A multi-granularity fuzzy zero-shot fault diagnosis method with progressive knowledge transfer

Abstract

The present invention discloses a multi-granularity fuzzy zero-sample fault diagnosis method with progressive knowledge transfer. First, the present invention divides the attribute vector describing the fault into multi-granularity attribute layers according to the information granularity from coarse to fine, and forms multiple attribute sub-vectors. Unlike the traditional method of directly predicting the attribute vector, the present invention designs a progressive knowledge transfer attribute prediction network to predict the attribute sub-vector layer by layer from coarse to fine granularity. At the same time, the knowledge learned in the prediction of the upper-layer attribute sub-vector is transferred to the prediction of the lower-layer attribute sub-vector to improve the attribute prediction accuracy. Based on the obtained predicted attribute sub-vector, the present invention designs a fuzzy inference strategy to obtain a combination of weighted coefficients, and realizes the probability result of inferring the fault from a single attribute layer. The attribute vector of the unseen fault category is obtained, and the fault probability results inferred from multiple attribute layers are fused through the designed multi-granularity attribute layer fusion strategy to finally provide a zero-sample fault diagnosis result. Through the technical solution of the present invention, the multi-granularity characteristics of attributes can be exploited, and the accuracy of attribute prediction and fault inference in zero-sample fault diagnosis of industrial processes can be improved.

Description

Technical Field

[0001] The present invention discloses a multi-granularity fuzzy zero-sample fault diagnosis method with progressive knowledge transfer. The present invention belongs to the field of industrial fault diagnosis, and in particular is aimed at zero-sample fault diagnosis of industrial process data with multi-granularity attributes. Background Art

[0002] Fault diagnosis plays a vital role in the safe maintenance of industrial processes. With the development of the Industrial Internet of Things (IIoT) and artificial intelligence (AI) technologies, data-driven fault diagnosis methods have garnered widespread attention from both industry and academia. Generally speaking, these methods approach fault diagnosis as a supervised classification problem. First, historical data on various faults must be collected using IIoT technologies. Then, during the training phase, a classification model is established, such as support vector machines (SVMs), Fisher discriminant analysis (FDA), and convolutional neural networks (CNNs). During the testing phase, the fault data is fed into the model to determine the fault category. While machine learning or deep learning-based fault diagnosis methods have achieved promising results, collecting sufficient samples for a wide range of faults is challenging due to the cost of data acquisition. Consequently, these supervised classification methods face significant limitations in practical applications.

[0003] In real industrial processes, there is an extreme case where the fault types in the test set do not appear in the training set. In this case, the above method will be ineffective because no new fault samples are provided. Zero-shot fault diagnosis (ZSD) is a feasible solution to this problem. Combining the attribute descriptions of seen faults and unseen faults, zero-shot fault diagnosis can be divided into two steps: first, an attribute prediction strategy is adopted to identify whether the fault has these attributes. Then, a fault reasoning strategy is used to find the fault category that best matches the attribute description. Since attributes can be converted between different categories, ZSD can transfer knowledge from seen categories to unseen categories, thereby enabling diagnosis of new fault categories.

[0004] However, attributes typically describe faults at different granularities, exhibiting hierarchical characteristics. Previous methods typically treat attributes of all granularities equally. Generally speaking, coarse-grained attributes reflect common fault information, while fine-grained attributes focus on specific and detailed fault information. This difference in attribute granularity significantly impacts attribute prediction and fault inference strategies in zero-shot fault diagnosis. For attribute prediction, fine-grained attributes are often more difficult to predict than coarse-grained attributes. In real industrial processes, a large number of collected variables are associated with coarse-grained attributes, making them easy to identify. However, some variables associated with fine-grained attributes are sparse, making fine-grained attribute identification a significant challenge. For fault inference, attributes of different granularities have different impacts on fault inference, leading to the following two issues. First, coarse-grained attributes influence the fault inference results of fine-grained attributes. Typically, coarse-grained attributes are used to first infer the approximate range of fault categories. Based on this range, fine-grained attributes are then used to gradually filter candidate faults. Traditional ZSD methods treat all attributes equally, which can lead to non-critical, detail-oriented attributes mistakenly dominating fault inference. Second, most ZSD methods can only provide simple binary results and cannot explore the different effects of multi-granularity attributes on fault inference. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the existing technology and propose a multi-granularity fuzzy zero-shot fault diagnosis method with progressive knowledge transfer. The invention can achieve zero-shot fault diagnosis by determining the membership of candidate faults through fuzzy reasoning while improving the accuracy of attribute prediction.

[0006] The object of the present invention is achieved through the following technical solutions:

[0007] A multi-granularity fuzzy zero-shot fault diagnosis method with progressive knowledge transfer includes the following steps:

[0008] Step 1: Collect the fault categories of various industrial processes, summarize the attribute vectors for each industrial process fault category based on the semantic description information of the fault category, and form an attribute vector set;

[0009] Step 2: Collect fault category samples and category labels for various industrial processes. Find the corresponding attribute vector in the attribute vector set in step 1 based on the category label. Divide the attribute vector into multiple attribute layers based on the information granularity, from coarse to fine. Starting from the first attribute layer, sequentially concatenate the attribute layers to form multiple attribute sub-vectors as training data.

[0010] Step 3: Construct a progressive knowledge transfer attribute prediction network, use the training data to train the progressive knowledge transfer attribute prediction network, predict the multiple attribute sub-vectors obtained in step 2, and obtain multiple predicted attribute sub-vectors;

[0011] Step 4: Based on the multiple prediction attribute sub-vectors from step 3, a fuzzy inference strategy is designed to infer the probability of the fault from each attribute layer in the multiple prediction attribute sub-vectors. A weighted coefficient combination of the probability results is calculated for each attribute layer based on the training data. Multiple weighted coefficient combinations are obtained for multiple attribute layers.

[0012] Step 5: Collect samples of unseen fault categories from various industrial processes. Unseen fault category samples refer to samples of fault categories that do not appear in the training data. Input the unseen fault category samples into the progressive knowledge transfer attribute prediction network to obtain multiple predicted attribute sub-vectors. Based on the multiple weighted coefficient combinations in step 4, multiple single-attribute layer comprehensive diagnosis results are obtained. Design a multi-granularity attribute layer fusion strategy to fuse the comprehensive diagnosis results of multiple single-attribute layers, calculate the probability of each unseen fault category, and diagnose the unseen fault category samples.

[0013] In step 2, we collect fault category samples and category labels for various industrial processes, including:

[0014] Collect samples of fault categories from various industrial processes and category labels Where N and J represent the number of fault category samples and the number of variables of fault category samples, respectively.

[0015] 3. The method according to claim 1, characterized in that in step 2, the corresponding attribute vector is found in the attribute vector set of step 1 according to the category label, and the attribute vector is divided into multiple attribute layers according to the information granularity from coarse to fine. Starting from the first attribute layer, the attribute layers are sequentially spliced ​​to form multiple attribute sub-vectors, specifically including:

[0016] 2.1) The attribute vector set is m represents the number of fault categories, and C represents the dimension of each attribute vector;

[0017] 2.2) According to the information granularity from coarse to fine, each attribute vector in the attribute vector set is divided into L attribute layers, Where a is the attribute vector in the attribute vector set, Indicates the splicing operation, a l Represents the l-th attribute layer of the attribute vector, where 1≤l≤L, and the smaller the l number, the coarser the granularity;

[0018] 2.3) Take the first attribute layer as the first attribute sub-vector, concatenate the first two attribute layers as the second attribute sub-vector, and so on, until the first L attribute layers are concatenated as the Lth attribute sub-vector, forming a total of L attribute sub-vectors.

[0019] In step 3, a progressive knowledge transfer attribute prediction network is constructed. The progressive knowledge transfer attribute prediction network is trained using training data. The multiple attribute sub-vectors obtained in step 2 are predicted to obtain multiple predicted attribute sub-vectors, specifically including:

[0020] A progressive knowledge transfer attribute prediction network is constructed to predict L attribute sub-vectors layer by layer. When predicting the next attribute sub-vector, the prior knowledge extracted from the prediction of the previous attribute sub-vector is transferred until all L attribute sub-vectors of the training data obtained in step 2 are predicted. Multiple predicted attribute sub-vectors are obtained, and the progressive knowledge transfer attribute prediction network is trained based on the mean square error between the predicted attribute sub-vectors of the training data and the attribute sub-vectors.

[0021] Construct a progressive knowledge transfer attribute prediction network, predict L attribute sub-vectors layer by layer, and transfer the prior knowledge extracted from the previous attribute sub-vector prediction when predicting the next attribute sub-vector. Specifically,

[0022] 3.1) For the first attribute sub-vector, build an attribute feature extractor and an attribute sub-vector predictor consisting of a multi-layer fully connected neural network. Use the attribute feature extractor to extract the attribute sub-vector features of the fault category samples in the training data. Input the extracted attribute sub-vector features into the attribute sub-vector predictor to predict the attribute sub-vector and obtain the first predicted attribute sub-vector.

[0023] 3.2) For the lth attribute sub-vector (2≤l≤L), build an attribute feature extractor and attribute sub-vector predictor composed of a multi-layer fully connected neural network. Use the parameters of the attribute feature extractor and attribute sub-vector predictor built for the l-1th attribute sub-vector to initialize the parameters of the constructed attribute feature extractor and attribute sub-vector predictor, respectively, to achieve parameter knowledge transfer; concatenate the attribute sub-vector features extracted when predicting the previous attribute sub-vector with the fault category samples in the training data, input them into the attribute feature extractor to extract attribute sub-vector features, and input the extracted attribute sub-vector features into the attribute sub-vector predictor to predict the attribute sub-vector, thereby obtaining the lth predicted attribute sub-vector;

[0024] 3.3) Repeat step 3.2) until all L attribute sub-vectors are predicted, and a total of L predicted attribute sub-vectors are obtained.

[0025] In step 4, based on the multiple prediction attribute sub-vectors in step 3, a fuzzy inference strategy is designed to infer the probability of failure from each attribute layer in the multiple prediction attribute sub-vectors, specifically including:

[0026] 5.1) For the i-th attribute layer (1≤i≤L), there are L-i+1 predicted attribute sub-vectors, each of which contains the i-th attribute layer, denoted as in represents the i-th attribute layer of the l-th predicted attribute subvector, where i≤l≤L;

[0027] 5.2) Calculate the i-th attribute layer of the l-th predicted attribute sub-vector The similarity with the i-th attribute layer of all attribute vectors in the attribute vector set A is normalized to obtain the probability result of inferring the fault from the i-th attribute layer of the l-th predicted attribute sub-vector, which is recorded as For L-i+1 predicted attribute sub-vectors, we obtain L-i+1 probability results, which are recorded as

[0028] In step 4, each attribute layer calculates a weighted coefficient combination of the probability result based on the training data. Multiple attribute layers obtain multiple weighted coefficient combinations, specifically including:

[0029] For the i-th attribute layer, each of the L-i+1 predicted attribute sub-vectors contains the i-th attribute layer, and a total of L-i+1 probability results are obtained. Set the optimization goal and search for a weighted coefficient combination w i , where the weighted coefficient combination w i Each element in is not less than 0, and the sum is 1. The weighted combination of the L-i+1 probability results is required to minimize the difference between the combined probability results and the optimization target; the weighted coefficient combination w i Solve through linear programming; for L attribute layers, obtain L weighted coefficient combinations, denoted as {w1,w2,...,w L}.

[0030] Set optimization goals, including:

[0031] Let the optimization target be p i , is a L-i+1 dimensional vector, and the setting rules for each element in the vector are as follows:

[0032] When the i-th attribute layer of the j-th attribute vector in the attribute vector set is the same as the i-th attribute layer of the attribute vector corresponding to the label category of the fault category sample, p i The j-th element of is denoted as 1;

[0033] When the i-th attribute layer of the j-th attribute vector in the attribute vector set is different from the i-th attribute layer of the attribute vector corresponding to the label category of the fault category sample, but is the same as the i-1-th attribute layer, p i The j-th element of is recorded as 0.5;

[0034] If it does not belong to the above two situations, p i The j-th element of is denoted as 0.

[0035] In step 5, samples of unseen fault categories from various industrial processes are collected and fed into the progressive knowledge transfer attribute prediction network to obtain multiple predicted attribute subvectors. Based on the multiple weighted coefficients from step 4, multiple single-attribute layer comprehensive diagnosis results are obtained. A multi-granularity attribute layer fusion strategy is designed to fuse the comprehensive diagnosis results of multiple single-attribute layers, calculate the probability of each unseen fault category, and diagnose the unseen fault category samples. Specifically, the following steps are performed:

[0036] 5.1) Collect samples of unseen fault categories of various industrial processes, summarize attribute vectors for unseen fault categories of each industrial process based on the semantic description information of the unseen fault categories, and form an unseen attribute vector set Where m represents the number of unseen fault categories;

[0037] 5.2) Input the unseen fault category samples into the progressive knowledge transfer attribute prediction network to obtain L predicted attribute sub-vectors;

[0038] 5.3) Calculate the relationship between each attribute layer of the L predicted attribute sub-vectors and the unseen attribute vector set A U The similarity of the attribute layers corresponding to all attribute vectors is normalized to obtain multiple fault probability results inferred from a single attribute layer;

[0039] 5.4) Based on the multiple weighted coefficient combinations obtained in step 4, perform weighted combinations on the fault probability results inferred from different attribute layers to obtain a comprehensive diagnosis result for a single attribute layer; for L attribute layers, obtain L comprehensive diagnosis results for each attribute layer;

[0040] 5.5) Design a multi-granularity attribute layer fusion strategy, fuse the L single attribute layer comprehensive diagnosis results obtained in step (5.4), calculate the probability of each unseen fault category, and obtain the diagnosis results of the unseen fault category samples.

[0041] In step 5.5), a multi-granularity attribute layer fusion strategy is designed to fuse the L single attribute layer comprehensive diagnosis results obtained in step 5.4), calculate the probability of each unseen fault category, obtain the diagnosis results of the unseen fault category samples, and diagnose the unseen fault category samples, specifically including:

[0042] For each unseen fault category, the diagnostic results that are the same as the unseen fault category in each single-attribute layer comprehensive diagnostic result are added together. For the L single-attribute layer comprehensive diagnostic results obtained, the added diagnostic results are multiplied and normalized to obtain the probability of each unseen fault category. The unseen fault category with the highest probability is taken as the diagnostic result of the unseen fault category sample.

[0043] Based on expert knowledge, the attributes of industrial process faults are generally divided into five attribute layers from coarse to fine granularity: "system attributes", "equipment attributes", "component attributes", "variable attributes", and "disturbance attributes". Attribute layers can be added or deleted based on actual industrial objects.

[0044] The beneficial effects of the present invention are:

[0045] The method proposed in the present invention is mainly aimed at industrial process data, and a multi-granularity fuzzy zero-sample fault diagnosis method with progressive knowledge transfer is proposed. In response to the problem of difficulty in identifying fine-grained attributes, the present invention designs a progressive knowledge transfer attribute prediction module, which transfers the dual knowledge based on models and features from coarse-grained attributes that are easy to identify to fine-grained attributes that are difficult to identify, realizes the sequential prediction of multi-granularity attributes, and improves the accuracy of attribute prediction. In response to the problem that it is difficult to quantitatively describe the impact of multi-granularity attributes on fault inference, the present invention proposes a multi-granularity attribute fuzzy inference module to quantitatively describe the impact of attributes of different granularities on fault inference. Unlike traditional zero-sample fault diagnosis methods that simply provide binary diagnosis results, this strategy gradually screens candidate faults through fuzzy reasoning and multi-granularity attribute fusion, improving the accuracy of fault diagnosis while having better interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The invention is further described below with reference to the accompanying drawings and embodiments:

[0047] Figure 1 An accompanying diagram for the abstract;

[0048] Figure 2 is the attribute description of various faults, where 1 indicates that the fault meets the attribute description, and 0 indicates that the fault does not meet the attribute description;

[0049] Figure 3 Framework diagram of multi-granularity fuzzy zero-shot fault diagnosis method for progressive knowledge transfer;

[0050] Figure 4 Network framework diagram for progressive knowledge transfer attribute prediction; DETAILED DESCRIPTION

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] This embodiment uses thermal power plant data as an example to verify the effectiveness of this method, but is not limited to this. Any industrial project with multi-granularity characteristics of attributes can use the method of the present invention for zero-sample fault diagnosis.

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] The thermal power generation process usually burns fuels such as oil, coal and natural gas to generate heat and turn water into steam. The high-temperature and high-pressure steam then drives the turbine blades to drive the generator. The thermal power generation process mainly consists of two subsystems, including the boiler system and the turbine system. Different systems contain different equipment, components and variables, and different disturbances of the variables also correspond to different fault types. In this example, a total of 68 process variables are included, including condenser circulating water temperature, condenser side A circulating water inlet temperature, condenser external drain expansion tank header temperature, circulating water pump motor current, circulating water pump outlet main pipe pressure, condensate flow, high-pressure steam inlet pressure, low-pressure steam inlet pressure, high-pressure steam inlet temperature, low-pressure steam inlet temperature, etc. This example collects 15 types of faults, each with 4,000 samples. These faults can be described by 17 attributes. These attributes are divided into five groups, namely "system attributes", "equipment attributes", "component attributes", "variable attributes", and "disturbance attributes". The specific attribute meanings and attribute descriptions of each fault are shown in Tables 1 and 2. Figure 2 shown.

[0055] Table 1 Specific meaning of fault attributes

[0056]

[0057]

[0058] Of the 15 fault types in this example, 12 were selected as known faults, and the remaining 3 were designated as unknown faults. This data was not used during training, and the unknown faults were presented as a test set to verify the effectiveness of zero-shot fault diagnosis. Nine different divisions of seen and unseen faults were designed for verification, as shown in Table 2.

[0059] Table 2 Classification of seen and unseen faults

[0060]

[0061] A multi-granularity fuzzy zero-shot fault diagnosis method with progressive knowledge transfer, the method framework is shown in the figure Figure 3 As shown, the specific modeling steps are as follows:

[0062] A multi-granularity fuzzy zero-shot fault diagnosis method with progressive knowledge transfer includes the following steps:

[0063] Step 1: Collect fault categories for various industrial processes. Based on the semantic description of the fault category, summarize attribute vectors for each industrial process fault to form an attribute vector set. Divide the attribute vectors in the attribute vector set into multiple attribute layers based on the information granularity from coarse to fine. Starting from the first attribute layer, the attribute layers are sequentially spliced ​​to form multiple attribute sub-vectors.

[0064] Step 2: Collect fault category samples and category labels for various industrial processes, and convert the attribute vector corresponding to each category label into multiple attribute sub-vectors according to step 1 as training data;

[0065] Step 3: Construct a progressive knowledge transfer attribute prediction network, use the training data to train the progressive knowledge transfer attribute prediction network, predict the multiple attribute sub-vectors obtained in step 2, and obtain multiple predicted attribute sub-vectors;

[0066] Step 4: Based on the multiple prediction attribute sub-vectors from step 3, a fuzzy inference strategy is designed to preliminarily infer multiple probability results of the fault from the perspective of each attribute layer in the multiple prediction attribute sub-vectors. Based on the training data, weighted coefficient combinations of the multiple probability results are solved. Multiple weighted coefficient combinations can be obtained for multiple attribute layers.

[0067] Step 5: Collect unseen fault categories from various industrial processes, input the unseen fault samples into the trained progressive knowledge transfer attribute prediction network, obtain multiple prediction attribute sub-vectors, and design a multi-granularity attribute layer fusion strategy based on the fuzzy inference strategy of step 4 and multiple weighted coefficient combinations. Diagnose the unseen fault categories based on the obtained multiple prediction attribute sub-vectors.

[0068] In step 1, the attribute vectors in the attribute vector set are divided into multiple attribute layers according to the information granularity from coarse to fine. Starting from the first attribute layer, the attribute layers are sequentially spliced ​​to form multiple attribute sub-vectors, including:

[0069] 2.1) The attribute vector set is in, represents the attribute vector of the i-th fault category, m represents the number of fault categories, and C represents the dimension of each attribute vector; a i,j (1≤j≤C) represents a i The jth attribute of ;

[0070] 2.2) According to the information granularity from coarse to fine, each attribute vector in the attribute vector set is divided into L attribute layers, in Indicates the splicing operation, a i represents the i-th attribute vector in the attribute vector set, It represents the lth attribute layer of the i-th attribute vector in the attribute vector set. The smaller the l number is, the coarser the granularity is. l Indicates the number of attributes in the lth attribute layer;

[0071] 2.3) For each attribute vector, starting from the first attribute layer, sequentially concatenate attribute layers until the Lth attribute layer is reached, forming L attribute sub-vectors.

[0072] In step 2, fault category samples and category labels of various industrial processes are collected, and the attribute vector corresponding to each category label is converted into multiple attribute sub-vectors as training data according to step 1, including:

[0073] Collect samples of fault categories from various industrial processes and category labels Where N and J represent the number of samples and the number of variables respectively, k (1≤k≤N) represents the kth sample, y k Represents the category label of the kth sample; obtain the corresponding L attribute sub-vectors according to the category label, record in Represents the lth attribute subvector corresponding to the category label of the kth sample;

[0074] In step 3, a progressive knowledge transfer attribute prediction network is constructed and the training data is used to train the progressive knowledge transfer attribute prediction network. The framework diagram is as follows: Figure 4 As shown, the multiple attribute sub-vectors obtained in step 2 are predicted to obtain multiple predicted attribute sub-vectors, specifically including:

[0075] A progressive knowledge transfer attribute prediction network is constructed, and the attribute sub-vectors are predicted layer by layer in the order of coarse to fine information granularity. When predicting the attribute sub-vector of the next layer, the prior knowledge extracted from the prediction of the attribute sub-vector of the upper layer is transferred until all the attribute sub-vectors of the training data obtained in step 2 are predicted. Multiple predicted attribute sub-vectors are obtained, and the progressive knowledge transfer attribute prediction network is trained based on the predicted attribute sub-vectors of the training data and the mean square error of the attribute sub-vectors.

[0076] In step 3, a progressive knowledge transfer attribute prediction network is constructed. The attribute sub-vectors are predicted layer by layer in the order of coarse to fine information granularity. When predicting the attribute sub-vector of the next layer, the prior knowledge extracted from the prediction of the attribute sub-vector of the upper layer is transferred until all the attribute sub-vectors of the training data obtained in step 2 are predicted. Multiple predicted attribute sub-vectors are obtained, and the progressive knowledge transfer attribute prediction network is trained based on the mean square error between the predicted attribute sub-vectors of the training data and the attribute sub-vectors. Specifically, the following steps are performed:

[0077] 5.1) For the kth sample x in the training data k , build the first attribute feature extractor and attribute sub-vector predictor composed of a multi-layer fully connected neural network, and initialize its network parameters randomly; k Input the first attribute feature extractor built to extract the first attribute sub-vector feature Record

[0078]

[0079] in, Represents the function corresponding to the first attribute feature extractor;

[0080] The first attribute sub-vector feature extracted Input the first attribute sub-vector predictor built and output the first predicted attribute sub-vector Record

[0081]

[0082] in, Represents the function corresponding to the first attribute sub-vector predictor;

[0083] The neural network parameters of the first attribute feature extractor and the first attribute subvector predictor are trained by minimizing the mean square error between the first predicted attribute subvector and the first attribute subvector corresponding to the category label of all training data, denoted as

[0084]

[0085] in, represents the two-norm, θ1 and φ1 represent the neural network parameters of the first attribute feature extractor and the first attribute sub-vector predictor, respectively;

[0086] 5.2) For the kth sample x in the training data k , construct the attribute feature extractor and attribute sub-vector predictor composed of the l-th multi-layer fully connected neural network to predict the l-th attribute sub-vector, where 2≤l≤L; use the network parameters θ in the attribute feature extractor and attribute sub-vector predictor trained in the previous step l-1 and φ l-1 The network parameters of the constructed l-th attribute feature extractor and attribute sub-vector predictor are aligned and migrated for initialization respectively. The network parameters of the unaligned l-th attribute feature extractor and attribute sub-vector predictor are initialized using the Kaiming initialization method, thereby realizing progressive parameter knowledge migration; the attribute sub-vector features extracted in the previous step are initialized using the Kaiming initialization method. With sample x k After splicing, the lth attribute feature extractor is input to realize progressive feature knowledge transfer and the lth attribute sub-vector feature is output. Record

[0087]

[0088] in, Represents the function corresponding to the lth attribute feature extractor;

[0089] The extracted l-th attribute sub-vector feature Input the built l-th attribute sub-vector predictor and output the l-th predicted attribute sub-vector Record

[0090]

[0091] in, represents the function corresponding to the l-th attribute sub-vector predictor;

[0092] The neural network parameters of the lth attribute feature extractor and the lth attribute subvector predictor are trained by minimizing the mean square error between the lth predicted attribute subvector and the lth attribute subvector corresponding to the category label of all training data, which is denoted as

[0093]

[0094] Among them, θ l and φ l denote the neural network parameters of the lth attribute feature extractor and the lth attribute sub-vector predictor respectively;

[0095] 5.3) Repeat step (5.2) until the L attribute sub-vectors corresponding to the category labels are predicted; for the kth sample x in the training data k , we can get L prediction attribute sub-vectors, recorded as

[0096] In step 4, based on the multiple prediction attribute sub-vectors in step 3, a fuzzy inference strategy is designed to preliminarily infer multiple probability results of the fault from the perspective of each attribute layer in the multiple prediction attribute sub-vectors. Based on the training data, weighted coefficient combinations of the multiple probability results are solved. Multiple attribute layers can obtain multiple weighted coefficient combinations, specifically including:

[0097] 6.1) For the kth sample in the training data, obtain L predicted attribute subvectors based on step 3 The i-th attribute layer of the l-th predicted attribute sub-vector is denoted as Where 1≤i≤L; for the i-th attribute layer, there are L-i+1 predicted attribute sub-vectors containing this attribute layer, recorded as For the i-th attribute layer of the l-th predicted attribute sub-vector Where i≤l≤L, the similarity between the attribute layer and the i-th attribute layer of all attribute vectors in the attribute vector set A is calculated to calculate the membership degree. Record

[0098]

[0099] in

[0100]

[0101] right Standardize to get

[0102]

[0103] in

[0104]

[0105] Indicates the inference of sample x from the i-th attribute layer of the l-th predicted attribute sub-vector k The probability results of m fault categories respectively; since there are L-i+1 prediction attribute sub-vectors containing the i-th attribute layer, L-i+1 probability results can be obtained

[0106] 6.2) For sample x k , the optimization result is recorded as When the i-th attribute layer of the j-th attribute vector of the attribute vector set is consistent with the sample x k When the attribute vectors corresponding to the label categories of the i-th attribute layer are the same, p i,k The jth element p i,k [j] is recorded as 1; when the i-th attribute layer of the j-th attribute vector of the attribute vector set is the same as the sample x k When the attribute vector corresponding to the label category of the i-th attribute layer is different from that of the i-th attribute layer, but is the same as that of the i-1-th attribute layer, p i,k The jth element p i,k [j] is recorded as 0.5; when it does not belong to the above two situations, p i,k The jth element p i,k [j] is recorded as 0; the optimization result is p i,k After that, its optimized form can be expressed as follows

[0107]

[0108]

[0109]

[0110] in is a weighted coefficient combination of the L-i+1 probability results, each element of which is not less than 0 and the sum is 1;

[0111] The optimization is solved by linear programming to obtain w i , and perform weighted summation on the Li-i+1 probability results to obtain the final probability result of the single attribute layer Record

[0112]

[0113] Since there are L attribute layers, a total of L weighted coefficient combinations can be obtained, denoted as [w1,w2,...,w L ], after weighting, we can get the final probability results of L single attribute layers, recorded as

[0114] In step 5, unseen fault categories from various industrial processes are collected and fed into the trained progressive knowledge transfer attribute prediction network to obtain multiple prediction attribute sub-vectors. Based on the fuzzy inference strategy from step 4 and a combination of multiple weighted coefficients, a multi-granularity attribute layer fusion strategy is designed to diagnose unseen fault categories based on the obtained multiple prediction attribute sub-vectors. Specifically, the strategy includes:

[0115] 7.1) Collect the unseen fault categories of various industrial processes, summarize the attribute vectors for the unseen fault categories of each industrial process based on the semantic description information of the unseen fault categories, and form an unseen attribute vector set in (1≤i≤m) represents the i-th attribute vector of the unseen attribute vector set, m represents the number of categories of unseen faults, and the unseen attribute vector set A U The attribute vectors in are different from the attribute vectors in the attribute vector set A;

[0116] 7.2) The unseen fault sample x U Input the trained progressive knowledge transfer attribute prediction network to obtain L predicted attribute sub-vectors, recorded as

[0117] 7.3) Based on the fuzzy inference strategy of step 4, calculate the relationship between each attribute layer of the L predicted attribute sub-vectors and the unseen attribute vector set A U The similarity of the attribute layer corresponding to all attribute vectors is normalized to obtain multiple probability results of the attribute layer, and the multiple probability results are weighted combined based on the weighted coefficient combination corresponding to the attribute layer to obtain the final probability result of the single attribute layer of the attribute layer; for L attribute layers, L single attribute layer final probability results can be obtained, which are recorded as in Represents the final probability result of a single attribute layer of the i-th (1≤i≤L) attribute layer;

[0118] 7.4) Based on the final probability results of L single attribute layers, calculate the unseen fault samples x U The probability of belonging to each unseen fault class is calculated as:

[0119] For the pth unseen fault category, its attribute vector is For the i-th attribute layer, in the unseen attribute vector set A U Find out The final probability result of the single attribute layer at the i-th attribute layer for the unseen fault category with the same i-th attribute layer In the above example, the probability corresponding to the unseen fault category is compared with the probability of the pth unseen fault category. The corresponding probabilities are added together to obtain the i-th attribute layer score s of the p-th unseen fault category p,i , for L attribute layers, the pth unseen fault category can obtain L attribute layer scores, recorded as [s p,1 ,s p,2 ,...,s p,L ], multiply the L attribute layer scores to obtain the total score s of the pth unseen fault category p ;for m Unseen fault categories are available m The total score is [s1,s2,...,s m ], normalize it and we can get m The probability of an unseen fault category is expressed as The unseen fault category with the highest probability is taken as the unseen fault sample x U The diagnosis results.

[0120] Based on the nine classification results in Table 2, the effectiveness of the present invention is demonstrated by testing the 12 faults in the training set and the three unknown faults. The effectiveness of the present invention is measured by the diagnostic accuracy.Here, Indirect Attribute Prediction (IAP) (Lampert C H, Nickisch H, Harmeling S. Attribute-based classification for zero-shot visual object categorization[J]. IEEE transactions on pattern analysis and machine intelligence, 2013, 36(3): 453-465.), Direct Attribute Prediction (DAP) (Lampert C H, Nickisch H, Harmeling S. Learning to detect unseen object classes by between-class attribute transfer[C] / / 2009 IEEE conference on computer vision and pattern recognition. IEEE, 2009: 951-958.), ESZSL (Romera-Paredes B, Torr P. An embarrassingly simple approach to zero-shot learning[C] / / International conference on machine learning. PMLR, 2015: 2152-2161.), ALE (Akata Z, Perronnin F, Harchaoui Z, et al. Label-embedding for image classification[J]. IEEE transactions on pattern analysis and machine intelligence, 2015, 38(7): 1425-1438.), and SJE (Akata Z, Reed S, Walter D, et al. Evaluation of output embeddings for fine-grained image classification[C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 2927-2936.) are selected for experimental comparison.In both IAP and DAP methods, replacing the base learner can achieve different results. Here, Random Forest (RF), Linear Support Vector Classifier (LSVC) and Radial Basis Function (RBF) are used as base learners, respectively. The results are shown in Table 3.

[0121] Experimental results show that the fault diagnosis accuracy of the proposed method ranges from 0.845 to 0.987 depending on the training / testing split. It can be seen that the proposed method achieves the best performance among all compared methods, with an average accuracy of 0.914, which is at least 2.3% higher than other methods. This demonstrates the feasibility and effectiveness of the proposed method.

[0122] Table 3 Comparison of fault diagnosis accuracy of 9 groups of fault classification using different methods (%)

[0123]

[0124] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications derived therefrom remain within the scope of protection of the present invention.

Claims

1. A multi-granularity fuzzy zero-sample fault diagnosis method with progressive knowledge transfer, characterized by: The following steps are involved: Step 1: Collect the fault categories of various industrial processes, summarize the attribute vectors for each industrial process fault category based on the semantic description information of the fault category, and form an attribute vector set; Step 2: Collect fault category samples and category labels for various industrial processes. Find the corresponding attribute vector in the attribute vector set in step 1 based on the category label. Divide the attribute vector into multiple attribute layers based on the information granularity, from coarse to fine. Starting from the first attribute layer, sequentially concatenate the attribute layers to form multiple attribute sub-vectors as training data. Step 3: Construct a progressive knowledge transfer attribute prediction network, use the training data to train the progressive knowledge transfer attribute prediction network, predict the multiple attribute sub-vectors obtained in step 2, and obtain multiple predicted attribute sub-vectors; Step 4: Based on the multiple prediction attribute sub-vectors from step 3, a fuzzy inference strategy is designed to infer the probability of the fault from each attribute layer in the multiple prediction attribute sub-vectors. A weighted coefficient combination of the probability results is calculated for each attribute layer based on the training data. Multiple weighted coefficient combinations are obtained for multiple attribute layers. Step 5: Collect samples of unseen fault categories from various industrial processes. Unseen fault category samples refer to samples of fault categories that do not appear in the training data. Input the unseen fault category samples into the progressive knowledge transfer attribute prediction network to obtain multiple predicted attribute sub-vectors. Based on the multiple weighted coefficient combinations in step 4, multiple single-attribute layer comprehensive diagnosis results are obtained. Design a multi-granularity attribute layer fusion strategy to fuse the comprehensive diagnosis results of multiple single-attribute layers, calculate the probability of each unseen fault category, and diagnose the unseen fault category samples.

2. The method according to claim 1, characterized in that In step 2, we collect fault category samples and category labels for various industrial processes, including: Collect samples of fault categories from various industrial processes and category labels , where N and J represent the number of fault category samples and the number of variables of fault category samples, respectively.

3. The method according to claim 1, characterized in that In step 2, the corresponding attribute vector is found in the attribute vector set of step 1 according to the category label. The attribute vector is divided into multiple attribute layers from coarse to fine information granularity. Starting from the first attribute layer, the attribute layers are sequentially spliced ​​to form multiple attribute sub-vectors, including: 2.1) The attribute vector set is , m represents the number of fault categories, Represents the dimension of each attribute vector; 2.2) According to the information granularity from coarse to fine, each attribute vector in the attribute vector set is divided into L attribute layers, record ,in is the attribute vector in the attribute vector set, Represents a splicing operation, represents the i-th attribute layer of the attribute vector, where , the smaller the i number is, the coarser the granularity is; 2.3) Take the first attribute layer as the first attribute sub-vector, concatenate the first two attribute layers as the second attribute sub-vector, and so on, until the first L attribute layers are concatenated as the Lth attribute sub-vector, forming a total of L attribute sub-vectors.

4. The method according to claim 1, characterized in that In step 3, a progressive knowledge transfer attribute prediction network is constructed. The progressive knowledge transfer attribute prediction network is trained using training data. The multiple attribute sub-vectors obtained in step 2 are predicted to obtain multiple predicted attribute sub-vectors, specifically including: A progressive knowledge transfer attribute prediction network is constructed to predict L attribute sub-vectors layer by layer. When predicting the next attribute sub-vector, the prior knowledge extracted from the prediction of the previous attribute sub-vector is transferred until all L attribute sub-vectors of the training data obtained in step 2 are predicted. Multiple predicted attribute sub-vectors are obtained, and the progressive knowledge transfer attribute prediction network is trained based on the mean square error between the predicted attribute sub-vectors of the training data and the attribute sub-vectors.

5. The method according to claim 4, characterized in that Construct a progressive knowledge transfer attribute prediction network, predict L attribute sub-vectors layer by layer, and transfer the prior knowledge extracted from the previous attribute sub-vector prediction when predicting the next attribute sub-vector. Specifically, 3.1) For the first attribute sub-vector, we build an attribute feature extractor and an attribute sub-vector predictor consisting of a multi-layer fully connected neural network. The attribute feature extractor extracts the attribute sub-vector features of the fault category samples in the training data. These extracted attribute sub-vector features are then input into the attribute sub-vector predictor to predict the attribute sub-vector, obtaining the first predicted attribute sub-vector. 3.2) For the lth attribute sub-vector , build an attribute feature extractor and an attribute sub-vector predictor composed of a multi-layer fully connected neural network, use the parameters of the attribute feature extractor and attribute sub-vector predictor built for the l-1th attribute sub-vector to initialize the parameters of the built attribute feature extractor and attribute sub-vector predictor respectively, and realize parameter knowledge transfer; splice the attribute sub-vector features extracted when predicting the previous attribute sub-vector with the fault category samples in the training data, input them into the attribute feature extractor to extract the attribute sub-vector features, input the extracted attribute sub-vector features into the attribute sub-vector predictor to predict the attribute sub-vector, and obtain the lth predicted attribute sub-vector; 3.3) Repeat step 3.2) until all L attribute sub-vectors are predicted, and a total of L predicted attribute sub-vectors are obtained.

6. The method according to claim 1, characterized in that In step 4, based on the multiple prediction attribute sub-vectors in step 3, a fuzzy inference strategy is designed to infer the probability of failure from each attribute layer in the multiple prediction attribute sub-vectors, specifically including: 5.1) For the i-th attribute layer , a total of Each of the predicted attribute sub-vectors contains the i-th attribute layer, denoted as ,in Indicates the The i-th attribute layer of the predicted attribute sub-vector, where ; 5.2) Calculate the The i-th attribute layer of the predicted attribute sub-vector With attribute vector set The similarity of the i-th attribute layer of all attribute vectors in is normalized and obtained from the i-th attribute layer. The probability result of the fault inference of the i-th attribute layer of the predicted attribute sub-vector is recorded as ;for prediction attribute sub-vectors, and obtain probability results, denoted as .

7. The method according to claim 1, characterized in that In step 4, each attribute layer calculates a weighted coefficient combination of the probability result based on the training data. Multiple attribute layers obtain multiple weighted coefficient combinations, specifically including: For the i-th attribute layer, Each of the predicted attribute sub-vectors contains the i-th attribute layer, and a total of Probability results ; Set the optimization goal and search for a weighted coefficient combination , where the weighted coefficient combination Each element in is not less than 0, and the sum is 1. The probability results are weighted combined, requiring the difference between the combined probability results and the optimization target to be minimized; the weighted coefficient combination Solve through linear programming; for L attribute layers, obtain L weighted coefficient combinations, recorded as .

8. The method according to claim 7, characterized in that Set optimization goals, including: The optimization goal is , is a dimensional vector, the setting rules for each element in the vector are as follows: When the i-th attribute layer of the j-th attribute vector of the attribute vector set is the same as the i-th attribute layer of the attribute vector corresponding to the label category of the fault category sample, The j-th element of is denoted as 1; When the i-th attribute layer of the j-th attribute vector of the attribute vector set is different from the i-th attribute layer of the attribute vector corresponding to the label category of the fault category sample, but is the same as the i-1-th attribute layer, The j-th element of is recorded as 0.5; If it does not fall into the above two situations, The j-th element of is denoted as 0.

9. The method according to claim 1, characterized in that In step 5, samples of unseen fault categories from various industrial processes are collected and fed into the progressive knowledge transfer attribute prediction network to obtain multiple predicted attribute subvectors. Based on the multiple weighted coefficients from step 4, multiple single-attribute layer comprehensive diagnosis results are obtained. A multi-granularity attribute layer fusion strategy is designed to fuse the comprehensive diagnosis results of multiple single-attribute layers, calculate the probability of each unseen fault category, and diagnose the unseen fault category samples. Specifically, the following steps are performed: 5.1) Collect samples of unseen fault categories of various industrial processes, summarize attribute vectors for unseen fault categories of each industrial process based on the semantic description information of the unseen fault categories, and form an unseen attribute vector set ,in represents the number of unseen fault categories; 5.2) Input the unseen fault category samples into the progressive knowledge transfer attribute prediction network to obtain L predicted attribute sub-vectors; 5.3) Calculate the L predicted attribute sub-vectors for each attribute layer and the set of unseen attribute vectors The similarity of the attribute layers corresponding to all attribute vectors is normalized to obtain multiple fault probability results inferred from a single attribute layer; 5.4) Based on the multiple weighted coefficient combinations obtained in step 4, perform a weighted combination of the fault probability results inferred from different attribute layers to obtain a comprehensive diagnosis result for a single attribute layer. For L attribute layers, obtain L comprehensive diagnosis results for each attribute layer. 5.5) Design a multi-granularity attribute layer fusion strategy to fuse the L single attribute layer comprehensive diagnosis results obtained in step (5.4), calculate the probability of each unseen fault category, and obtain the diagnosis results of the unseen fault category samples.

10. The method according to claim 9, characterized in that In step 5.5), a multi-granularity attribute layer fusion strategy is designed to fuse the L single attribute layer comprehensive diagnosis results obtained in step 5.4), calculate the probability of each unseen fault category, obtain the diagnosis results of the unseen fault category samples, and diagnose the unseen fault category samples. Specifically, the following steps are performed: For each unseen fault category, the diagnostic results that are the same as the unseen fault category in each single-attribute layer comprehensive diagnostic result are added together. For the L single-attribute layer comprehensive diagnostic results obtained, the added diagnostic results are multiplied and normalized to obtain the probability of each unseen fault category. The unseen fault category with the highest probability is taken as the diagnostic result of the unseen fault category sample.

Images