Zero sample equipment fault diagnosis method based on attribute migration

Abstract

The invention provides a zero sample equipment fault diagnosis method based on attribute migration, which comprises the following steps: dividing each attribute into S attribute subsets, and calculating S probability values obtained by attribute migration under a multi-Gaussian hypothesis and a first attribute predicted value under a single-Gaussian hypothesis according to the attribute subsets; fusing the S probability values and the first attribute prediction value through an attribute fusion network to obtain a final attribute prediction vector; and finally, combining a plurality of final attribute predicted values to obtain a predicted attribute vector, and selecting a fault category with the minimum Euclidean space distance from the predicted attribute vector in a known fault attribute space by utilizing a nearest neighbor search inference rule, namely, determining the fault category of the input sample. According to the method, the capability difference of predicting the same fault attribute by various attribute transfer methods is considered, the problem that the attribute transfer capability of fault diagnosis in a zero sample environment is still relatively weak is solved, and the zero sample diagnosis performance of industrial equipment is improved.

Description

Technical Field

[0001] The present invention relates to the field of zero-sample diagnosis of industrial equipment, and in particular to a zero-sample equipment fault diagnosis method based on attribute migration. Background Art

[0002] Industrial equipment is the cornerstone of modern industrial production. Its performance and reliability are directly related to production efficiency, product quality and the economic benefits of enterprises. Advanced industrial equipment can significantly improve the automation level of the production process and reduce labor costs. Through timely fault diagnosis, equipment problems can be quickly identified and repaired, downtime can be reduced, and the overall efficiency of the production line can be improved. Current research methods mainly focus on fault diagnosis techniques under supervised samples, which rely on a large amount of labeled data for model training. However, in practical applications, obtaining enough labeled data is often costly and time-consuming. Zero-sample fault diagnosis methods can diagnose potential faults by analyzing data during normal operation of the equipment without historical fault data. At present, zero-sample equipment fault diagnosis technology does not consider the differences in the migration capabilities of different fault attribute migrators for the same attribute, resulting in poor zero-sample fault diagnosis performance. Summary of the invention

[0003] The purpose of the present invention is to provide a zero-sample equipment fault diagnosis method based on attribute migration, which takes into account the ability differences of various attribute transfer methods to predict the same fault attributes, solves the problem that the attribute migration ability of fault diagnosis in a zero-sample environment is still weak, and improves the zero-sample diagnosis performance of industrial equipment.

[0004] To achieve the above object, the present invention provides a zero-sample device fault diagnosis method based on attribute migration, comprising the following steps:

[0005] First, each attribute is divided into S attribute subsets, and the S probability values ​​obtained by attribute migration under the multi-Gaussian hypothesis and the first attribute prediction value under the single Gaussian hypothesis are calculated based on the attribute subsets;

[0006] The attribute fusion network fuses S probability values ​​and the first attribute prediction value to obtain the final attribute prediction vector:

[0007] When S is greater than 1, the proposed attribute fusion network is used to fuse the S probability values ​​obtained by attribute migration under the multi-Gaussian hypothesis and the first attribute prediction value obtained under the single Gaussian hypothesis through multiple fully connected layers and connections to calculate the final attribute prediction value;

[0008] When S is not greater than 1, the first attribute prediction value based on the single Gaussian hypothesis is directly used as the final attribute prediction value;

[0009] Finally, multiple final attribute prediction values ​​are combined to obtain the predicted attribute vector and the known fault attribute space By using the nearest neighbor search inference rule, the fault category with the smallest distance to the predicted attribute vector in the Euclidean space in the fault attribute space is selected, that is, the fault category of the input sample is determined.

[0010] Preferably, the first attribute prediction value obtained by calculating the attribute transfer equation under the single Gaussian assumption includes:

[0011] First, Divide into S subsets, is the attribute space of the known fault category with the jth attribute value of 1, A represents the known fault category, Indicates T A Middle C j The part with attribute value equal to 1, T A represents the attribute space of known fault classes, j represents the index value of the jth fault attribute, M represents the fault attribute number, N represents the number of fault categories with the jth attribute value of 1 in the known fault category. A Indicates the number of known fault classes;

[0012] The expression of the predicted value of the first attribute is:

[0013]

[0014] Among them, V j (x) represents the jth attribute prediction value of sample x, and P(·) refers to the probability of occurrence; in the attribute space of known fault classes In the example, the jth attribute is represented by {C j |j∈(0,1,...,M-1), Refers to T A Middle C j The part with attribute value equal to 0,

[0015] Preferably, S probability values ​​obtained by attribute migration under the multi-Gaussian assumption are calculated, including

[0016] In addition to the attribute C j In addition, sample x has M-1 other attributes, It is divided into S sets by other M-1 attributes;

[0017]

[0018] in, is the Sth attribute set, n S-1is the element number of the Sth attribute set, n0+n1+…+n S-1 =N A ;

[0019] S probability values ​​are h0(x),h1(x),…h S-1 (x), where

[0020]

[0021] Among them, h S-1 (x) represents the Sth probability value.

[0022] Preferably, the input of the attribute fusion network is x p , x p Represents the sample x after principal component analysis, and the proposed attribute fusion network is:

[0023] c1=Relu(W1x p +b1);

[0024] c2=Relu(W2c1+b2);

[0025] V = Relu(W3c2+b3);

[0026] P=Concat(Output0,(h0(x p ),h1(x p ),...,h S-1 (x p )));

[0027] Output1 = V·P;

[0028] Among them, c1 represents the output of the first fully connected layer, c2 represents the output of the second fully connected layer, and V represents the weight vector of attribute fusion. P represents the probability vector, Relu represents a nonlinear correction unit, Concat represents a connection operation, and x p represents the input of the attribute fusion network, W1, W2 and W3 represent the weights of the first fully connected layer, the second fully connected layer and the third fully connected layer respectively, b1, b2 and b3 represent the biases of the first fully connected layer, the second fully connected layer and the third fully connected layer respectively, Output1 represents the final attribute prediction value, Output0 represents the first attribute prediction value V j (x).

[0029] Therefore, the present invention adopts the above-mentioned zero-sample device fault diagnosis method based on attribute migration, and the technical effects are as follows:

[0030] (1) Considering the differences in the ability of various attribute transfer methods to predict the same fault attributes, the problem that the attribute transfer ability of fault diagnosis in zero-sample environment is still weak is solved, and the zero-sample diagnosis performance of industrial equipment is improved.

[0031] (2) The attribute fusion network is used to fuse the probability value obtained by attribute migration under the multi-Gaussian hypothesis and the attribute prediction value obtained under the single Gaussian hypothesis, so as to obtain a more accurate fault attribute prediction value, thereby obtaining considerable zero-sample fault diagnosis capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a zero-sample device fault diagnosis method based on attribute migration according to the present invention;

[0033] Figure 2 This is a performance test with zero samples. DETAILED DESCRIPTION

[0034] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0035] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0036] Embodiment 1

[0037] like Figure 1 As shown, a zero-sample device fault diagnosis method based on attribute migration includes the following steps:

[0038] First, each attribute is divided into S attribute subsets, and the S probability values ​​obtained by attribute migration under the multi-Gaussian hypothesis and the first attribute prediction value obtained under the single Gaussian hypothesis are calculated based on the attribute subsets;

[0039] The first attribute prediction value obtained under the single Gaussian assumption includes

[0040] During the preprocessing process, is the attribute space of the known fault category with the jth attribute value of 1, is the number of fault categories with the jth attribute value of 1 in the known fault category, x p represents the sample x after principal component analysis (PCA). In the preprocessing work, Divide into S subsets.

[0041] In the attribute space of known fault classes In the example, the jth attribute is represented as Refers to T A Middle Cj The part where the attribute value is equal to 0. Refers to T A Middle C j The part with attribute value equal to 1. M refers to the fault attribute number. j refers to the index value of the jth fault attribute. The jth attribute prediction value of sample x is recorded as V j (x), where P(·) refers to the probability of occurrence.

[0042] The expression of the predicted value of the first attribute is:

[0043]

[0044] Calculate the S probability values ​​h0(x),h1(x),...h obtained by attribute migration under the Gaussian hypothesis S-1 (x) including:

[0045] In addition to the attribute C j In addition, sample x has M-1 other attributes. Assume Divided into S sets by other M-1 attributes.

[0046]

[0047] in, is the Sth attribute set, n S-1 is the element number of the Sth attribute set, n0+n1+…+n S-1 =N A .

[0048] in this case, in

[0049]

[0050] in is the predicted value of the jth attribute of sample x in the hypothesis.

[0051] Step 2: Fuse the S probability values ​​and the first attribute prediction value through the attribute fusion network to obtain the final attribute prediction vector:

[0052] In the attribute migration process, the attribute prediction values ​​obtained by attribute migration under the multi-Gaussian hypothesis and the attribute prediction values ​​obtained by the Bayesian fault attribute transfer equation have different accuracy distributions on fault attributes. Different assumptions are likely to have advantages in the prediction of different fault attributes. Therefore, the attribute prediction results that combine the two assumptions are likely to obtain more accurate attribute predictions.

[0053] When S is greater than 1, the proposed attribute fusion network is used to fuse the S probability values ​​obtained by attribute migration under the multi-Gaussian hypothesis and the first attribute prediction value obtained under the single Gaussian hypothesis to calculate the final attribute prediction value;

[0054] When S is not greater than 1, the first attribute prediction value based on the single Gaussian hypothesis is directly used as the final attribute prediction value;

[0055] Step 3: Finally, combine multiple final attribute prediction values ​​to obtain the predicted attribute vector and the known fault attribute space By using the nearest neighbor search inference rule, the fault category with the smallest distance to the predicted attribute vector in the Euclidean space in the fault attribute space is selected, that is, the fault category of the input sample is determined.

[0056] The proposed attribute fusion network is:

[0057] c1=Relu(W1x p +b1);

[0058] c2=Relu(W2c1+b2);

[0059] V = Relu(W3c2+b3);

[0060] P=Concat(Output0,(h0(x p ),h1(x p ),...,h S-1 (x p )));

[0061] Output1 = V·P;

[0062] Among them, c1 represents the output of the first fully connected layer, c2 represents the output of the second fully connected layer, and V represents the weight vector of attribute fusion. P represents the probability vector, Relu represents a nonlinear correction unit, Concat represents a connection operation, and x p represents the input of the attribute fusion network, W1, W2 and W3 represent the weights of the first fully connected layer, the second fully connected layer and the third fully connected layer respectively, b1, b2 and b3 represent the biases of the first fully connected layer, the second fully connected layer and the third fully connected layer respectively, Output1 represents the final attribute prediction value, Output0 represents the first attribute prediction value V j (x).

[0063] Based on the spyder tool, the Tennessee dataset is used and divided into a training set and a test set.

[0064] like Figure 2As shown, the zero-sample performance test graph provided by this embodiment. This method is compared with the method with training samples (the number of training samples is 30-600). A total of ten experiments were conducted and the average value was taken. The figure shows that the proposed method is effective for zero-sample diagnosis of industrial equipment and has a high diagnostic accuracy.

[0065] In the preprocessing work, first Convert to S subsets. is the attribute space of the known fault category with the jth attribute value of 1. It is worth noting that different known fault categories and will be combined into different The experiment used 25 combinations of known fault categories, each with 18 This means that the experiment has 450 different For different The experiment uses the following simple rules to divide the subsets.

[0066]

[0067] in is the attribute space of known fault categories The element of , the attribute value is 1. i represents the i-th fault category. j represents the j-th attribute.

[0068] The fault descriptions of Fault #0 (A / C feed ratio, B component constant, step size change) and Fault #1 (B component, A / C ratio constant, step size change) are similar and very different from other fault descriptions. In addition, the fault attribute vectors of Fault #0 and Fault #1 are highly similar, and have low similarity to other faults. From the fault descriptions and attribute vectors, it can be seen that Fault #0 and Fault #1 are similar and different from other faults. Therefore, the faults belonging to Fault #0 and Fault #1 are In the experiment, all All follow this simple rule.

[0069] Therefore, the present invention adopts the above-mentioned zero-sample equipment fault diagnosis method based on attribute migration, takes into account the ability differences of various attribute transfer methods to predict the same fault attributes, solves the problem that the attribute migration ability of fault diagnosis in a zero-sample environment is still weak, and improves the zero-sample diagnosis performance of industrial equipment.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A zero-sample device fault diagnosis method based on attribute migration, characterized in that: The following steps are involved: First, each attribute is divided into S attribute subsets, and the S probability values ​​obtained by attribute migration under the multi-Gaussian hypothesis and the first attribute prediction value under the single Gaussian hypothesis are calculated based on the attribute subsets; The attribute fusion network fuses S probability values ​​and the first attribute prediction value to obtain the final attribute prediction vector: When S is greater than 1, the proposed attribute fusion network is used to fuse the S probability values ​​obtained by attribute migration under the multi-Gaussian hypothesis and the first attribute prediction value obtained under the single Gaussian hypothesis through multiple fully connected layers and connections to calculate the final attribute prediction value; When S is not greater than 1, the first attribute prediction value based on the single Gaussian hypothesis is directly used as the final attribute prediction value; Finally, multiple final attribute prediction values ​​are combined to obtain the predicted attribute vector and the known fault attribute space By using the nearest neighbor search inference rule, the fault category with the smallest distance to the predicted attribute vector in the Euclidean space in the fault attribute space is selected, that is, the fault category of the input sample is determined.

2. The zero-sample device fault diagnosis method based on attribute migration according to claim 1, characterized in that: Calculate the first attribute prediction value obtained by the attribute transfer equation under the single Gaussian assumption, including First, Divide into S subsets, is the attribute space of the known fault category with the jth attribute value of 1, A represents the known fault category, Indicates T A Middle C j The part with attribute value equal to 1, T A represents the attribute space of known fault classes, j represents the index value of the jth fault attribute, M represents the fault attribute number, N represents the number of fault categories with the jth attribute value of 1 in the known fault category. A Indicates the number of known fault classes; The expression of the predicted value of the first attribute is: Among them, V j (x) represents the jth attribute prediction value of sample x, and P(·) refers to the probability of occurrence; in the attribute space of known fault classes In the example, the jth attribute is represented by {C j |j∈(0,1,...,M-1), Refers to T A Middle C j The part with attribute value equal to 0, 3. The zero-sample device fault diagnosis method based on attribute migration according to claim 1, characterized in that: Calculate the S probability values ​​obtained by attribute migration under the Gaussian hypothesis, including In addition to the attribute C j In addition, sample x has M-1 other attributes, It is divided into S sets by other M-1 attributes; in, is the Sth attribute set, n S-1 is the element number of the Sth attribute set, n0+n1+...+n S-1 =N A ; S probability values ​​are h0(x),h1(x),...h S-1 (x), where Among them, h S-1 (x) represents the Sth probability value.

4. The zero-sample device fault diagnosis method based on attribute migration according to claim 1, characterized in that: The input of the attribute fusion network is x p , x p Represents the sample x after principal component analysis, and the proposed attribute fusion network is: c1=Relu(W1x p +b1); c2=Relu(W2c1+b2); V = Relu(W3c2+b3); P=Concat(Output0,(h0(x p ),h1(x p ),...,h S-1 (x p ))); Output1 = V·P; Among them, c1 represents the output of the first fully connected layer, c2 represents the output of the second fully connected layer, and V represents the weight vector of attribute fusion. P represents the probability vector, Relu represents a nonlinear correction unit, Concat represents a connection operation, and x p represents the input of the attribute fusion network, W1, W2 and W3 represent the weights of the first fully connected layer, the second fully connected layer and the third fully connected layer respectively, b1, b2 and b3 represent the biases of the first fully connected layer, the second fully connected layer and the third fully connected layer respectively, Output1 represents the final attribute prediction value, Output0 represents the first attribute prediction value V j (x).

Images