Industrial process zero sample fault diagnosis method based on similarity consistency relation

Abstract

The invention provides an industrial process zero sample fault diagnosis method based on a similarity consistency relation, and relates to the technical field of fault diagnosis and classification. From the perspective of similarity between faults, on one hand, understanding of experts on the faults is deeply mined, and the similarity between fault categories is constructed based on the knowledge level; and on the other hand, the model is trained by using the known fault data, the data features of the known faults are extracted on the data level, and the similarity between the fault features is quantified from multiple angles. By solving a group of consistency parameters, mapping of faults from knowledge similarity to data similarity is constructed. Through the mapping, the abstract and comprehensive similarity relation of the knowledge level can be specifically explained by utilizing the similarity measurement of the data level, and the diagnosis of unseen class faults can be effectively supported.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis and classification, and in particular to an industrial process zero-sample fault diagnosis method based on similarity consistency relationship. Background Art

[0002] With the improvement of industrial automation, equipment fault diagnosis has become a key task to ensure production safety and stable operation of equipment. Traditional fault diagnosis methods generally rely on a large number of fault data samples for training. However, in actual industrial scenarios, since equipment failures may cause serious losses, factories usually do not allow equipment to continue to operate when a failure occurs to collect samples for model training. Therefore, it is difficult to obtain sufficient fault data, and it may even be impossible to obtain usable data.

[0003] To solve this problem, zero-shot learning methods have emerged. Zero-shot learning aims to solve the problem of how to effectively classify and identify samples without training for specific categories. Unlike traditional supervised learning methods, zero-shot learning can identify categories that have never been seen and have not appeared in the training set during the test phase by using auxiliary information for reasoning. Existing zero-shot fault diagnosis methods for industrial processes mainly rely on the semantic / attribute features of the fault, and achieve reasoning and classification of unseen fault categories by sharing attribute mapping with data of seen categories.

[0004] For example, the Chinese patent "Semantic knowledge base, construction method and zero-sample fault diagnosis method for thermal power equipment" proposes a zero-sample fault diagnosis method that uses fault attribute information to train an attribute discriminator. This method extracts the fault semantic attribute information in the thermal power generation process and encodes it into an attribute vector, trains the attribute discriminator with the seen class fault data, and constructs a "data-attribute-attribute discriminator" knowledge base. When encountering an unseen class fault, the attribute discriminator identifies its attribute vector, and the fault mode is determined and knowledge transfer is realized based on shared attributes, thereby realizing zero-sample fault diagnosis of high-end thermal power equipment.

[0005] Another related technical solution is the Chinese patent "A generalized zero-shot fault diagnosis method for industrial processes based on DSECMR-VAE". This method uses variational autoencoders (VAE) to embed features of fault samples and semantic attribute vectors, generate latent variables, and constrain the consistency of embedding by constructing Barlow matrices. This method improves the representation ability of features through cross-modal reconstruction, and finally trains classifiers based on latent variables to achieve the diagnosis of unseen, seen, and mixed faults.

[0006] In summary, the zero-shot fault diagnosis method overcomes the limitations of traditional methods in the absence of data by using semantic attribute information, and provides a new solution for industrial equipment fault diagnosis.

[0007] Chinese patents such as "Semantic knowledge base, construction method and zero-sample fault diagnosis method for thermal power equipment" and "A generalized zero-sample fault diagnosis method for industrial processes based on DSECMR-VAE" have proposed zero-sample fault diagnosis methods based on fault attributes for industrial processes. However, such methods still have certain limitations:

[0008] First, this type of method completely relies on fault semantic information. However, it is very difficult to obtain detailed and accurate fault semantic information and achieve precise attribute decomposition in real situations. Especially when facing complex industrial process failures, the fault semantic description obtained by relying solely on expert experience may not cover all possible fault types.

[0009] Secondly, the existing semantic description of faults is usually based on experience or surface information, which belongs to "shallow knowledge" and cannot fully reflect the complexity and multi-dimensional characteristics of faults. In order to achieve more accurate fault diagnosis, it is necessary to dig deeper into "deep knowledge" and expand the scope of application of zero-sample fault diagnosis problems in complex industrial environments.

[0010] In summary, although the existing zero-shot fault diagnosis methods provide preliminary solutions, their limitations in practice are still key issues that need to be improved. Summary of the invention

[0011] In view of the shortcomings of the prior art, the purpose of the present invention is to propose an industrial process zero-sample fault diagnosis method based on similarity consistency relationship, comprising:

[0012] Step 1: Establish the fault similarity matrix S between the seen fault classes and the seen fault classes in the industrial process E , and the fault similarity matrix S between the seen faults and the unseen faults E u ;

[0013] Step 2: Obtain seen class fault data as input samples, and obtain the category label corresponding to the seen class fault data as output samples, the input samples and the output samples are combined to obtain training samples, and the initial convolutional network is trained based on the training samples. After the training is completed, a fault classification model is obtained, and the fault classification model includes an input layer, a first convolutional layer, a batch normalization layer, a ReLU activation layer, a pooling layer, a second convolutional layer, a batch normalization layer, a ReLU activation layer, a pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer, a third fully connected layer, a Softmax layer and an output layer in the order of data processing;

[0014] Step 3: Input all categories of seen class fault data into the fault classification model, obtain the feature vector output by the second fully connected layer in the fault classification model, and obtain the feature vector of all categories of seen class fault data;

[0015] Step 4: According to the category labels of the seen class fault data, the feature vectors of all categories of the seen class fault data are divided to obtain the feature vectors corresponding to each category of the seen class fault data, and the average value of the feature vectors corresponding to each category of the seen class fault data is calculated to obtain the target vectors f1, f2, ..., f p ;

[0016] Step 5: Target vector f1,f2,...,f according to the seen class faults p , calculate the feature similarity matrix S of the seen class faults and the seen class faults D ;

[0017] Specifically, the target vector f is calculated using five algorithms: Euclidean distance, cosine similarity, Pearson correlation coefficient, Spearman correlation coefficient, and mutual information. i With f j The feature similarity between the seen faults and the seen faults is obtained. D , which is specifically expressed by the following formula:

[0018]

[0019] Among them, the feature similarity matrix S D The vector contained in is denoted as d ij , d ij represents the feature similarity between the seen class faults and the seen class faults, d ij Contains the normalized feature similarity values ​​corresponding to the five algorithms;

[0020] Step 6: Calculate the fault similarity matrix S E and the feature similarity matrix S D The consistency parameter ω=[ω1,ω2,...,ω5];

[0021] Step 7: Input the unseen class fault data into the fault classification model, obtain the feature vector output by the second fully connected layer in the fault classification model, and obtain the feature vector f of the unseen class fault data u ;

[0022] Step 8: Calculate the feature vector f of the unseen fault data using five algorithms: Euclidean distance, cosine similarity, Pearson correlation coefficient, Spearman correlation coefficient, and mutual information.u The target vector f1,f2,...,f p The feature similarity matrix between Among them, the feature similarity matrix Each vector in is represented as d ua , d ua The feature vector f represents the unseen class fault data u With the target vector f a The similarity of ua Contains the normalized feature similarity values ​​corresponding to the five algorithms, a=1,2,…,p;

[0023] Step 9: Use the uniformity parameter ω to adjust the feature similarity matrix Perform a consistent transformation to obtain the fault similarity vector between the unseen class faults and the seen class faults

[0024] Step 10: Target the Fault Similarity Matrix For each row of values ​​in , calculate its similarity vector S with the fault E u* The Euclidean distance is calculated to obtain a target value, and then multiple target values ​​are obtained. The minimum value is obtained from the multiple target values. The category of the unseen class fault data corresponding to the minimum value is the category of the unseen class fault data in step 7.

[0025] Optionally, the fault similarity matrix S between the seen class faults and the seen class faults in step 1 E It is expressed by the following formula:

[0026]

[0027] Among them, the fault similarity matrix S E The numerical value contained in is represented by s ij ,s ij represents the similarity between the i-th seen class fault and the j-th seen class fault, i,j=1,2,…,p, where p is the total number of seen class faults;

[0028] Among them, the fault similarity matrix S between the seen faults and the unseen faults is E u It is expressed by the following formula:

[0029]

[0030] Among them, the fault similarity matrix S E u The numerical value contained in is represented by s ab ,s abIt represents the similarity between the a-th seen fault type and the b-th unseen fault type, a=1,2,…,p, b=1,2,…,q, where q is the total number of unseen fault types.

[0031] Optionally, step 6 specifically includes:

[0032] The consistency parameter is expressed as ω = [ω1, ω2, ..., ω5], and the specific value of the consistency parameter is calculated so that the feature similarity matrix S D The matrix calculated by the consistency parameter is similar to the fault similarity matrix S E The Euclidean distance is the smallest, which is specifically expressed by the following formula:

[0033]

[0034] Among them, C represents the number of algorithms, which is 5;

[0035] Minimize the above formula and obtain the value of each parameter in the consistency parameter ω = [ω1, ω2, ..., ω5].

[0036] Optionally, step 9 is implemented by the following formula:

[0037]

[0038] Optionally, in step 5, the target vector f of the i-th type of seen fault is calculated by Euclidean distance i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula:

[0039]

[0040] Among them, d1(f i ,f j ) is the target vector f of the seen class faults under the Euclidean distance algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, represents the target vector of the i-th type of seen fault in the k-th dimension, represents the target vector of the jth type of seen fault in the kth dimension;

[0041] For d1(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

[0042]

[0043] Optionally, in step 5, the target vector f of the i-th type of seen fault is calculated by cosine similarity i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula:

[0044]

[0045] Among them, d2(f i ,f j ) is the target vector f of the seen faults under the cosine similarity algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, represents the target vector of the i-th type of seen fault in the k-th dimension, represents the target vector of the jth type of seen fault in the kth dimension;

[0046] For d2(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

[0047]

[0048] Optionally, in step 5, the target vector f of the i-th type of seen fault is calculated by using the Pearson correlation coefficient i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula:

[0049]

[0050] Among them, d3(f i ,f j ) is the target vector f of the observed faults under the Pearson correlation coefficient algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, represents the target vector of the i-th type of seen fault in the k-th dimension, represents the target vector of the j-th type of seen fault in the k-th dimension, represents the average value of the target vector of the i-th eigenvector in n dimensions, Represents the average value of the target vector of the i-th feature vector in n dimensions;

[0051] For d3(fi ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

[0052]

[0053] Optionally, in step 5, the target vector f of the i-th type of seen fault is calculated by the Spearman correlation coefficient i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula:

[0054]

[0055] Among them, d4(f i ,f j ) is the target vector f of the observed faults under the Spearman correlation coefficient algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, represents the target vector of the i-th type of seen fault in the k-th dimension, represents the target vector of the j-th type of seen fault in the k-th dimension, is the target vector f i middle the sort position in the target vector;

[0056] For d4(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

[0057]

[0058] Optionally, in step 5, the target vector f of the i-th type of seen fault is calculated by mutual information i and the target vector f of the jth seen fault j The feature similarity value between them is expressed by the following formula:

[0059]

[0060] Among them, d4(f i ,f j ) is the target vector f of the seen faults under the mutual information algorithm i and the target vector f of the jth seen fault j The feature similarity value between represents the normalized feature similarity value. Under the mutual information algorithm, d4(f i ,f j )and equal, p(x,y) is f i and f j The joint probability of taking specific values ​​of x and y; p(x) and p(y) are f i and f j Their respective marginal probabilities.

[0061] The beneficial effects of adopting the above technical solution are:

[0062] The present invention starts from the perspective of similarity between faults. On the one hand, it deeply mines the experts' understanding of faults and constructs the similarity between fault categories based on the knowledge level. On the other hand, it uses the data training model of seen fault types to extract the data features of seen fault types at the data level, and quantifies the similarity between fault features from multiple perspectives. By solving a set of consistency parameters, a mapping of faults from knowledge similarity to data similarity is constructed. Through this mapping, not only can the similarity measurement at the data level be used to realize the specific interpretation of the abstract and comprehensive similarity relationship at the knowledge level, but it can also effectively support the diagnosis of unseen fault types. The core goal of the present invention is to overcome the dependence of existing methods on the semantic description of faults from the perspective of similarity between faults, deeply mine the similarity relationship between faults, and provide a new perspective of zero-sample diagnosis methods to promote the popularization and effect improvement of zero-sample fault diagnosis in actual industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of a process flow of a zero-sample fault diagnosis method for an industrial process based on a similarity consistency relationship in an embodiment of the present invention;

[0064] Figure 2 Schematic diagram of the convolutional neural network structure in an embodiment of the present invention;

[0065] Figure 3 TE process flow diagram in an embodiment of the present invention;

[0066] Figure 4 is a schematic diagram of a fault category similarity matrix based on expert knowledge in an embodiment of the present invention;

[0067] Figure 5Schematic diagram of the confusion matrix of the diagnosis results of group 2 in the comparative test in the embodiment of the present invention, wherein Figure a is a schematic diagram of the confusion matrix of the diagnosis results of the unseen fault classes 1, 14, and 15, Figure b is a schematic diagram of the confusion matrix of the diagnosis results of the unseen fault classes 2, 3, and 6, Figure c is a schematic diagram of the confusion matrix of the diagnosis results of the unseen fault classes 4, 10, and 13, Figure d is a schematic diagram of the confusion matrix of the diagnosis results of the unseen fault classes 5, 8, and 9, and Figure e is a schematic diagram of the confusion matrix of the diagnosis results of the unseen fault classes 7, 11, and 12. DETAILED DESCRIPTION

[0068] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0069] In view of the problems existing in the prior art, the present invention starts from the perspective of similarity between faults. On the one hand, it deeply mines the experts' understanding of faults and constructs the similarity between fault categories based on the knowledge level; on the other hand, it uses the data of seen faults to train the model, extracts the data features of seen faults at the data level, and quantifies the similarity between fault features from multiple perspectives. By solving a set of consistency parameters, a mapping of faults from knowledge similarity to data similarity is constructed. Through this mapping, not only can the similarity measurement at the data level be used to realize the specific interpretation of the abstract and comprehensive similarity relationship at the knowledge level, but it can also effectively support the diagnosis of unseen faults. The core goal of the present invention is to overcome the dependence of existing methods on the semantic description of faults from the perspective of similarity between faults, deeply mine the similarity relationship between faults, provide a new perspective of zero-sample diagnosis methods, and promote the popularization and effect improvement of zero-sample fault diagnosis in actual industrial applications.

[0070] Specifically, the present invention provides an industrial process zero-sample fault diagnosis method based on similarity consistency relationship, combined with Figure 1 , which may include the following steps:

[0071] Step 1: Establish the fault similarity matrix S between the seen fault classes and the seen fault classes in the industrial process E , and the fault similarity matrix S between the seen faults and the unseen faults E u ;

[0072] Among them, this step can be established based on expert knowledge. In this step, the location of the fault, influencing factors, manifestations (such as pressure, flow, temperature, etc.) and other expert experience and cognition information are comprehensively considered to quantify the similarity between each fault category.

[0073] Among them, the fault similarity matrix S between the seen class faults and the seen class faults is EIt is expressed by the following formula:

[0074]

[0075] Among them, the fault similarity matrix S E The numerical value contained in is represented by s ij ,s ij represents the similarity between the i-th seen class fault and the j-th seen class fault, i,j=1,2,…,p, where p is the total number of seen class faults;

[0076] Among them, the fault similarity matrix between the seen faults and the unseen faults is It is expressed by the following formula:

[0077]

[0078] Among them, the fault similarity matrix The numerical value contained in is represented by s ab ,s ab It represents the similarity between the a-th seen fault type and the b-th unseen fault type, a=1,2,…,p, b=1,2,…,q, where q is the total number of unseen fault types.

[0079] Step 2: Obtain the seen class fault data as input samples, and obtain the category label corresponding to the seen class fault data as output samples. The input samples and the output samples are combined to obtain training samples. Based on the training samples, the initial convolutional neural network is trained. After the training is completed, a fault classification model is obtained. Specifically, the training samples are divided into a training set and a test set. The initial convolutional neural network is trained using the training set, and the accuracy of the classification model is evaluated using the test set. When the accuracy reaches the preset threshold, the model parameters are determined.

[0080] Combination Figure 2 The fault classification model includes, in the order of data processing, an input layer, a first convolutional layer, a batch normalization layer, a ReLU activation layer, a pooling layer, a second convolutional layer, a batch normalization layer, a ReLU activation layer, a pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer, a third fully connected layer, a Softmax layer and an output layer; wherein the second fully connected layer is used as a feature extraction layer to extract feature vectors of fault samples.

[0081] Step 3: Input all categories of seen class fault data into the fault classification model, obtain the feature vector output by the second fully connected layer in the fault classification model, and obtain the feature vector of all categories of seen class fault data;

[0082] Step 4: According to the category labels of the seen class fault data, the feature vectors of all categories of the seen class fault data are divided to obtain the feature vectors corresponding to each category of the seen class fault data, and the average value of the feature vectors corresponding to each category of the seen class fault data is calculated to obtain the target vectors f1, f2, ..., f p ;

[0083] In the specific implementation process, after obtaining the feature vector corresponding to each type of fault data, the feature matrix F1, F2, ..., F p , calculate the average value of each feature matrix column by column.

[0084] Step 5: Target vector f1,f2,...,f according to the seen class faults p , calculate the feature similarity matrix S of the seen class faults and the seen class faults D ;

[0085] Specifically, the target vector f is calculated using five algorithms: Euclidean distance, cosine similarity, Pearson correlation coefficient, Spearman correlation coefficient, and mutual information. i With f j The feature similarity between the seen faults and the seen faults is obtained. D , which is specifically expressed by the following formula:

[0086]

[0087] Among them, the feature similarity matrix S D The vector contained in is denoted as d ij , d ij represents the feature similarity between the seen class faults and the seen class faults, d ij Contains the normalized feature similarity values ​​corresponding to the five algorithms;

[0088] Among them, the target vector f of the i-th type of seen fault is calculated by Euclidean distance i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula:

[0089]

[0090] Among them, d1(f i ,f j ) is the target vector f of the seen class faults under the Euclidean distance algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, represents the target vector of the i-th type of seen fault in the k-th dimension, represents the target vector of the jth type of seen fault in the kth dimension;

[0091] For d1(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

[0092]

[0093] Among them, the target vector f of the i-th type of seen fault is calculated by cosine similarity i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula:

[0094]

[0095] Among them, d2(f i ,f j ) is the target vector f of the seen faults under the cosine similarity algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, represents the target vector of the i-th type of seen fault in the k-th dimension, represents the target vector of the jth type of seen fault in the kth dimension;

[0096] For d2(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

[0097]

[0098] Among them, the target vector f of the i-th type of seen fault is calculated by Pearson correlation coefficient i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula:

[0099]

[0100] Among them, d3(f i ,f j ) is the target vector f of the observed faults under the Pearson correlation coefficient algorithm i and the target vector f of the jth seen fault jThe feature similarity value between them, n represents the dimension of the target vector, represents the target vector of the i-th type of seen fault in the k-th dimension, represents the target vector of the j-th type of seen fault in the k-th dimension, represents the average value of the target vector of the i-th eigenvector in n dimensions, Represents the average value of the target vector of the i-th feature vector in n dimensions;

[0101] For d3(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

[0102]

[0103] Among them, the target vector f of the i-th type of seen fault is calculated by the Spearman correlation coefficient i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula:

[0104]

[0105] Among them, d4(f i ,f j ) is the target vector f of the observed faults under the Spearman correlation coefficient algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, represents the target vector of the i-th type of seen fault in the k-th dimension, represents the target vector of the j-th type of seen fault in the k-th dimension, is the target vector f i middle the sort position in the target vector;

[0106] For d4(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

[0107]

[0108] Among them, the target vector f of the i-th type of seen fault is calculated through mutual information i and the target vector f of the jth seen fault jThe feature similarity value between them is expressed by the following formula:

[0109]

[0110] Among them, d4(f i ,f j ) is the target vector f of the seen faults under the mutual information algorithm i and the target vector f of the jth seen fault j The feature similarity value between represents the normalized feature similarity value. Under the mutual information algorithm, d4(f i ,f j )and equal, p(x,y) is f i and f j The joint probability of taking specific values ​​of x and y; p(x) and p(y) are f i and f j Their respective marginal probabilities.

[0111] Step 6: Calculate the fault similarity matrix S E and the feature similarity matrix S D The consistency parameter ω=[ω1,ω2,...,ω5];

[0112] Specifically, the consistency parameter is expressed as ω = [ω1, ω2, ..., ω5], and the specific value of the consistency parameter is calculated so that the feature similarity matrix S D The matrix calculated by the consistency parameter is similar to the fault similarity matrix S E The Euclidean distance is the smallest, which is specifically expressed by the following formula:

[0113]

[0114] Among them, C represents the number of algorithms, which is 5;

[0115] Minimize the above formula and obtain the value of each parameter in the consistency parameter ω = [ω1, ω2, ..., ω5].

[0116] Step 7: Input the unseen class fault data into the fault classification model, obtain the feature vector output by the second fully connected layer in the fault classification model, and obtain the feature vector f of the unseen class fault data u ;

[0117] Step 8: Calculate the feature vector f of the unseen fault data using five algorithms: Euclidean distance, cosine similarity, Pearson correlation coefficient, Spearman correlation coefficient, and mutual information. u The target vector f1,f2,...,fp The feature similarity matrix between Among them, the feature similarity matrix Each vector in is represented as d ua , d ua The feature vector f represents the unseen class fault data u With the target vector f a The similarity of ua Contains the normalized feature similarity values ​​corresponding to the five algorithms, a=1,2,…,p;

[0118] The calculation method in this step is similar to that in step 5.

[0119] Step 9: Use the uniformity parameter ω to adjust the feature similarity matrix Perform a consistent transformation to obtain the fault similarity vector between the unseen class faults and the seen class faults This is achieved specifically through the following formula:

[0120]

[0121] Step 10: Target the Fault Similarity Matrix For each row of values ​​in , calculate its similarity vector with the fault The Euclidean distance is calculated to obtain a target value, and then multiple target values ​​are obtained. The minimum value is obtained from the multiple target values. The category of the unseen class fault data corresponding to the minimum value is the category of the unseen class fault data in step 7.

[0122] The key innovation of the present invention is to propose a zero-sample fault diagnosis method based on knowledge and data similarity consistency, which aims to overcome the existing technology's complete reliance on fault attributes / semantic descriptions for zero-sample fault diagnosis, and provide a new perspective - to achieve zero-sample fault diagnosis from the perspective of similarity between faults.

[0123] Knowledge-based fault similarity uses experts' deep understanding of industrial processes and combines factors such as the specific location of the fault, the scope of the fault's impact, and potential manifestations to measure the similarity relationship between faults. Compared with fault attributes and semantic descriptions, similarity belongs to a more abstract and comprehensive "deep knowledge".

[0124] The core innovation of the present invention is to obtain the consistency relationship by using the similarity at the knowledge and data levels between the seen types of faults through this deep similarity analysis, and to realize the conversion from the data similarity between the unseen type of faults and the seen type of faults to the knowledge similarity when diagnosing the unseen type of faults, thereby realizing zero-sample diagnosis of the unseen type of faults.

[0125] This method provides a new perspective for zero-sample fault diagnosis. By utilizing deeper similarity information, it improves the accuracy and reliability of zero-sample fault diagnosis and has important industrial application value.

[0126] The present invention will be further described in detail below through specific examples.

[0127] ①Introduction to the dataset

[0128] The Tennessee-Eastman (TE) process is a simulation data set that simulates an actual chemical process established by Downs et al. in 1993. This process provides an experimental platform for studying the control, monitoring, and fault diagnosis techniques of this type of process, and helps to fairly demonstrate the effectiveness of its methods.

[0129] TE process structure diagram Figure 3 As shown in FIG. 1 , the process mainly includes five operating units: reactor, condenser, circulating compressor, separator and stripping tower. The fault data set includes 21 types of faults, each of which is described by 41 measured variables and 11 manipulated variables. Each type of fault includes 480 fault data samples. The first 15 types of faults are selected to verify the effectiveness of the present invention. The specific fault description is shown in Table 1.

[0130] Table 1TE process fault description

[0131]

[0132] ②Fault similarity matrix

[0133] The similarity between the faults is quantified by comprehensively considering the location, influencing factors, manifestations (such as pressure, flow, temperature, etc.) of the fault, and other information such as expert experience and cognition. A fault category similarity matrix is ​​established for 15 fault types, such as Figure 4 shown.

[0134] ③Convolutional neural network structure and parameters

[0135] A convolutional neural network model is established as a feature extraction and fault classification model. The designed network architecture includes two convolutional layers with a kernel size of 2×1, generating 32 and 64 convolutional features respectively. The convolutional layer is followed by a batch normalization layer and a ReLU activation layer. Between the convolutional layers, a maximum pooling layer with a window size of 2×1 and a step size of 1 is used. In addition, the fully connected layer contains 100 neurons, the feature output layer contains 30 neurons, and the classification output layer needs to be set according to the number of categories of the specific task (p-class training faults). The training parameters use the Adam optimization algorithm, the maximum training round is 50, the batch size is 128, the initial learning rate is 0.001, L2 regularization is introduced, the regularization parameter is 0.0001, and the segmented learning rate reduction strategy is adopted. The learning rate reduction factor is 0.1 and the reduction period is 47. The dataset is randomly shuffled in each round of training.

[0136] On this basis, steps one to ten are used to realize zero-sample fault diagnosis of TE process.

[0137] To verify the effectiveness of the present invention, the seen faults are divided into 12 categories, the unseen faults are divided into 3 categories, and a total of 15 faults are divided into 5 groups. Tables 2-7 show the diagnostic experimental results of 6 different grouping methods.

[0138] Table 2 Group 1

[0139]

[0140] Table 3 Group 2

[0141]

[0142] Table 4 Group 3

[0143]

[0144] Table 5 Group 4

[0145]

[0146] Table 6 Group 5

[0147]

[0148] Table 7 Group 6

[0149]

[0150] In group 5, when the seen faults were 2-5, 7-10, and 12-15, the diagnostic accuracy rates of the unseen faults were 1, 6, and 11 were 0.5896, 0.9604, and 0.9896, respectively, with an average value of 0.8465, the highest among the 6 groups. The average diagnostic accuracy rate of group 2 was the highest, 0.6739, and the average accuracy rates of the other groups were 0.6133, 0.6632, 0.5898, 0.5464, and 0.6271, respectively.

[0151] The experimental results show that the method of the present invention exhibits effective diagnostic capabilities when dealing with unknown faults, and the grouping of different known faults and unknown faults has a significant impact on the diagnostic results. When the unknown fault is more similar to the known fault, the accuracy of the diagnostic result is higher.

[0152] In order to further verify the effectiveness of the zero-sample fault diagnosis method proposed in this patent, the method of the present invention is experimentally compared with the Chinese patent "Semantic knowledge base, construction method and zero-sample fault diagnosis method for thermal power equipment". The method proposed in this patent uses auxiliary information to decompose the attributes of various faults and trains an attribute learner for each attribute. The types of attribute learners include long short-term memory network (LSTM), naive Bayes (NB) and random forest (RF).

[0153] Two different grouping methods were selected in the comparative experiment, as shown in Tables 8-9, where “-” represents no results.

[0154] Table 8 Comparative experimental results of zero-sample fault diagnosis for group 1

[0155]

[0156] Table 9 Comparative experimental results of zero-sample fault diagnosis for group 2

[0157]

[0158] The experimental results show that in group 1, the method of the present invention has the highest diagnostic accuracy of 0.6827 when processing the unseen faults 9, 13, and 15; the average accuracy of the five groups of experiments is higher than that of LSTM and RF, but lower than that of NB. In group 2, the diagnostic accuracy of the present method is ahead of other methods, especially in the comparison between LSTM and RF, showing a higher average accuracy. Among them, the confusion matrix of the diagnosis results of group 2 is shown in Figure 5As shown, Figure a is a schematic diagram of the confusion matrix of the diagnosis results of unseen fault classes 1, 14, and 15, Figure b is a schematic diagram of the confusion matrix of the diagnosis results of unseen fault classes 2, 3, and 6, Figure c is a schematic diagram of the confusion matrix of the diagnosis results of unseen fault classes 4, 10, and 13, Figure d is a schematic diagram of the confusion matrix of the diagnosis results of unseen fault classes 5, 8, and 9, and Figure e is a schematic diagram of the confusion matrix of the diagnosis results of unseen fault classes 7, 11, and 12, wherein the columns of each figure represent the true fault categories, the rows represent the diagnosed diagnostic categories, the values ​​in the matrix represent the accuracy of the diagnosis, taking the first row of Figure a as an example, when the true category is 1, the accuracy of the diagnostic category 1 is 0.93, the accuracy of the diagnostic category 14 is 0.06, and the accuracy of the diagnostic category 15 is 0.01.

[0159] In summary, the method of the present invention can effectively utilize the fault similarity at the knowledge and data level and successfully realize zero-sample fault diagnosis of unseen faults. From the perspective of similarity between faults, when fault attributes are difficult to obtain, it provides another perspective for realizing zero-sample fault diagnosis by using fault similarity. Of course, this method also has certain limitations, that is, there must be similarity between seen faults and unseen faults and the degree of similarity must be high, otherwise the proposed method will perform poorly.

[0160] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.

Claims

1. A zero-sample fault diagnosis method for industrial processes based on similarity consistency relations, characterized in that: include: Step 1: Establish the fault similarity matrix S between the seen fault classes and the seen fault classes in the industrial process E , and the fault similarity matrix between seen and unseen faults Step 2: Obtain seen class fault data as input samples, and obtain the category label corresponding to the seen class fault data as output samples, the input samples and the output samples are combined to obtain training samples, and the initial convolutional network is trained based on the training samples. After the training is completed, a fault classification model is obtained, and the fault classification model includes an input layer, a first convolutional layer, a batch normalization layer, a ReLU activation layer, a pooling layer, a second convolutional layer, a batch normalization layer, a ReLU activation layer, a pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer, a third fully connected layer, a Softmax layer and an output layer in the order of data processing; Step 3: Input all categories of seen class fault data into the fault classification model, obtain the feature vector output by the second fully connected layer in the fault classification model, and obtain the feature vector of all categories of seen class fault data; Step 4: According to the category labels of the seen class fault data, the feature vectors of all categories of the seen class fault data are divided to obtain the feature vectors corresponding to each category of the seen class fault data, and the average value of the feature vectors corresponding to each category of the seen class fault data is calculated to obtain the target vectors f1, f2, ..., f p ; Step 5: Target vector f1,f2,...,f according to the seen class faults p , calculate the feature similarity matrix S of the seen class faults and the seen class faults D ; Specifically, the target vector f is calculated using five algorithms: Euclidean distance, cosine similarity, Pearson correlation coefficient, Spearman correlation coefficient, and mutual information. i With f j The feature similarity between the seen faults and the seen faults is obtained. D , which is specifically expressed by the following formula: Among them, the feature similarity matrix S D The vector contained in is denoted as d ij , d ij represents the feature similarity between the seen class faults and the seen class faults, d ij Contains the normalized feature similarity values ​​corresponding to the five algorithms; Step 6: Calculate the fault similarity matrix S E and the feature similarity matrix S D The consistency parameter ω=[ω1,ω2,...,ω5]; Step 7: Input the unseen class fault data into the fault classification model, obtain the feature vector output by the second fully connected layer in the fault classification model, and obtain the feature vector f of the unseen class fault data u ; Step 8: Calculate the feature vector f of the unseen fault data using five algorithms: Euclidean distance, cosine similarity, Pearson correlation coefficient, Spearman correlation coefficient, and mutual information. u The target vector f1,f2,...,f p The feature similarity matrix between Among them, the feature similarity matrix Each vector in is represented as d ua , d ua The feature vector f represents the unseen class fault data u With the target vector f a The similarity of ua Contains the normalized feature similarity values ​​corresponding to the five algorithms, a=1,2,…,p; Step 9: Use the uniformity parameter ω to adjust the feature similarity matrix Perform a consistent transformation to obtain the fault similarity vector between the unseen class faults and the seen class faults Step 10: Target the Fault Similarity Matrix For each row of values ​​in , calculate its similarity vector with the fault The Euclidean distance is calculated to obtain a target value, and then multiple target values ​​are obtained. The minimum value is obtained from the multiple target values. The category of the unseen class fault data corresponding to the minimum value is the category of the unseen class fault data in step 7.

2. The method for zero-sample fault diagnosis of industrial processes based on similarity consistency relationship according to claim 1, characterized in that: The fault similarity matrix S between the seen class faults and the seen class faults described in step 1 E It is expressed by the following formula: Among them, the fault similarity matrix S E The numerical value contained in is represented by s ij ,s ij represents the similarity between the i-th seen class fault and the j-th seen class fault, i,j=1,2,…,p, where p is the total number of seen class faults; Among them, the fault similarity matrix between the seen faults and the unseen faults is It is expressed by the following formula: Among them, the fault similarity matrix The numerical value contained in is represented by s ab ,s ab It represents the similarity between the a-th seen fault type and the b-th unseen fault type, a=1,2,…,p, b=1,2,…,q, where q is the total number of unseen fault types.

3. The method for zero-sample fault diagnosis of industrial processes based on similarity consistency relationship according to claim 1, characterized in that: Step 6 specifically includes: The consistency parameter is expressed as ω = [ω1, ω2, ..., ω5], and the specific value of the consistency parameter is calculated so that the feature similarity matrix S D The matrix calculated by the consistency parameter is similar to the fault similarity matrix S E The Euclidean distance is the smallest, which is specifically expressed by the following formula: Among them, C represents the number of algorithms, which is 5; Minimize the above formula and obtain the value of each parameter in the consistency parameter ω = [ω1, ω2, ..., ω5].

4. The method for zero-sample fault diagnosis of industrial processes based on similarity consistency relationship according to claim 1, characterized in that: Step 9 is implemented by the following formula:

5. The method for zero-sample fault diagnosis of industrial processes based on similarity consistency relationship according to claim 1, characterized in that: In step 5, the target vector f of the i-th type of seen fault is calculated by Euclidean distance i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula: Among them, d1(f i ,f j ) is the target vector f of the seen class faults under the Euclidean distance algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, f i k represents the target vector of the i-th type of seen fault in the k-th dimension, f j k represents the target vector of the jth type of seen fault in the kth dimension; For d1(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

6. The method for zero-sample fault diagnosis of industrial processes based on similarity consistency relationship according to claim 1, characterized in that: In step 5, the target vector f of the i-th type of seen fault is calculated by cosine similarity. i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula: Among them, d2(f i ,f j ) is the target vector f of the seen faults under the cosine similarity algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, f i k represents the target vector of the i-th type of seen fault in the k-th dimension, f j k represents the target vector of the jth type of seen fault in the kth dimension; For d2(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

7. The method for zero-sample fault diagnosis of industrial processes based on similarity consistency relationship according to claim 1, characterized in that: In step 5, the target vector f of the i-th type of seen fault is calculated by using the Pearson correlation coefficient. i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula: Among them, d3(f i ,f j ) is the target vector f of the observed faults under the Pearson correlation coefficient algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, f i k represents the target vector of the i-th type of seen fault in the k-th dimension, f j k represents the target vector of the j-th type of seen fault in the k-th dimension, represents the average value of the target vector of the i-th eigenvector in n dimensions, Represents the average value of the target vector of the i-th feature vector in n dimensions; For d3(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

8. The method for zero-sample fault diagnosis of industrial processes based on similarity consistency relationship according to claim 1, characterized in that: In step 5, the target vector f of the i-th type of seen fault is calculated by using the Spearman correlation coefficient i and the target vector f of the jth seen fault j The feature similarity value between is realized by the following formula: Among them, d4(f i ,f j ) is the target vector f of the observed faults under the Spearman correlation coefficient algorithm i and the target vector f of the jth seen fault j The feature similarity value between them, n represents the dimension of the target vector, f i k represents the target vector of the i-th type of seen fault in the k-th dimension, f j k represents the target vector of the jth type of seen fault in the kth dimension, R(f i k ) is the target vector f i Medium i k the sort position in the target vector; For d4(f i ,f j ) to normalize and obtain the normalized feature similarity value This is achieved specifically through the following formula:

9. The method for zero-sample fault diagnosis of industrial processes based on similarity consistency relationship according to claim 1, characterized in that: In step 5, the target vector f of the i-th type of seen fault is calculated through mutual information i and the target vector f of the jth seen fault j The feature similarity value between them is expressed by the following formula: Among them, d4(f i ,f j ) is the target vector f of the seen faults under the mutual information algorithm i and the target vector f of the jth seen fault j The feature similarity value between represents the normalized feature similarity value. Under the mutual information algorithm, d4(f i ,f j )and equal, p(x,y) is f i and f j The joint probability of taking specific values ​​of x and y; p(x) and p(y) are f i and f j Their respective marginal probabilities.

Images