Metalearning fault diagnosis and identification method based on Brown covariance

Abstract

The invention discloses a meta-learning fault diagnosis and identification method based on Brown covariance. The method comprises the following steps: firstly, dividing input source domain data into a support set and a query set according to meta-tasks, and taking target domain data as the query set; then, deep fault features with identification capability are extracted through a Brown covariance feature extraction method; then, the extracted features are input into a prototype network, and fault prototype representations of all categories are generated; and finally, calculating the similarity between the query sample and the category prototype through a high-dimensional measurer, and outputting a fault category. According to the method, the difference of equipment data under different working conditions is fully considered, and efficient classification and accurate prediction in a complex environment are ensured in combination with a Brown covariance feature extraction method, a prototype network and a high-dimensional measurer.

Description

Technical Field

[0001] The invention belongs to the field of artificial intelligence, and in particular relates to a meta-learning fault diagnosis and identification method based on Brownian covariance. Background Art

[0002] Rotating machinery failures are crucial in the daily operation of industrial systems and directly affect the stability, production efficiency and safety of equipment. Fault diagnosis technology identifies and predicts potential fault states by analyzing equipment operation data (such as vibration, temperature and acoustic signals). In recent years, deep learning has attracted widespread attention in the field of fault diagnosis due to its excellent feature extraction and recognition capabilities, and has shown good development prospects in application scenarios such as intelligent manufacturing, transportation and energy equipment monitoring. However, deep learning methods rely on a large amount of labeled data, which leads to many challenges in industrial applications, especially in data scarcity and cross-domain environments. Industrial data is often interfered by noise and is non-stationary. The data distribution between the source domain and the target domain is inconsistent due to differences in acquisition equipment, environmental changes and equipment working conditions. In order to improve the performance of the model in small sample and cross-domain scenarios, researchers gradually combine deep learning with meta-learning to enhance the generalization ability and adaptability of the model.

[0003] Nevertheless, these methods still face three major challenges when dealing with rotating machinery fault diagnosis: First, the vibration signal data of normally operating bearings is relatively abundant, while the fault vibration signal data is relatively scarce. Second, there are significant data distribution differences between the source domain and the target domain, which poses severe challenges to intelligent fault diagnosis. Finally, current feature extraction methods often focus on marginal distributions and ignore the joint distribution between features. Therefore, how to train efficient fault diagnosis models under the conditions of scarce samples, cross-domain environmental differences, and insufficient consideration of the joint distribution of features has become a key issue that needs to be solved urgently. Summary of the invention

[0004] To solve the above problems, the present invention discloses a meta-learning fault diagnosis and recognition method based on Brownian covariance. The method first divides the input source domain data into a support set and a query set according to the meta-task, and uses the target domain data as the query set; then, a feature extraction technology of Brownian covariance is designed to extract key features from the support set and query set samples, and obtain feature representations with strong discriminative power through Brownian correlation analysis between samples. Next, the extracted features are input into the prototype network, and a prototype is defined for each category in the prototype network by the representative feature vector calculated by the support set. Finally, in order to achieve more accurate classification, the present invention introduces a high-dimensional measurer to further measure the feature differences between the query sample and the category prototype to quantify their similarities and achieve efficient and robust fault diagnosis classification.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A meta-learning fault diagnosis and recognition method based on Brownian covariance mainly includes four core steps: meta-task division, feature extraction of Brownian covariance, category prototype generation and high-dimensional quantifier measurement. The specific steps are as follows:

[0007] a. Meta-task division

[0008] The original vibration signals collected from rotating machinery are divided into source domain dataset D S and the target domain dataset D T , where D S ≠D T . D S The dataset is divided into support sets and queryset Among them, x represents data, y represents label, and N <M,S S Used to train the model and save the trained model, S Q For model validation, D T The dataset is divided into query sets Used for model prediction.

[0009] b. Feature extraction of Brownian covariance

[0010] b.1 The vibration data first enters three convolution modules. Each convolution consists of a one-dimensional convolution layer, a maximum pooling layer, a batch normalization layer, and a rectified linear unit. The features are extracted according to the following formula to obtain low-dimensional feature information. Where h is the height, w is the width, and s is the number of spatial channels.

[0011]

[0012] f2=RELU(BN(Maxpool(Conv(f1)))) (2)

[0013] U=RELU(BN(Maxpool(Conv(f2)))) (3)

[0014] Among them, BN(·) represents the batch normalization layer and RELU(·) represents the activation function.

[0015] b.2 Then, the low-dimensional feature information U is input into the feature extraction module of Brownian covariance, and then the feature information is reshaped into and Structure, where represents each column of U, Indicates that each row after transposition is selected Observe the computation as a random tensor X;

[0016] b.3X h and X k The calculation process is the same as X h Take for example, first calculate the feature matrix X h The squared Euclidean distance between the hth column and the kth row in The calculation is as follows:

[0017]

[0018] in, Represents that each element is 1, I is the unit matrix, represents the Hadamard product, T represents the matrix transpose, (P) sym =(P+P T ),here,

[0019] b.4 Calculation by square root The Euclidean distance The feature matrix A is obtained by subtracting The row average, column average and average of all elements are calculated. The calculation process is as follows:

[0020]

[0021] b.5 Calculate the Brownian matrix of the sample after feature extraction using formula (6), and then input A into the 2×2 maximum pooling layer:

[0022] A'=Maxpool(A) (7)

[0023] c. Category prototype generation

[0024] Use the feature vectors in the support set to generate the prototype of each category. Specifically, the feature vectors of the samples of the same category in A' are averaged to obtain the prototype representation c of the category. k , the formula is as follows:

[0025]

[0026] in, represents the kth prototype, represents the dimension of m feature embedding, N represents N classification tasks, S k represents the query set data of category k, f s (·) represents the feature vector of the query set data after feature extraction.

[0027] d. High-dimensional metric measurement

[0028] d.1 Calculate the distance from each sample of the query set in the target domain to each prototype in turn to evaluate the closest prototype and predict the query set subset label of the target domain, and then use the Softmax function to calculate the support set data sample embedding and prototype c k The distance prediction P between them is expressed as follows:

[0029]

[0030] Among them, f q (·) represents the feature vector of the support set data after feature extraction, represents the i-th prototype;

[0031] d.2 Predicted probability of classification based on each prototype The maximum value P max To predict the final classification result. Finally, the loss function is optimized by the negative log probability between the predicted label and the true label between the target domain. The calculation is as follows:

[0032]

[0033] Beneficial effects of the present invention:

[0034] First: Combining small sample learning with meta-learning, by constructing support sets and query sets to increase the diversity of training data, the model can still learn rich feature information under limited sample conditions.

[0035] Second: A new prototype network classification framework is proposed to effectively integrate the multi-dimensional features extracted from the support set and the query set. By generating category prototypes and performing similarity evaluation on query samples, this framework can capture feature interaction information and significantly improve the consistency and reliability of feature representation.

[0036] Third: In order to solve the problem that the feature extraction method ignores the joint distribution, a feature extraction method based on Brownian covariance is designed. By analyzing the marginal distribution and joint distribution between sample features, it can effectively identify and extract features with significant discriminative ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of a meta-learning fault diagnosis and identification method based on Brownian covariance according to the present invention;

[0038] Figure 2 is a structural schematic diagram of a meta-learning fault diagnosis and identification method based on Brownian covariance in an embodiment of the present invention;

[0039] Figure 3is a schematic diagram of the structure of a feature extraction module of Brownian covariance in an embodiment of the present invention;

[0040] Figure 4 is a schematic diagram of accuracy values ​​on a CWRU data set in an embodiment of the present invention;

[0041] Figure 5 is a schematic diagram of the accuracy value on the PHM2012 dataset in an embodiment of the present invention;

[0042] Figure 6 is a schematic diagram of the F1 score on the CWRU dataset in an embodiment of the present invention;

[0043] Figure 7 It is a schematic diagram of the F1 score on the PHM2012 dataset in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0045] As shown in the figure, a meta-learning fault diagnosis and identification method based on Brownian covariance described in the present invention includes the following steps:

[0046] a. Meta-task division

[0047] The original vibration signals collected from rotating machinery are divided into source domain dataset D S and the target domain dataset D T , where D S ≠D T . D S The dataset is divided into support sets and queryset Among them, x represents data, y represents label, and N <M,S S Used to train the model and save the trained model, S Q For model validation, D T The dataset is divided into query sets Used for model prediction.

[0048] b. Feature extraction of Brownian covariance

[0049] b.1 The vibration data first enters three convolution modules. Each convolution consists of a one-dimensional convolution layer, a maximum pooling layer, a batch normalization layer, and a rectified linear unit. The features are extracted according to the following formula to obtain low-dimensional feature information. Where h is the height, w is the width, and d is the number of spatial channels.

[0050]

[0051] f2=RELU(BN(Maxpool(Conv(f1)))) (2)

[0052] U=RELU(BN(Maxpool(Conv(f2)))) (3)

[0053] Among them, BN(·) represents the batch normalization layer, RELU(·) represents the activation function

[0054] b.2 Then, the low-dimensional feature information U is input into the feature extraction module of Brownian covariance, and then the feature information is reshaped into and Structure, where represents each column of U, Indicates that each row after transposition is selected Observe the computation as a random tensor X;

[0055] b.3X h and X k The calculation process is the same. h Take for example, first calculate the feature matrix X h The squared Euclidean distance between the hth column and the kth row in The calculation is as follows:

[0056]

[0057] in, Represents that each element is 1, I is the unit matrix, represents the Hadamard product, T represents the matrix transpose, (P) sym =(P+P T ),here,

[0058] b.4 Calculation by square root The Euclidean distance The feature matrix A is obtained by subtracting The row average, column average and average of all elements are calculated. The calculation process is as follows:

[0059]

[0060] b.5 Through formula (6), we calculate the Brownian matrix of the sample after feature extraction. Then input A into the 2×2 maximum pooling layer:

[0061] A'=Maxpool(A) (7)

[0062] c. Category prototype generation

[0063] Use the feature vectors in the support set to generate the prototype of each category. Specifically, the feature vectors of samples of the same category in A' are averaged to obtain the prototype representation c of the category. k , the formula is as follows:

[0064]

[0065] in, represents the kth prototype, represents the dimension of m feature embedding, N represents N classification tasks, S k represents the query set data of category k, f s (·) represents the feature vector of the query set data after feature extraction.

[0066] d. High-dimensional metric measurement

[0067] d.1 Calculate the distance from each sample of the query set in the target domain to each prototype in turn to evaluate the closest prototype and predict the query set subset label of the target domain, and then use the Softmax function to calculate the support set data sample embedding and prototype c k The distance prediction P between them is expressed as follows:

[0068]

[0069] Among them, f q (·) represents the feature vector of the support set data after feature extraction, represents the i-th prototype;

[0070] d.2 Predicted probability of classification based on each prototype The maximum value P max To predict the final classification result. Finally, the loss function is optimized by the negative log probability between the predicted label and the true label between the target domain. The calculation is as follows:

[0071]

[0072] For the embodiment of the present invention, the experimental analysis is performed as follows:

[0073] The proposed method was implemented using PyTorch 1.7.0 and NVIDIA

[0074] GeForce RTX 4070 GPU was used for training. In the classification task, accuracy was selected as the evaluation indicator:

[0075]

[0076] Among them, TP means predicting the positive class as the positive class number, TN means predicting the negative class as the negative class number, FP means predicting the negative class as the positive class number, and FP means predicting the positive class as the negative class number.

[0077] In order to verify the effectiveness of the JDBPN proposed in this paper, JDBPN and other existing methods are compared on two public bearing datasets CWRU and PHM2012. The results are shown in Table 1.

[0078] Table 1 Performance comparison of different algorithms on CWRU and PHM2012 datasets (%)

[0079]

[0080]

[0081] It can be seen that JDBPN performs well in comparison with other existing deep learning algorithms and demonstrates more excellent performance. Compared with traditional methods, JDBPN not only fully considers the differences between different working conditions, but also provides effective solutions to the challenges faced in the field of fault diagnosis through meta-learning methods and effective use of Brownian covariance. This achievement fully demonstrates the effectiveness of the proposed method and shows its wide adaptability and superiority in a variety of application scenarios. In addition, the experimental results of the embodiments of the present invention in various task scenarios of the two data sets, including indicators such as accuracy and F1 score, have been shown in detail. Figure 4 , Figure 5 , Figure 6 and Figure 7 It is worth noting that in the CWRU dataset, we selected the most difficult task scenarios 0HP-3HP in cross-domain tasks as the main research objects.

[0082] In summary, compared with other advanced deep learning methods, the method proposed in the present invention shows significant effectiveness. The present invention successfully alleviates the differences between the source domain and the target domain by combining deep learning with meta-learning methods. The Brownian covariance is used to enhance the data feature extraction capability, so that the model can learn more discriminative feature representations. In addition, the introduction of high-dimensional measurers further improves the performance of the model in prototype classification, thereby effectively achieving accurate classification. These innovative measures not only improve the classification performance of the model, but also provide strong support for practical applications in the field of fault diagnosis.

[0083] The present invention combines small sample-based learning with meta-learning, and increases the diversity of training data by constructing support sets and query sets. This method enables the model to learn richer feature information when faced with limited training samples.

[0084] This paper designs a new prototype network classification framework that effectively integrates multi-dimensional features extracted from the support set and the query set. By constructing prototypes of each category and performing similarity evaluation on query samples, the framework can capture the interactive information between different features and significantly improve the consistency and reliability of feature representation.

[0085] In order to solve the problem that the feature extraction method ignores the joint distribution, the present invention designs a feature extraction method of Brownian covariance, which can effectively identify and extract features with significant discriminative ability by analyzing the marginal distribution and joint distribution between sample features.

[0086] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A meta-learning fault diagnosis and identification method based on Brownian covariance, characterized by: The steps include: (1) Meta-task division: The data is divided into source domain and target domain. The source domain data is used to train and generate category prototypes, and the target domain data is used to test and verify the cross-domain adaptability of the model. (2) Feature extraction of Brownian covariance: Extract discriminative features from samples in the source and target domains, and improve the cross-domain sensitivity of features by analyzing the correlation between marginal and joint distributions; (3) Category prototype generation: Obtain the category prototype c for the same type of samples in the support set. k , enhance the classification stability and generalization ability of the model; (4) High-dimensional metric measurement: Evaluate the similarity between query samples and category prototypes in high-dimensional space, significantly improve the discrimination and robustness of classification, and ensure efficient and accurate fault diagnosis.

2. The meta-learning fault diagnosis and identification method based on Brownian covariance according to claim 1 is characterized in that: The meta-task division module described in step (1) is specifically as follows: The original vibration signals collected from rotating machinery are divided into source domain dataset D s and the target domain dataset D T , where D S ≠D T ;D S The dataset is divided into support sets and queryset And N <M,S S Used to train the model and save the trained model, S Q For model validation, D T The dataset is divided into query sets Used for model prediction.

3. The meta-learning fault diagnosis and identification method based on Brownian covariance according to claim 1 is characterized in that: The feature extraction module of Brownian covariance described in step (2) is specifically as follows: a.1 The vibration data first enters three convolution modules. Each convolution consists of a one-dimensional convolution layer, a maximum pooling layer, a batch normalization layer, and a rectified linear unit. The features are extracted according to the following formula to obtain low-dimensional feature information. Where h is the height, w is the width, and d is the number of spatial channels; f2=RELU(BN(Maxpool(Conv(f1)))) (2) U=RELU(BN(Maxpool(Conv(f2)))) (3) Among them, BN(·) represents the batch normalization layer, RELU(·) represents the activation function; a.2 Then, the low-dimensional feature information U is input into the feature extraction module of Brownian covariance, and then the feature information is reshaped into and Structure, where represents each column of U, Indicates that each row after transposition is selected Observe the computation as a random tensor X; a.3X h and X k The calculation process is the same as X h Take for example, first calculate the feature matrix X h The squared Euclidean distance between the hth column and the kth row in The calculation is as follows: in, Represents that each element is 1, I is the unit matrix, represents the Hadamard product, T represents the matrix transpose, (P) sym =(P+P T ),here, a.4 Calculation by square root The Euclidean distance The feature matrix A is obtained by subtracting The row average, column average and average of all elements are calculated; the calculation process is as follows: a.5 Calculate the Brownian matrix of the sample after feature extraction through formula (6); then input A into the 2×2 maximum pooling layer: A′=Maxpool(A) (7) 4. The meta-learning fault diagnosis and identification method based on Brownian covariance according to claim 1 is characterized in that: The category prototype generation module described in step (3) is specifically as follows: Use the feature vectors in the support set to generate the prototype of each category. Specifically, the feature vectors of the samples of the same category in A′ are averaged to obtain the prototype representation c of the category. k , the formula is as follows: in, represents the kth prototype, represents the dimension of m feature embedding, N represents N classification tasks, S k represents the query set data of category k, f s (·) represents the feature vector of the query set data after feature extraction.

5. The meta-learning fault diagnosis and identification method based on Brownian covariance according to claim 1 is characterized in that: The high-dimensional measurer measurement module described in step (4) is specifically as follows: c.1 Compute each embedding sample of the query set in the target domain to each prototype c k to evaluate the closest prototype and predict the labels of the query set subset of the target domain; Then use the Softmax function to calculate the support set data sample embedding and prototype c k The distance prediction P between them is expressed as follows: Among them, f q (·) represents the feature vector of the support set data after feature extraction; represents the i-th prototype; c.2 Prediction probability of each prototype classification The maximum value P max To predict the final classification result; finally, the loss function is optimized by the negative logarithmic probability between the predicted label and the true label between the target domain. The calculation is as follows:

Images