traditional inter-domain distribution measurement methods typically only consider the differences in the overall sample distribution, neglecting the differences between individual samples. However, sample similarity weights based on cosine similarity fully consider the similarity between each sample and other samples. This implementation incorporates the differences between individual samples, helping to more accurately assess the similarity between source and target domain data. The cosine similarity-based sample similarity weight matrix can provide personalized weight assignments for each target domain sample to be classified. By calculating the cosine similarity between target and source domain samples, it is possible to determine which source domain samples have greater similarity to target domain samples, thus assigning different weights to different samples. Such personalized weight assignments can better distinguish and utilize samples in the source domain data that are similar to the target domain data, improving the accuracy of fault mode recognition. In the case of small samples, the quantity and diversity of data are limited, and traditional classification methods may face problems such as overfitting and underfitting. Cosine similarity-based sample similarity weights, by introducing the similarity relationship between target and source domain samples, enable target domain samples to better utilize the information from source domain samples and capture the differences between samples at the individual sample level, effectively compensating for the shortcomings of small sample data.

[0105] Implementation Method Seven: This implementation method further defines the small-sample bearing fault mode recognition method based on multi-source data integration described in Implementation Method One. Step S7 specifically includes:

[0106]

[0107] in, For the classification results, k represents the category label. This represents the classification results of different base classifiers.

[0108] Implementation Method 8: A small-sample bearing fault mode recognition system based on multi-source data integration, as described in this implementation method, includes:

[0109] Sample building unit: used to build multi-source data samples based on public datasets;

[0110] Feature extraction unit: used to extract features from multi-source data samples, and construct n source domain training sets and 1 target domain training set based on the features;

[0111] Sample recombination unit: used to combine the one target domain training set and n or more source domain training sets to obtain the combined source domain training set;

[0112] Training unit: used to train a base classifier based on the target domain training set and the combined multi-source domain sample set;

[0113] Calculation unit: used to calculate inter-domain distribution metrics and sample similarity based on the target domain test set and the combined multi-source domain sample set;

[0114] Weight matrix construction unit: used to construct a weight matrix based on the inter-domain distribution metric and sample similarity;

[0115] Multi-source integration unit: used to classify and identify the target domain test set to be identified according to the base classifier, and obtain the class probability of the test set samples;

[0116] Identification unit: used to perform weighted integration of the class probabilities of the test set samples according to the weight matrix, obtain the classification results of the target domain test set, and complete the fault mode identification.

[0117] Implementation Method Nine: A computer-readable storage medium according to this implementation method, the computer-readable storage medium being used to store a computer program, the computer program executing any one of Implementation Methods One to Seven, a small sample bearing fault mode recognition method based on multi-source data integration.

[0118] Implementation Method 10: A computer device according to this implementation method includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a small sample bearing fault mode recognition method based on multi-source data integration according to any one of Implementation Methods 1 to 7.

[0119] Implementation Method 11, see below Figure 2 and Figure 3 This embodiment describes a specific implementation of the small-sample bearing fault mode recognition method based on multi-source data integration described in Embodiment 1. It also serves to explain Embodiments 2 through 7. Specifically:

[0120] In this embodiment, four publicly available datasets from different institutions are selected for cross-domain validation to comprehensively compare the performance of the method described in this invention. The selected datasets include the JNU bearing dataset, the IMS bearing dataset, the NUAA bearing dataset, and the CWRU bearing dataset. Each dataset contains 200 samples of each of four different fault types: normal condition, outer race fault, inner race fault, and rolling element fault, for a total of 800 samples. One dataset is selected as the target domain dataset, and the other three datasets are selected as the source domain datasets.

[0121] Figure 2The comparison results of the small sample bearing fault mode recognition method based on multi-source data integration provided in the embodiments of this application with various algorithms are shown. Among them, the ELM-MSDI model is the model established by the method provided in this application. It can be seen that when the sample size of the target domain training set is small, the method provided in this application has the best classification accuracy.

[0122] In this example, the number of neurons is a core parameter affecting the classification accuracy of the method provided in this application. To investigate the impact of the number of neurons on the proposed model, this embodiment trains the model using 20, 30, and 40 sets of target domain data, respectively, and obtains the global average results obtained by training the model using the target domain training set and using the proposed multi-source ensemble strategy under different numbers of neurons, as shown below. Figure 3 As shown.

[0123] from Figure 3 As can be seen, the influence of the number of neurons on the accuracy of the method provided in this application trained with 20, 30, and 40 sets of target domain training sets follows a consistent trend. When 100 neurons are selected, the prediction accuracy of the model is relatively low. When the number of neurons is between 200 and 1000, the influence of the number of neurons on the model accuracy is not significant, indicating that the method and the constructed model provided by this invention have good stability.

[0124] This implementation addresses the challenge of finding readily available labeled datasets for different working conditions across various domains in real-world applications, while datasets related to different working conditions within the same domain are readily available. It proposes an effective multi-source ensemble model. When calculating the combined weights of classifiers trained on multi-source domain datasets, it considers not only inter-domain distribution metrics but also the impact of individual sample similarity across different datasets. By introducing a sample similarity matrix, the recognition results of the multi-classification model are effectively integrated, thereby improving the accuracy of target domain sample classification. Finally, the method is validated in different classification tasks and compared with multiple different strategies and classification algorithms, demonstrating that the proposed method achieves higher classification accuracy in fault mode recognition on small target domain training sets.

[0125] It should be noted that the methods and detailed examples provided in the above embodiments can be incorporated into the devices and equipment provided in the embodiments, and are referred to each other without further elaboration. Those skilled in the art will recognize that the units and algorithm steps of the various instances described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. In the embodiments provided in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / equipment embodiments described above are merely illustrative. For example, the division of the above modules or units is merely a logical functional division, and in actual implementation, it can be divided in other ways. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A small-sample bearing fault mode recognition method based on multi-source data integration, characterized in that, The method includes: S1: Construct multi-source data samples based on publicly available datasets; S2: Extract features from multi-source data samples, and construct n source domain training sets and 1 target domain training set based on the features; S3: Combine the target domain training set and n source domain training sets to obtain the combined source domain training set; S4: Train a base classifier based on the target domain training set and the combined multi-source domain sample set; S5: Calculate the inter-domain distribution metric and sample similarity based on the target domain test set and the combined multi-source domain sample set; S6: Construct a weight matrix based on the inter-domain distribution metric and sample similarity; S7: Classify and identify the target domain test set to be identified according to the base classifier to obtain the class probability of the test set samples; S8: The class probabilities of the test set samples are weighted and integrated according to the weight matrix to obtain the classification results of the target domain test set and complete the fault mode identification. Step S3 includes: ， in, This represents the initial source domain training set. This represents a labeled target domain training set. This represents the recombined source domain training set. n The number of training sets in the source domain; The inter-domain distribution metric and sample similarity in step S5 are specifically as follows: The inter-domain distribution metric is: ， in, MMD is a measure of inter-domain distribution, where X and Y represent different sample sets, and MMD represents the maximum mean difference. The sample similarity is: ， in, For sample similarity The number of samples in the source domain. For a target domain sample to be classified, Cosine similarity; Step S8 specifically involves: in, For the classification results, k Indicates category label, This represents the classification results of different base classifiers.

2. The method for identifying small-sample bearing fault modes based on multi-source data integration according to claim 1, characterized in that, The extraction of multi-source data sample features in step S2 includes: Extract time-domain features, frequency-domain features, and time-frequency-domain features from multi-source data samples; The frequency domain features include the average frequency, root mean square frequency, frequency center, and root variance frequency.

3. The method for identifying small-sample bearing fault modes based on multi-source data integration according to claim 2, characterized in that, The time-frequency domain features are obtained by the empirical mode decomposition algorithm, which decomposes the original vibration signal into a set of IMF components.

4. The method for identifying small-sample bearing fault modes based on multi-source data integration according to claim 1, characterized in that, In step S4, the extreme learning machine is used to train the base classifier.

5. A small-sample bearing fault mode recognition system based on multi-source data integration, characterized in that, The system includes: Sample building unit: used to build multi-source data samples based on public datasets; Feature extraction unit: used to extract features from multi-source data samples, and construct n source domain training sets and 1 target domain training set based on the features; Sample recombination unit: used to combine the target domain training set and n source domain training sets to obtain a combined source domain training set; including: ， in, This represents the initial source domain training set. This represents a labeled target domain training set. This represents the recombined source domain training set. n The number of training sets in the source domain; Training unit: used to train a base classifier based on the target domain training set and the combined multi-source domain sample set; Calculation unit: used to calculate inter-domain distribution metrics and sample similarity based on the target domain test set and the combined multi-source domain sample set; Weight matrix construction unit: used to construct a weight matrix based on the inter-domain distribution metric and sample similarity; the inter-domain distribution metric and sample similarity are specifically as follows: The inter-domain distribution metric is: ， in, MMD is a measure of inter-domain distribution, where X and Y represent different sample sets, and MMD represents the maximum mean difference. The sample similarity is: ， in, For sample similarity The number of samples in the source domain. For a target domain sample to be classified, Cosine similarity; Multi-source integration unit: used to classify and identify the target domain test set to be identified according to the base classifier, and obtain the class probability of the test set samples; The identification unit is used to perform weighted ensemble analysis on the class probabilities of the test set samples according to the weight matrix, obtain the classification results of the target domain test set, and complete the fault mode identification. Specifically: in, For the classification results, k Indicates category label, This represents the classification results of different base classifiers.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes the small-sample bearing fault mode recognition method based on multi-source data integration as described in any one of claims 1-4.

7. A computer device, characterized in that, The system includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a small-sample bearing fault mode recognition method based on multi-source data integration according to any one of claims 1-4.

Images