A small sample fault diagnosis method based on twin consistency fault representation self-learning

Abstract

The present invention provides a small sample fault diagnosis method based on self-learning of twin consistency fault representation, which belongs to the field of fault diagnosis technology under small sample data conditions. First, the sample data set is preprocessed to obtain training set samples and test set samples; sample association matching is performed in the training set samples to construct a fault sample association data set, and the sample space is simplified. Secondly, based on the fault sample association data set, the training and optimization of the fault diagnosis model are completed by setting the training hyperparameters, and the fault features of the samples to be tested and the training set samples are extracted. Finally, the fault feature twin similarity measurement is performed to identify the fault mode of the sample to be tested. The small sample fault diagnosis method provided by the present invention can improve the generalization ability and diagnostic performance of the fault diagnosis model under the condition of unbalanced fault information; realize the adaptive measurement of twin feature similarity, and improve the accuracy of fault mode identification.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis under small sample data conditions, and relates to a small sample fault diagnosis method based on twin consistency fault characterization self-learning. Background Art

[0002] As the automation level and complexity of industrial equipment gradually increase, the demand for safety, reliability, and stability during operation is also gradually increasing. Timely and accurate fault diagnosis technology can provide effective technical support for autonomous perception of equipment status and autonomous decision-making in task planning. Deep learning-based fault diagnosis methods have become the mainstream fault diagnosis method due to their flexible models and strong feature extraction capabilities. The performance of deep learning-based fault diagnosis methods is highly dependent on the quality of equipment monitoring data. However, during operation, equipment is mostly in normal state, so the number of normal samples collected is far greater than the number of fault samples [Chinese Invention Patent CN113281029A]. When there is a large difference in the number of samples of different equipment failure modes, the deep learning-based fault diagnosis model will perform biased fault information mining, making it impossible to accurately extract highly generalized equipment status representations, ultimately resulting in reduced diagnostic performance.

[0003] Currently, the main approaches to addressing the poor diagnostic performance caused by an imbalance in the number of samples for different fault modes include expert knowledge-based fault diagnosis methods, generative model-based fault diagnosis methods, and resampling-based fault diagnosis methods. Expert knowledge-based fault diagnosis methods primarily leverage the experience and rules of domain experts to establish fault inference rules to guide the model learning process and thus perform fault identification. However, this approach relies heavily on expert knowledge, making fault knowledge often difficult to construct, and the model's generalization and flexibility are poor. Generative model-based fault diagnosis methods, based on generative adversarial networks, simulate real-world data distributions to generate new samples similar to the original data, thereby alleviating the problem of insufficient samples for some faults. However, this approach places high demands on the original samples and is prone to generating redundant samples containing a large amount of repeated information. Resampling-based fault diagnosis methods balance the number of samples for different categories by oversampling or undersampling the original sample data, thereby improving classification accuracy. However, the main drawback of this approach is that excessive resampling can introduce noise, reducing model learning efficiency. Summary of the Invention

[0004] To address the problem of weak fault diagnosis capabilities in existing fault diagnosis models when the distribution of fault samples is unbalanced and effective fault information is scarce, the present invention provides a small-sample fault diagnosis method based on self-learning of twin-consistent fault representation. This method draws on the twin network architecture to construct a twin-consistent fault representation self-learning network. By learning the fault correlation between samples in the original sample data set, it extracts and mines robust fault features with strong representation capabilities, ultimately improving the generalization ability and diagnostic performance of the fault diagnosis model under conditions of unbalanced fault information.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A small sample fault diagnosis method based on twin consistency fault representation self-learning, the small sample fault diagnosis comprising the following steps:

[0007] Step 1: Collect equipment status monitoring data to construct a sample data set, and preprocess the sample data set to obtain training set samples and test set samples.

[0008] Step 2: Based on the training set samples obtained in step 1, the fault label information of different samples is used to perform sample association matching, thereby constructing fault-associated samples and obtaining a fault sample association data set.

[0009] Step 3: Based on the fault sample association data set obtained in step 2, perform sample space simplification. By setting a simplification scale, the fault association samples are simplified, thereby constructing a simplified fault sample association data set.

[0010] Step 4. Construct and train a fault diagnosis model based on the twin consistency fault representation self-learning network. By selecting the model architecture and parameters, construct a fault diagnosis model based on the twin consistency fault representation self-learning network. Based on the fault sample association data set obtained in step 3, complete the training and optimization of the fault diagnosis model by setting the training hyperparameters.

[0011] Step 5: Extract fault features of the test samples and training set samples. After the fault diagnosis model is trained, fault features are extracted for the test samples and training set samples based on the feature encoder in the trained fault diagnosis model.

[0012] Step 6: Measure the twin similarity of fault features. Based on the fault features of the test sample and the training set samples obtained in step 5, use the feature fusion measurement module in the trained fault diagnosis model to measure the twin similarity of the fault features between the features.

[0013] Step 7: Identify the fault mode of the sample to be tested. Based on the twin similarity measurement results of the fault characteristics obtained in step 6, the fault mode of the sample to be tested is identified by setting a label recognition strategy based on the unbalanced sample distribution.

[0014] The details are as follows:

[0015] Furthermore, the sample data set construction method described in step 1 includes using a sliding window to perform sample division on the condition monitoring data, and performing maximum and minimum value normalization processing, to provide a data basis for the subsequent fault diagnosis model based on the twin consistency fault characterization self-learning network to learn fault characteristics; and includes the following sub-steps:

[0016] Step 1.1, dividing the training set and test set samples;

[0017] The sample data set is constructed for the equipment's condition monitoring data. First, for the collected condition monitoring data of each fault mode, multiple samples are generated using sliding window cutting. Suppose the condition monitoring data under a certain fault mode is Length is , sample cutting is performed with a window of length w and a moving step of s, and we get Generate samples and then select training samples, and the remaining samples as test samples.

[0018] Step 1.2, data normalization;

[0019] Normalize each sample S in the training data to obtain the normalized training sample set and test sample set. The normalization formula is:

[0020]

[0021] in, is a signal sequence consisting of training samples; is the maximum value in the signal sequence composed of training samples, is the minimum value in the signal sequence composed of training samples, The normalized sample is obtained after the original sample is normalized.

[0022] Finally, we get the training data set and test dataset , a total of N failure modes, among which and Represents the training sample set and test sample set with failure mode N.

[0023] Furthermore, the sample association matching described in step 2 is specifically as follows: according to the failure mode of the sample, the samples are matched with each other, the samples with the same failure mode are matched as positively associated samples, and the samples with different failure modes are matched as negatively associated samples; the combination label of the samples with the same label is 1, and the combination label of the samples with different labels is 0. It includes the following sub-steps:

[0024] Step 2.1, correlation matching of samples with the same fault mode;

[0025] For the training sample set with fault mode i, two samples are randomly selected from it , since they belong to the same failure mode, we have ,in Represents the fault correlation operator; for samples with the same fault mode, its correlation attribute is 1.

[0026] Step 2.2, correlation matching of different fault mode samples;

[0027] For failure modes and The training sample set 、 , take one sample from each , because they belong to different failure modes, there are , for samples of different failure modes, their associated attributes are 0.

[0028] After matching the fault samples of all training samples, the fault sample association dataset is finally obtained. 、 , where the subscript 、 are the number of associated sample pairs whose associated attributes are 1 and 0, respectively.

[0029] Furthermore, the sample space reduction described in step 3 is specifically as follows: for the fault sample association dataset obtained in step 2, considering the distribution of the number of samples of different fault modes in the training dataset obtained in step 1, setting a reduction scale, and reducing the samples of different fault modes involved in sample association, thereby achieving simplification of the fault association sample space; and includes the following sub-steps:

[0030] Step 3.1, construct the reduction scale;

[0031] For the training sample set , N is the number of failure modes. The present invention considers the relative size of the distribution of the number of samples of different failure modes in the training data set, so there is , .in, The failure mode is The training sample set, The failure mode is The training sample set, is the minimum number of samples for each failure mode in the training data set, is the median number of samples of each failure mode in the training data set, and Each represents a set of training samples of a fault mode. For the associated sample set with an associated attribute of 1, the reduction scale ; For the associated sample set with an associated attribute of 0, the scale is simplified .

[0032] Step 3.2, sample space reduction;

[0033] For the fault sample association dataset constructed in step 2 、 ,in, For the associated sample set constructed for the same failure mode sample, since there are N failure modes in total, Randomly select and keep number of samples, and remove the remaining samples to complete The sample space of simplifies; The associated sample set constructed for different fault mode samples, among N kinds of fault modes, the fault associated samples obtained by sample association with normal samples are randomly selected and retained from these associated samples. The number of samples is eliminated and the rest of the samples are completed The sample space is simplified. Through the above sample reduction operation, redundant fault-related information is reduced while ensuring the diversity and maximization of fault information as much as possible. Finally, the simplified fault-related sample set is obtained.

[0034] Furthermore, the fault diagnosis model based on the twin-consistent fault representation self-learning network is constructed and trained as described in step 4. Specifically, the model uses the twin neural network as the fault feature encoding module and the deep convolutional neural network as the feature fusion measurement module. Based on the simplified fault-related sample set obtained in step 3, the root mean square error is used as the training loss function to complete the training of the fault diagnosis model. It includes the following sub-steps:

[0035] Step 4.1, design the fault diagnosis model architecture;

[0036] The present invention constructs a fault diagnosis model based on a twin-consistent fault representation self-learning network. The fault diagnosis model consists of a fault feature encoding module and a feature fusion metric module. The fault feature encoder is composed of a two-branch network with shared weights, and the feature fusion metric module is composed of a convolutional network.

[0037] First, for a pair of related samples and Use the feature encoding module to extract features and obtain fault features and ,in Represents the linear transformation operation of the input sample using the encoder module, Represents a nonlinear transformation operation. Then, the feature fusion measurement module is used to measure the twin similarity of the above fault features, and the measurement result is , where For the feature splicing operation, the two fault features are spliced along the sample dimension in a way that ensures the feature dimension remains unchanged. In order to utilize the feature measurement module based on convolutional neural network, the twin similarity of fault features is measured by convolution operation on the input. Finally, the twin similarity of fault features of this pair of associated samples is calculated. .

[0038] Step 4.2, training the fault diagnosis model;

[0039] Based on the fault diagnosis model constructed in step 4.1, the present invention uses the root mean square error is the loss function, where for the kth group of fault-related samples , is the twin similarity prediction value, is the twin similarity label value. This method uses the gradient descent strategy to update the parameters of the fault diagnosis model. The principle is to assume that the parameters of the fault diagnosis model are , the loss function used for training the fault diagnosis model is , then the optimization objective is :

[0040] The parameter update strategy is:

[0041]

[0042] Where, is the initial parameter, is the updated parameter; is the learning rate.

[0043] The learning rate value can control the update amplitude of the parameters in each iteration. When the learning rate value is large, the gradient of the parameters decreases greatly and the loss function converges faster. However, when the learning rate is too large, the loss function will oscillate near its extreme point, thus hindering convergence. When the learning rate value is small, the gradient of the parameters decreases less, which will lead to a slower convergence of the model. When it gradually converges to 0, it means that the fault diagnosis model has optimized a set of suitable parameters. , so that the loss function Get the minimum value.

[0044] Furthermore, the feature extraction described in step 5 is specifically as follows: after completing the training of the fault diagnosis model, based on the fault feature encoding module in the fault diagnosis model, the training data set constructed in step 1 is extracted. All training samples in the training set are extracted to obtain the fault features of the training set , the sample to be tested Extract fault features and obtain fault features of the sample to be tested. .

[0045] Furthermore, the fault feature twin similarity measurement described in step 6 is specifically:

[0046] Step 6.1, reconstructing the fault characteristics of the test sample;

[0047] The fault characteristics of the sample to be tested obtained in step 5 Fault characteristics of the training set All sample fault features in the fault feature set are spliced together. The splicing method is to splice along the sample dimension to ensure that the feature dimension remains unchanged. Finally, the reconstructed fault feature set is obtained through feature splicing. ,in Represents the failure mode N The training samples and failure modes are No. The splicing fault characteristics are composed of test samples.

[0048] Step 6.2, fault feature twin similarity measurement;

[0049] Using the feature fusion metric module in the fault diagnosis model, calculate the fault feature set obtained in step 6.1 The twin similarity of each splicing fault feature in the training set is obtained, and the twin similarity between the test sample and the training set sample is defined as ,in , represents the total number of failure modes N training samples and test samples The twin similarity set of .

[0050] Furthermore, the fault mode identification of the test sample described in step 7 is specifically as follows:

[0051] Based on the twin similarity results between the test sample and the training set sample obtained in step 6 , the label recognition result of the sample to be tested is , that is, calculating the mean twin similarity between the training set samples and the test samples for different fault modes, and then taking the fault mode with the largest mean similarity as the fault mode identification result for the test sample. This completes the fault mode identification of the test sample; finally, steps 5 to 7 are performed on all test set samples to achieve fault diagnosis for all test samples.

[0052] The beneficial effects of the present invention are:

[0053] (1) To address the problems of weak generalization ability and poor feature robustness of existing fault diagnosis models caused by the imbalance in the number of samples of different fault modes, a fault diagnosis method based on self-learning of twin consistency representation is proposed.

[0054] (2) The sample association matching method in the fault diagnosis method of the present invention can utilize the fault information of different samples, establish the fault association relationship between different samples by performing association matching on different samples, and realize sample association matching.

[0055] (3) The sample reduction method in the fault diagnosis method of the present invention can simplify and redistribute the unbalanced fault information distribution for the constructed fault-related sample set, thereby helping the model learn more robust fault characteristics.

[0056] (4) The twin consistency fault characterization self-learning network in the fault diagnosis method of the present invention realizes the adaptive measurement of twin feature similarity by fusing the feature encoding module based on the twin network and the feature fusion measurement module based on the convolutional network, thereby avoiding the subjective dependence of the traditional contrast loss function on the selection of hyperparameters during the training process.

[0057] (5) The method for identifying the fault mode of the test sample proposed in the present invention can use the fault information of the existing training samples to comprehensively weigh and measure the fault category to which the test sample belongs, thereby improving the accuracy of fault mode identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of the small sample fault diagnosis method based on self-learning of twin consistency fault representation;

[0059] Figure 2 It is the correlation diagram of samples with the same failure mode;

[0060] Figure 3 It is the correlation diagram of different failure mode samples;

[0061] Figure 4 It is the correlation diagram of samples with the same fault mode after simplification;

[0062] Figure 5 is the correlation diagram of different failure mode samples after simplification;

[0063] Figure 6 This is a structural diagram of the fault diagnosis model based on the twin consistency fault representation self-learning network;

[0064] Figure 7 Loss function curve for the fault diagnosis model training process;

[0065] Figure 8 The twin similarity results of the associated samples in the training set;

[0066] Figure 9 is the fault feature result of the training set samples;

[0067] Figure 10 is the fault feature result of the test set sample;

[0068] Figure 11 It is the similarity measurement result of the fault characteristic twin of the tested samples with different fault modes and the original samples;

[0069] Figure 12 Confusion matrix results for fault diagnosis of test set samples;

[0070] Figure 13 Comparison of the accuracy of the proposed method and the fault diagnosis model based on the traditional twin network;

[0071] Figure 14 This is a comparison of the diagnostic performance of the method proposed in this invention under different data conditions. DETAILED DESCRIPTION

[0072] The present invention is further described below with reference to specific implementation cases.

[0073] Step 1: Collect equipment status monitoring data to construct a sample data set, and preprocess the sample data set to obtain training set samples and test set samples;

[0074] The present invention uses the vibration signal data of the hydraulic pump test bench as an example to verify the case of the method of the present invention. The data are collected from the vibration signal data of the hydraulic pump in three different states: normal state, wear state of the slipper and the swash plate (fault 1), and wear state of the original rotor of the distribution plate (fault 2). The sample set construction method described in step 1 is adopted, and 100 training samples and 400 test samples are generated under each fault state through methods such as training set and test set sample division and data normalization. The present invention verifies the effectiveness of the method of the present invention by designing different degrees of scarcity conditions of fault samples. The specific data conditions are shown in Table 1:

[0075] Table 1. Training dataset under sample scarcity conditions

[0076]

[0077] Step 2: Based on the training set samples obtained in step 1, the fault label information of different samples is used to perform sample association matching, thereby constructing fault association samples and obtaining a fault sample association data set. Taking data condition 1 as an example, sample association matching is first performed on the original unbalanced sample data set. According to the different fault modes of the samples, the types of associated samples are divided into two categories: one category is obtained based on the association matching of samples with the same fault mode, and the other category is obtained based on the association matching of samples with different fault modes.

[0078] Step 2.1, correlation matching of samples with the same fault mode;

[0079] Correlation matching of samples with the same fault mode helps the model identify different sample manifestations of the same fault mode, thereby mining the intrinsic components of the fault characteristics. The correlation matching results of samples with the same fault mode are as follows: Figure 2 As shown. Figure 2 As can be seen, samples of different fault modes are marked with different symbols. Since each sample is related to itself as a sample of the same type, it is also associated with the samples. By building associations between samples, more sample association results will be generated. However, since the number of original normal samples is far greater than the number of fault samples, associating samples with fault modes will lead to further increase in fault information for that fault mode, thereby exacerbating the imbalance of fault information.

[0080] Step 2.2, correlation matching of different fault mode samples;

[0081] Correlation matching of samples with different fault modes helps the model to identify the different components of different fault modes, thereby achieving more accurate fault feature extraction. Figure 3 As shown. Figure 3 It can be seen that for the sample association of different fault modes, in the data space after sample association, more associated samples are used to characterize the association relationship between fault samples and normal samples, while only a small number of associated samples are used to characterize the association relationship between different fault modes, which also aggravates the imbalance of fault information in the data space.

[0082] Step 3: Based on the fault sample association dataset obtained in step 2, perform sample space simplification. By setting the simplification scale, the fault association samples are simplified, thereby constructing a simplified fault sample association dataset. Figure 2 and Figure 3 It can be seen that all the original sample data before simplification participated in the sample association, but both sample association methods exacerbated the imbalance of the fault information in the entire sample space. Therefore, it is necessary to simplify the sample space and balance the fault information in the sample space to improve the data quality.

[0083] Step 3.1, construct the reduction scale;

[0084] According to the reduction scale calculation method described in step 3, according to the data set conditions set in Table 1, the reduction scale calculated for the associated sample sets with the associated attribute of 1 and the associated attribute of 0 under different data conditions is They are (400,400), (225,300), (100,200), (225,225), (100,150), and (100,100) respectively.

[0085] Step 3.2, sample space reduction;

[0086] The results of the association matching of the same and different fault mode samples after reduction are shown in the figure below. Figure 4 and Figure 5 As shown in the figure, after sample reduction, for the three types of fault samples with fault labels (normal, fault-1, and fault-2), when correlating samples with the same fault mode, the three types of samples involved in the sample association are reduced to varying degrees. The correlation between fault samples of the same type is significantly reduced, which ensures that the number of associated sample pairs of the same type after the final association is relatively balanced. For scenarios where samples are correlated between different fault modes, the number of normal class samples involved in the sample association is also significantly reduced. This allows the sample association process to retain as much information as possible from the minority sample types (fault 1 and fault 2), while reducing redundant information from the majority sample types (normal), thereby balancing the distribution of fault information in the sample space.

[0087] In order to quantitatively measure the fault information before and after sample space reduction, the present invention performs statistics on sample association pairs before and after reduction. The results are shown in Table 2:

[0088] Table 2. Comparison results of fault-related sample data before and after simplification

[0089]

[0090] Table 2 shows that for the six data scenarios, when no sample reduction is performed, the ratio of samples associated with the same fault to samples associated with different faults increases with increasing data imbalance. For extremely unbalanced data condition 6, the ratio reaches 2.40, meaning that each sample associated with different faults corresponds to 2.4 samples associated with the same fault. Even after sample association, the imbalance in fault information persists. In the reduced sample space, however, the ratio between samples associated with the same fault and samples associated with different faults remains relatively stable, showing no significant fluctuations with increasing data imbalance, and the average ratio remains around 0.5:1. The reduced fault association dataset maintains a relatively stable distribution of fault information for datasets of varying degrees of imbalance, providing a sound data foundation for subsequent robust fault feature extraction.

[0091] Step 4. Construct and train a fault diagnosis model based on the twin consistency fault representation self-learning network. By selecting the model architecture and parameters, construct a fault diagnosis model based on the twin consistency fault representation self-learning network. Based on the simplified fault sample association data set obtained in step 3, complete the training and optimization of the fault diagnosis model by setting the training hyperparameters.

[0092] Step 4.1, design the fault diagnosis model architecture;

[0093] This paper proposes a fault diagnosis model based on twin consistency fault characterization self-learning network. The architecture of the model is as follows: Figure 6 The model structure parameters are shown in Table 3:

[0094] Table 3. Structural parameters of the fault diagnosis model based on the twin consistency fault representation self-learning network

[0095]

[0096] The present invention is based on Python language and uses the PyTorch deep learning framework to build a fault diagnosis model based on the above parameter configuration. The hyperparameters of the fault diagnosis model training are shown in Table 4:

[0097] Table 4. Fault diagnosis model training hyperparameter configuration

[0098]

[0099] Step 4.2, training the fault diagnosis model;

[0100] Based on the fault diagnosis model architecture parameters and training parameter configurations in Tables 3 and 4, taking data condition 2 (100 normal - 20 fault 1 - 20 fault 2) as an example, the model training process is as follows: Figure 7 Shown: By Figure 7It can be seen that the loss function based on the root mean square error (MSE) steadily decreases during the model training process and eventually converges to near 0, indicating that the model parameters are updated and optimized normally. After the model training is completed, all training data are matched based on sample labels, and the fault-related samples are input into the trained fault model in pairs. The twin similarity of the fault-related samples in the training set is as follows: Figure 8 Shown: By Figure 8 As can be seen, all samples in the training set are labeled Normal, Fault 1, and Fault 2, respectively. Therefore, there are three types of fault associations when correlating samples with different fault labels. For samples associated with the same fault label, their twin similarity is close to 1. For samples with different fault labels, the twin similarity curve is divided into three levels, all close to 0, due to the three types of association. This indicates that the trained model is able to effectively extract consistent twin fault features for associated samples with the same fault label. However, for samples associated with different fault labels, the extracted fault features lack twin consistency information, demonstrating that the model is able to specifically extract better generalized fault features for samples with different fault modes.

[0101] Step 5: After completing the fault diagnosis model training, based on the feature encoder in the trained fault diagnosis model, extract the fault features for the training set samples. In the case of the sample situation of 100 normal - 20 fault 1 - 20 fault 2, the feature encoding module in the fault diagnosis model is used to extract features for all the samples in the training set and the test set. The results are as follows: Figure 9 and Figure 10 As shown in the figure, after the fault diagnosis model proposed in this invention extracts fault features from the training and test sets, the distribution of each fault feature in the low-dimensional space is uniform. The fault features of the same type of test samples are compactly distributed, while the fault features of different types of test samples are evenly distributed in the feature space. Therefore, the proposed method can extract accurate and robust fault features for samples, effectively supporting the subsequent fault sample label identification work.

[0102] Step 6: Measure the twin similarity of fault features. Based on the fault features of the test sample and the training set samples obtained in step 5, use the feature fusion measurement module in the trained fault diagnosis model to measure the twin similarity of the fault features between the features.

[0103] Step 6.1, reconstructing the fault characteristics of the test sample;

[0104] In this case, taking test sample No. 30 in the test set under three fault modes, namely normal state, wear state of slipper and swash plate (fault 1), and wear state of original rotor of valve plate (fault 2), its fault features are spliced with the fault features of the training set samples to reconstruct the fault features of each test sample.

[0105] Step 6.2, fault feature twin similarity measurement;

[0106] The reconstructed fault features are input into the trained fault diagnosis model, and the twin similarities between the fault features of the 30th test sample under the three fault modes and the fault features of all known samples are obtained as follows: Figure 11 shown.

[0107] Depend on Figure 11 It can be seen that for the test sample with the fault label of normal, the fault feature twin similarity between this sample and all known training samples with the fault label of normal is the highest and the similarity is close to 1, while for the other two types of known training samples, their fault feature twin similarities are small and close to 0, so it can be judged that the fault label of the test sample is normal; for the test sample with the fault label of slipper and inclined plate wear (fault 1), the fault feature twin similarity between this sample and all known training samples with the fault label of fault 1 is the highest and the similarity is close to 1, while with the known training samples of normal and fault 2, its fault feature twin similarity is small and close to 0, so it is judged that the fault label of the test sample is fault 1; similarly, the fault label of the last type of test sample can be judged to be fault 2.

[0108] Step 7: Identify the fault mode of the test set samples. Based on the above fault mode identification process of the test samples, the fault diagnosis of the test samples under data condition 2 is performed, and the confusion matrix of the diagnosis results is as follows: Figure 12 As shown in the figure, the vertical axis represents the true label, and the horizontal axis represents the predicted label. The figure shows that for test samples with the true label of Normal, 94% of the samples were diagnosed as Normal, and 6% of the samples were diagnosed as Fault 1. For test samples with the true label of Fault 1, all were diagnosed as Fault 1. The same applies to Fault 2. Therefore, for data case 1, the final test accuracy is 98.17%.

[0109] Based on the various unbalanced data conditions set in Table 1, the fault diagnosis performance of the fault diagnosis method proposed in the present invention under different data conditions is verified. The proposed method is compared with the fault diagnosis model based on the traditional neural network (T-Siam) and the influence of different hyperparameter values (margin) on the fault diagnosis accuracy under various unbalanced data conditions in the fault diagnosis model based on the traditional twin network is analyzed. The results are as follows: Figure 13 shown.

[0110] Depend on Figure 13 It can be seen that with the reduction of fault samples in the small sample data set and the deepening of the imbalance in the number of samples of different fault modes, the fault diagnosis accuracy of the two fault diagnosis methods shows a downward trend. In addition, for the fault diagnosis model based on the traditional neural network (T-Siam), different hyperparameter setting values have a significant impact on the performance of the method. It can be seen from Table 5 that when the hyperparameter value is set to 0.5, a higher diagnostic accuracy can be achieved under each small sample data set, and when the hyperparameter value is greater than or less than 0.5, the method performance decreases. The optimal hyperparameter interval is determined to be [0.1, 1.0], but it is still difficult to determine the specific value of the optimal hyperparameter. The comparison results of the diagnostic accuracy of the method of the present invention and the fault diagnosis method based on the traditional twin network under different data conditions are shown in Table 5:

[0111] Table 5. Comparison of diagnostic accuracy between the proposed method and the traditional twin network-based fault diagnosis method under different data conditions

[0112]

[0113] As shown in Table 5, the method proposed in the present invention achieves higher fault diagnosis accuracy under various data conditions and can still achieve higher diagnosis accuracy under the extreme data condition of data condition 6.

[0114] In addition, in order to verify the effectiveness of the present invention, a comparative analysis was conducted with the current mainstream fault diagnosis model based on deep learning. Finally, the fault diagnosis performance of the method proposed in the present invention, the fault diagnosis model based on the traditional twin network (comparison method 1), and the autoencoder model based on the deep convolutional neural network (comparison method 2) were compared and analyzed. The results are as follows: Figure 14 As shown in Table 6.

[0115] Table 6. Comparison of diagnostic accuracy of different fault diagnosis methods

[0116]

[0117] Depend on Figure 14 As can be seen, when fault samples are sufficient, all three fault diagnosis methods achieve high fault diagnosis accuracy. However, as the number of fault samples decreases, the performance of the deep convolutional neural network-based autoencoder model declines significantly, demonstrating that data-driven deep learning-based fault diagnosis models are highly sensitive to data changes. Table 6 shows that the fault diagnosis model proposed by this method maintains superior fault diagnosis accuracy compared to both methods under various imbalanced data conditions. This demonstrates that the proposed method exhibits excellent fault diagnosis performance and model adaptability in diverse sample imbalance scenarios and under small sample data conditions.

[0118] In summary, the present invention proposes a small sample fault diagnosis method based on self-learning of twin consistency fault representation. This method can perform sample correlation matching based on the fault correlation of different samples for the original unbalanced small sample data set, and redistribute the fault information of the sample space through sample reduction. Furthermore, a fault diagnosis model based on the twin consistency fault representation self-learning network is constructed. By learning and automatically measuring the twin similarity of fault features between related samples, the twin consistency fault representation extraction of the same sample is realized, and a more robust fault diagnosis model is obtained. Finally, the model is used to autonomously measure the twin similarity of fault features of the sample to be tested and the known sample, and the fault diagnosis of the sample to be tested is completed in combination with the measurement results. The method has advanced technology and strong innovation, and has good practicality and operability.

[0119] The above-described embodiments merely express the implementation methods of the present invention, but should not be understood as limiting the scope of the present invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A small sample fault diagnosis method based on twin consistency fault representation self-learning, characterized by: The small sample fault diagnosis includes the following steps: Step 1: Collect equipment status monitoring data to construct a sample data set, preprocess the sample data set to obtain training set samples and test set samples; Step 2: Based on the training set samples obtained in step 1, the fault label information of different samples is used to perform sample association matching, construct fault association samples, and obtain a fault sample association data set; wherein, the sample association matching is specifically as follows: according to the fault mode of the sample, the samples with the same fault mode are matched as positively associated samples, and the samples with different fault modes are matched as negatively associated samples; the combination label of the samples with the same label is 1, and the combination label of the samples with different labels is 0; Step 3: Based on the fault sample association dataset obtained in step 2, perform sample space simplification. By setting a simplification scale, the fault association samples are simplified to construct a simplified fault sample association dataset. This includes the following sub-steps: Step 3.1, construct the reduction scale; For the training sample set , N is the number of failure modes. Considering the relative size of the distribution of the number of samples of different failure modes in the training data set, we have , ;in, The failure mode is The training sample set, The failure mode is The training sample set, is the minimum number of samples for each failure mode in the training data set, is the median number of samples of each failure mode in the training data set, and Each represents a set of training samples of a fault mode; for the associated sample set with an associated attribute of 1, the reduction scale ; For the associated sample set with an associated attribute of 0, the scale is simplified ; Step 3.2, sample space reduction; For the fault sample association dataset constructed in step 2 、 ,in, The associated sample set constructed for the same fault mode sample, since there are N kinds of fault modes in total, Randomly select and keep N number of samples, and remove the remaining samples to complete The sample space of simplifies; For the associated sample sets constructed for different fault mode samples, among N kinds of fault modes, for the fault associated samples obtained by sample association with normal samples, randomly select and retain (N-1)* The number of samples is eliminated and the rest of the samples are completed The sample space is simplified; the simplified fault-related sample set is obtained; Step 4: Construct and train a fault diagnosis model based on the twin-consistent fault representation self-learning network. Specifically, by selecting the model architecture and parameters, a fault diagnosis model based on the twin-consistent fault representation self-learning network is constructed. Based on the fault sample association dataset obtained in step 3, the training hyperparameters are set to complete the training and optimization of the fault diagnosis model. This includes the following sub-steps: Step 4.1: Design a fault diagnosis model architecture based on a twin-consistent fault representation self-learning network; the fault diagnosis model consists of a fault feature encoding module and a feature fusion measurement module; First, for a pair of related samples and Use the feature encoding module to extract features and obtain fault features and ,in Represents the use of the encoder module to perform linear transformation operations on the input samples. represents the nonlinear transformation operation; then, the feature fusion measurement module is used to measure the twin similarity of the above fault features, and the measurement result is , where For the feature splicing operation, the two fault features are spliced along the sample dimension in a way that ensures the feature dimension remains unchanged. In order to use the feature measurement module based on convolutional neural network, the twin similarity of fault features is measured by convolution operation on the input; finally, the twin similarity of fault features of this pair of associated samples is calculated. ; Step 4.2, training the fault diagnosis model; Based on the fault diagnosis model constructed in step 4.1, the root mean square error is the loss function, where for the kth group of fault-related samples , is the twin similarity prediction value, is the twin similarity label value; and the parameters of the fault diagnosis model are updated; Step 5: Extract fault features of the test samples and training set samples. Specifically: After completing the fault diagnosis model training, based on the trained fault diagnosis model, extract fault features for the test samples and training set samples respectively; Step 6: Measure the fault feature twin similarity. Specifically, for the fault features of the test sample and the training set sample obtained in step 5, use the feature fusion measurement module in the trained fault diagnosis model to measure the fault feature twin similarity between the features. Step 7: Based on the fault feature twin similarity measurement results obtained in step 6, the fault mode of the sample to be tested is identified by setting a label recognition strategy based on the unbalanced sample distribution.

2. A small sample fault diagnosis method based on twin consistency fault characterization self-learning according to claim 1, characterized in that: In the step 1, the sample data set construction method includes using a sliding window to perform sample division on the condition monitoring data, and normalization processing; The following sub-steps are included: Step 1.1, dividing the training set and test set samples; The sample data set is constructed for the equipment's condition monitoring data. First, for the collected condition monitoring data of each fault mode, multiple samples are generated using sliding window cutting. Suppose the condition monitoring data under a certain fault mode Length is , sample cutting is performed with a window of length w and a moving step of s, and we get Generate samples and select training samples, and the remaining samples as test samples; Step 1.2, data normalization; Normalize each sample S in the training data to obtain the normalized training sample set and test sample set. The normalization formula is: ， in, is a signal sequence consisting of training samples; is the maximum value in the signal sequence composed of training samples, is the minimum value in the signal sequence composed of training samples, The normalized sample is obtained after the original sample is normalized; Finally, we get the training data set and test dataset , a total of N failure modes, among which and Represents the training sample set and test sample set with failure mode N.

3. A small sample fault diagnosis method based on twin consistency fault characterization self-learning according to claim 2, characterized in that: The second step includes the following sub-steps: Step 2.1, correlation matching of samples with the same fault mode; For the training sample set with fault mode i, two samples are randomly selected from it , ,in represents the fault correlation operator; for samples with the same fault mode, its correlation attribute is 1; Step 2.2, correlation matching of different fault mode samples; For failure modes and The training sample set 、 , take one sample from each , ,For samples with different failure modes, their associated attributes are 0; After matching the fault samples of all training samples, the fault sample association dataset is finally obtained. 、 , where the subscript 、 are the number of associated sample pairs whose associated attributes are 1 and 0, respectively.

4. The small sample fault diagnosis method based on twin consistency fault characterization self-learning according to claim 1 is characterized by: In step 4.1, the fault feature encoder is composed of a dual-branch network with shared weights, and the feature fusion measurement module is composed of a convolutional network; In step 4.2, the gradient descent strategy is used to update the parameters of the fault diagnosis model. Assume that the parameters of the fault diagnosis model are , the loss function used for training the fault diagnosis model is , then the optimization objective is : The parameter update strategy is: ， Where, is the initial parameter, is the updated parameter; is the learning rate; when When it gradually converges to 0, it means that the fault diagnosis model has optimized a set of suitable parameters. , so that the loss function Get the minimum value.

5. A small sample fault diagnosis method based on twin consistency fault characterization self-learning according to claim 4, characterized in that: In step 5, feature extraction is specifically as follows: after completing the training of the fault diagnosis model, based on the fault feature coding module in the fault diagnosis model, the training data set constructed in step 1 is extracted. All training samples in the training set are extracted to obtain the fault features of the training set , the sample to be tested Extract fault features and obtain fault features of the sample to be tested. .

6. A small sample fault diagnosis method based on twin consistency fault characterization self-learning according to claim 5, characterized in that: In step 6, the fault feature twin similarity measurement is specifically as follows: Step 6.1, reconstructing the fault characteristics of the test sample; The fault characteristics of the sample to be tested obtained in step 5 Fault characteristics of the training set All sample fault features in the fault feature set are spliced together. The splicing method is to splice along the sample dimension to ensure that the feature dimension remains unchanged. Finally, the reconstructed fault feature set is obtained through feature splicing. ,in Represents the failure mode N The training samples and failure modes are No. The splicing fault characteristics composed of test samples; Step 6.2, fault feature twin similarity measurement; Using the feature fusion metric module in the fault diagnosis model, calculate the fault feature set obtained in step 6.1 The twin similarity of each splicing fault feature in the training set is obtained, and the twin similarity between the test sample and the training set sample is defined as ,in , represents the total number of failure modes N training samples and test samples The twin similarity set of .

7. A small sample fault diagnosis method based on twin consistency fault characterization self-learning according to claim 6, characterized in that: In step seven, the failure mode identification of the test sample is specifically as follows: Based on the twin similarity results between the test sample and the training set sample obtained in step 6 , the label recognition result of the sample to be tested is , calculate the twin similarity mean of the training set samples and the test samples of different fault modes, and then take the fault mode with the largest similarity mean as the fault mode identification result of the test sample; complete the fault mode identification of the test sample; finally, perform steps five to seven on all test set samples to achieve fault diagnosis of all test samples.

Images