Category tree and multi-granularity fault diagnosis model construction method and hierarchical zero-shot diagnosis method for thermal power equipment faults

Abstract

The present invention discloses a method for constructing a category tree and a multi-granularity fault diagnosis model for thermal power equipment faults, as well as a hierarchical zero-sample diagnosis method. The present invention designs a category tree construction mechanism for data attribute fusion, utilizes data similarity to fuse attribute similarity, constructs a fault category tree from the bottom up, and constructs a multi-granularity fault diagnosis model, designs a layer-by-layer knowledge transfer strategy, regards the coarse-grained task model parameters as knowledge, provides guidance for fine-grained tasks, and works together with the designed attention module to improve the multi-granularity fault diagnosis model's ability to distinguish similar and unseen faults, thereby obtaining a trained multi-granularity fault diagnosis model. Utilizing the trained model, a probability fusion strategy is designed to integrate the coarse-grained task diagnosis results into the fine-grained task, thereby achieving zero-sample fault diagnosis. The present invention can mine the multi-granularity characteristics of fault types and improve the accuracy of similar and unseen fault classification in zero-sample fault diagnosis of industrial processes.

Description

Technical Field

[0001] The present invention belongs to the field of industrial fault diagnosis, and in particular aims at zero-sample fault diagnosis of industrial process data with similar unseen faults. Background Art

[0002] Fault diagnosis plays an important role in the maintenance process of industrial systems. With the development of sensors and information technology, industrial data has become increasingly accessible, which has driven the research of data-driven fault diagnosis. Generally speaking, these studies regard the fault diagnosis task as a supervised classification problem, where sufficient historical samples are provided for each category. Therefore, various supervised diagnosis models, such as support vector machines (SVM) and convolutional neural networks (CNN), can be trained to identify target fault samples. When sufficient fault samples are collected, these supervised models achieve satisfactory performance. However, considering the huge losses caused by some faults, factories usually do not let their machines run until they fail. Therefore, it is difficult and expensive to collect samples of these faults to train supervised diagnosis models.

[0003] For traditional classification tasks, missing samples for certain categories is a significant issue. In real-world industrial processes, there's an extreme case where process data for a fault type in the test set doesn't appear in the training set. In this case, the aforementioned supervised diagnosis methods fail due to the lack of fault samples in the test set. Zero-shot fault diagnosis (ZSD) is a viable solution to this problem. Zero-shot fault diagnosis can be divided into three steps: First, experts summarize the attribute descriptions of all seen and unseen faults from the fault description, including fault type, cause, and consequences. Then, combining the attribute descriptions of each fault, an attribute prediction strategy is employed to identify whether the fault possesses these attributes. Finally, a fault inference strategy is used to find the fault category that best matches the attribute description. Because attributes exist across different categories, zero-shot fault diagnosis can transfer knowledge from seen categories to unseen categories, enabling diagnosis of unseen fault categories.

[0004] However, most existing zero-shot fault diagnosis works are based on a hidden assumption: the diagnostic tasks for all faults are independent and of similar complexity. In reality, this assumption may not be correct in real industrial scenarios, as similarities between faults are typical. Industrial systems typically exhibit a tree structure, and therefore, faults occurring in industrial systems will inherit this tree structure. Some faults belong to the same coarse-grained fault, while others belong to different coarse-grained faults, resulting in different similarities between them. Identifying these similar faults is a challenging task, and therefore, existing zero-shot fault diagnosis works have difficulty separating them when they appear simultaneously in a set of unseen faults.

[0005] Existing zero-shot fault diagnosis methods have difficulty distinguishing unseen faults with high similarity. The reasons are as follows: (1) Unseen faults belonging to the same coarse-grained fault (sibling nodes) often have similar attributes, making them indistinguishable. In this case, compared with direct diagnosis, adopting a strategy of diagnosing from coarse-grained faults to fine-grained faults to gradually narrow the selection range will achieve higher accuracy. (2) For unseen faults belonging to different coarse-grained faults but with similar detailed attributes, existing zero-shot fault diagnosis models may focus too much on learning these attributes and ignore the discriminative attributes at the coarse-grained fault level. In this case, diagnosis at the coarse-grained level is a simpler and better choice. Both reasons indicate that it is possible to consider introducing a coarse-to-fine diagnosis framework in zero-shot fault diagnosis to alleviate the problem of difficulty in identifying similar faults.

[0006] Applying a coarse-to-fine diagnostic framework to zero-shot fault diagnosis presents certain difficulties. On the one hand, the prerequisite for applying this framework is to construct a tree structure for the fault. However, existing zero-shot learning methods construct category trees entirely from semantic knowledge, including attributes, pre-trained word vectors, and manual construction methods, which are not suitable for industrial scenarios. More specifically, because attributes cannot represent all the characteristics of each category and are usually labeled as discrete values, attributes alone are insufficient to comprehensively and accurately construct a fault category tree that is close to reality. In addition, due to the scarcity of expert knowledge in industrial scenarios, pre-trained word vectors or manually constructed fault category trees are difficult to obtain. On the other hand, for unseen faults belonging to the same coarse-grained fault, even if a coarse-to-fine diagnostic framework is deployed to narrow the range of options, identifying the remaining options remains a difficult task. Summary of the Invention

[0007] This invention addresses the shortcomings of existing technologies by proposing a method for constructing a category tree and multi-granularity fault diagnosis model for thermal power equipment faults, as well as a hierarchical zero-shot diagnosis method. This invention exploits the multi-granularity characteristics of fault types, improving the accuracy of similar fault classification in zero-shot fault diagnosis of industrial processes.

[0008] The object of the present invention is achieved through the following technical solutions:

[0009] A method for constructing a category tree and multi-granularity fault diagnosis model for thermal power equipment faults includes the following steps:

[0010] Step 1: Collect various types of fault data of thermal power equipment to construct a training dataset. The training dataset includes fault data of both observed faults and unobserved faults. The observed fault data includes the fault type and its corresponding industrial process data and attribute vector; the unobserved fault data includes the fault type and its corresponding attribute vector.

[0011] Step 2: Construct a category tree for thermal power equipment faults based on the training data set constructed in step 1; the height of the category tree is n, each node of the category tree corresponds to a fault type, the nodes of the first layer of the category tree are composed of all the fault types in the training data set, and the nodes of the 2nd to nth layers of the category tree are respectively composed of the clustering results of the fault data corresponding to the nodes of the previous layer; wherein, the nodes of the kth layer correspond one-to-one to the cluster categories of the fault data corresponding to the nodes of the k-1th layer, the fault type corresponding to the node is the cluster category label, and the attribute vector is the mean of the attribute vectors of all fault types in the cluster category; at the same time, mark the node type, for the nodes of the first layer, mark according to whether the fault type corresponding to the node in the training stage contains industrial process data, if not, mark it as unseen, otherwise mark it as seen; for the nodes of the 2nd to nth layers of the category tree, if all the child nodes constituting the node of the kth layer are unseen nodes, then the node is marked as unseen, if all the child nodes constituting the node of the kth layer are seen nodes, then the node is marked as seen, otherwise it is marked as other;

[0012] Step 3: Construct a multi-granularity fault diagnosis model based on the category tree constructed in step 2; the multi-granularity fault diagnosis model includes a branch shared feature extractor and n multi-granularity fault diagnosis branches corresponding one-to-one to the category tree layers, each multi-granularity fault diagnosis branch includes a branch feature extractor and a branch attribute predictor; the branch shared feature extractor is used to extract shared features of industrial process data, the branch feature extractor is used to extract branch features of the corresponding category tree layer based on the shared features, and the branch attribute predictor is used to output a predicted attribute vector of the corresponding category tree layer based on the branch features;

[0013] Step 4: Use the industrial process data in step 1 as the input of the multi-granularity fault diagnosis model, and obtain the classification probability of each fault type based on the attention vector of each branch and the predicted attribute vector of the corresponding category tree layer output by each branch attribute predictor; construct a loss function based on the classification probability for training to obtain a trained multi-granularity fault diagnosis model.

[0014] Furthermore, the clustering result of the fault data corresponding to the upper layer nodes is obtained by the following method:

[0015] Extract features from the industrial process data corresponding to each k-th node in the previous layer and calculate the centroid c of the fault type features corresponding to each k-th node and other nodes. i :

[0016]

[0017] in, Represents the process data set of the fault type corresponding to the i-th node in the k-th layer (needs to be a seen or other node), represents the industrial process data and its label of the fault type corresponding to the i-th node in the k-th layer (needs to be seen and other nodes), |*| represents the number of samples in the set *; f(*) represents the output of the feature extractor; C k represents the centroid set of the kth layer; Represents the set of fault types corresponding to the k-th layer’s seen nodes and other nodes;

[0018] Calculate the k-th layer data similarity matrix by Euclidean distance: If the i-th and j-th nodes are both seen nodes and other nodes, and the Euclidean distance between the centroids of their features is less than the first threshold Then the data similarity matrix The element value of row i and column j is set to 1 otherwise it is 0;

[0019] Calculate attribute similarity: Calculate the k-th layer attribute similarity matrix through cosine distance and Euclidean distance: If the i-th or j-th node is an unseen node, and the cosine distance of the attribute vectors corresponding to the i-th and j-th nodes is greater than the set threshold Or if the i-th and j-th nodes are both seen nodes and other nodes, and the cosine distance of the corresponding attribute vectors is greater than the set threshold The Euclidean distance between the centroid of the feature is less than the second threshold Then the attribute similarity matrix The element value of row i and column j is set to 1, otherwise it is 0;

[0020] Calculate the fusion similarity matrix Fusion similarity matrix The element value of the i-th row and j-th column is determined by the data similarity matrix and attribute similarity matrix The corresponding position element value is determined when and The element values ​​of row i and column j are all 0. The element value of row i and column j is 0, otherwise it is 1;

[0021] The fusion similarity matrix is ​​transformed into Rearrange to Then, traverse along the diagonal The first diagonal element and the diagonal element whose sum of values ​​from the diagonal to the end of the row is 1 are used as the partitioning elements. The set of k-th layer fault types corresponding to the elements between any two adjacent split elements (excluding the first diagonal element, excluding the previous split element and including the next split element) is regarded as a k-th layer fault cluster, and also represents a k+1-th layer parent class fault.

[0022] Furthermore, the classification probability of each seen or other fault type is obtained based on the attention vector of each branch and the prediction attribute vector of the corresponding category tree layer output by each branch attribute predictor, as follows:

[0023] The attention vector m of each branch k Obtained through the following methods:

[0024]

[0025] where η k+1 Indicates the total number of parent nodes in the k+1th layer that have two or more unseen or other child nodes in the kth layer, each is the standard deviation of the fault attribute vectors of all k-th layer unseen or other nodes belonging to the same parent node, and M is the dimension of the attribute vector;

[0026] The attribute vector of the k-th branch based on the prediction and the attention vector m of the kth branch k , calculate the classification probability of the i-th fault (need to be a seen or other fault) Record

[0027]

[0028] in represents the set of fault types corresponding to the k-th layer’s seen nodes and other nodes, Represents the Hadamard product.

[0029] Furthermore, the loss function is composed of a weighted cross entropy loss of each branch.

[0030] Furthermore, the loss weight of the k-th branch is expressed as:

[0031] λ k [e]=unit[ee k ]

[0032]

[0033] Among them, unit[] represents the discrete unit step function, e represents a certain training round, e k is the equidistant division point of the total number of training rounds, and ξ is the total number of training rounds.

[0034] Furthermore, when the number of training rounds reaches e k When , the parameters of the k+1th layer branch feature extractor are used to initialize the kth layer branch feature extractor, and the learning rates of the shared feature extractor and the k+1th to nth layer branch feature extractors and branch attribute predictors are reset. The reset learning rate is smaller than the original learning rate.

[0035] A hierarchical zero-shot diagnosis method for thermal power equipment faults is proposed, specifically:

[0036] Obtaining industrial process data of unseen faults belonging to the training data set and inputting them into the multi-granularity fault diagnosis model to obtain n prediction attribute vectors; calculating the classification probability of each unseen or other fault type based on the similarity between each prediction attribute vector and the attribute vector corresponding to each unseen or other node in the corresponding classification tree

[0037]

[0038] in represents the set of fault types corresponding to the unseen nodes and other nodes in the kth layer, Represents the predicted probability of the parent fault of the k-th layer fault i (needs to be an unseen or other fault) predicted by the k+1-th branch;

[0039] Take the first branch to calculate the classification probability on the unseen fault (no other faults on the first layer) The fault corresponding to the highest value is the diagnosis result.

[0040] The beneficial effects of the present invention are:

[0041] This invention primarily targets industrial process data and addresses the difficulty of classifying similar, unseen faults. It proposes a method for constructing a category tree and multi-granularity fault diagnosis model for thermal power equipment faults. This method provides a paradigm for automatically constructing fault category trees in zero-shot scenarios and introduces the concept of coarse-to-fine classification into zero-shot learning. To comprehensively and accurately construct fault category trees when expert knowledge is scarce, the invention designs a category tree construction mechanism that integrates data attributes. This mechanism goes beyond simply using attributes and instead leverages data carrying richer information to supplement fault similarity assessment. To achieve coarse-to-fine zero-shot fault diagnosis, the invention also designs a layer-by-layer knowledge transfer zero-shot fault diagnosis mechanism. Specifically, to address the difficulty of distinguishing similar, unseen faults between sibling nodes, the invention proposes a layer-by-layer knowledge transfer strategy that transfers knowledge from coarse-grained diagnostic tasks to fine-grained diagnostic tasks, providing guidance and facilitating the training of the latter. Complementing this strategy, the invention designs an attention module that focuses on discriminative attributes during training, further improving diagnostic performance for similar, unseen faults. The present invention can mine the multi-granularity characteristics of fault types and improve the accuracy of similar fault classification in zero-sample fault diagnosis of industrial processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The invention is further described below with reference to the accompanying drawings and embodiments:

[0043] Figure 1 This is a flow chart of a category tree and multi-granularity fault diagnosis model construction and a hierarchical zero-sample diagnosis method for thermal power equipment faults according to the present invention;

[0044] 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;

[0045] Figure 3 This is a framework diagram for the category tree construction and hierarchical zero-shot diagnosis method for thermal power equipment faults, which consists of a category tree construction mechanism based on data attribute fusion and a layer-by-layer knowledge transfer zero-shot fault diagnosis mechanism. DETAILED DESCRIPTION

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] This embodiment verifies the effectiveness of the method based on measured data from a thermal power plant, but is not limited thereto. Any industrial project with multi-granularity characteristics can use the method of the present invention for zero-sample fault diagnosis.

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Thermal power plants, designed to convert chemical energy into electrical energy, primarily consist of two subsystems: the boiler system and the steam turbine system. The boiler system first burns fuel to generate heat, which converts water into high-temperature, high-pressure steam, which propels the turbine blades to drive the generator. Thermal power generation is a large-scale process, and collecting samples for certain faults is costly. Therefore, zero-sample fault diagnosis is of great significance.

[0050] A method for constructing a category tree and multi-granularity fault diagnosis model for thermal power equipment faults. The method framework is shown in the figure below. Figure 1 As shown on the left, the specific modeling steps are as follows:

[0051] Step 1: Collect various types of fault data of thermal power equipment to build a training dataset

[0052] Collect fault data of various fault types to form a process data set for the training set (see category process data set) D s ={(x s ,y s )|x s ∈X s ,y s ∈Y s} and the process data set of the test set (unseen category process data set) D u ={(x u ,y u )|x u ∈X u ,y u ∈Y u}, where X s and X u They represent the sets of training samples (seen category samples) and test samples (unseen category samples), Y s and Y u Represents the set of seen and unseen fault types respectively. The training set includes the fault data of seen faults and unseen faults, where the fault data of seen faults include the fault type and its corresponding industrial process data; the fault data of unseen faults include the fault type; Based on expert knowledge, the attribute description of each fault type is obtained where a i represents the attribute vector of the i-th fault, L represents the number of fault types, and M represents the dimension of the attribute; a i Each element in takes the value "0" or "1", indicating that the fault does not meet or meets the description of the attribute; this example collects 15 types of faults, each of which contains 4000 samples, and each sample contains 68 process variables, including the temperature of the coal mill motor coil and the steam pressure at the high-pressure cylinder inlet. These faults can be described by 17 attributes, the meanings and values ​​of these attributes are shown in Tables 1 and Figure 2 shown.

[0053] Table 1 Specific Meanings of Fault Attributes

[0054]

[0055]

[0056] For the 15 types of faults in this embodiment, an 80%-20% division strategy is adopted to divide seen and unseen faults. That is, 12 types of faults are selected as seen faults, and the remaining 3 types of faults are used as unseen faults. Five different cases of seen and unseen fault divisions are designed in this embodiment to verify the zero-sample fault diagnosis effect. The division cases are shown in Table 2. Finally, the training data set contains 15 types of fault types, among which 12 are seen faults and 3 are unseen faults. The fault data of seen and unseen faults, where the fault data of seen faults includes the fault type and its corresponding industrial process data and attribute vector; the fault data of unseen faults includes the fault type and its corresponding attribute vector;

[0057] Table 2 Division of Seen and Unseen Faults

[0058]

[0059] Step 2: Construct a category tree for thermal power equipment faults based on the training data set constructed in Step 1;

[0060] By fusing data similarity and attribute similarity, construct a fault category tree with a height set to n: Each node of the category tree corresponds to a fault type. First, regard the 15 types of fault types in the above training data set as the fault nodes of the first layer (the bottom layer) of the fault category tree, and regard the process data of the corresponding seen faults and the attributes of all fault types as the process data and attributes of the first layer, that is, fault data; for the 2-nth layer, use the fault data corresponding to the nodes of the kth layer (k < n) to design a data attribute fusion strategy to obtain a fusion similarity matrix; design a matrix rearrangement and segmentation strategy to obtain a fault clustering result, form the faults of the k + 1th layer, and reassign the fault types of the nodes and calculate the attribute vectors for them. Repeat this process until an n-layer fault category tree is constructed. Among them, the nodes of the k + 1th layer correspond one-to-one with the clustering categories of the fault data corresponding to the nodes of the kth layer, the fault type corresponding to the node is the clustering category label, and the attribute vector is the mean value of the attributes of all fault types in the clustering category; at the same time, mark the node type. For the nodes of the first layer, mark them according to whether the fault type corresponding to the node in the training stage contains industrial process data. If it does not contain, mark it as unseen, otherwise mark it as seen; for the nodes of the 2-nth layer of the category tree, if all the child nodes forming the kth layer nodes are unseen nodes, then this node is marked as unseen. If all the child nodes forming the kth layer nodes are seen nodes, then this node is marked as seen, otherwise it is marked as other;

[0061] In this embodiment, the data attribute fusion strategy is calculated as follows:

[0062] Calculate data similarity: Assume that the process data set of the seen nodes and other nodes in the k-th layer of the fault category tree (k is a positive integer and k≤n) is in represents the set of fault types corresponding to the nodes seen in the kth layer and other nodes, and the jth fault in the kth layer (j∈[1,2,…,L k ], L k The attribute vector of (is the total number of faults in the kth layer) is The first time this step is performed, only the process data set of the first level categories is available. and the attribute vector of all faults in layer 1 First, feature extraction is performed on the data, and the feature extractor is represented as f(*). The feature extractor is obtained using SimSiam (Chen X, He K. Exploring simple siamese representation learning[C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2021: 15750-15758.) training method. Since only the seen nodes and other nodes have process data, the centroid set C of the fault type features corresponding to the seen and other nodes is calculated. k , recorded as

[0063]

[0064] in, Represents the process data set of the i-th node in the k-th layer (needs to be seen or other nodes), |*| represents the number of samples in the set *, and the k-th layer data similarity matrix is ​​calculated by Euclidean distance (the matrix size is L k ×L k ): If the i-th and j-th nodes are both seen nodes and other nodes, and the Euclidean distance between the centroids of their features is less than the first threshold Then the data similarity matrix The element value of row i and column j is set to 1 otherwise it is 0, denoted as

[0065]

[0066] where c i ∈C k , c j ∈Ck ,||c i ,c j || represents c i ,c j The Euclidean distance between Represents the set Euclidean distance threshold of the k-th layer centroid;

[0067] Calculate attribute similarity: Calculate the attribute similarity matrix of the kth layer by cosine distance and Euclidean distance (the matrix size is L k ×L k ): If the i-th and j-th nodes are both seen nodes and other nodes, and the cosine distance of the corresponding attribute vectors is greater than the set threshold The Euclidean distance between the centroid of the feature is less than the second threshold Or the i-th or j-th node is an unseen node, and the cosine distance of the attribute vectors corresponding to the i-th and j-th nodes is greater than the set threshold Then the attribute similarity matrix The element value of row i and column j is set to 1, otherwise it is 0, denoted as

[0068]

[0069] in represents the i-th fault in the k-th layer (i∈{1,2,…,L k}), Representative vector and The cosine distance, and Respectively represent the set k-th layer attribute cosine distance threshold and centroid Euclidean distance threshold (with different);

[0070] Calculate the fusion similarity matrix (The matrix size is L k ×L k ): The value of the element in row i and column j is given by and The corresponding position element value is determined when and The element values ​​of row i and column j are all 0. The element value of row i and column j is 0, otherwise it is 1, denoted as

[0071]

[0072] The matrix rearrangement and segmentation strategy is calculated as follows:

[0073] First, the RCM (Reverse Cuthill McKee) matrix rearrangement method is used to Rearrange to Then, traverse along the diagonal Take the first diagonal element and the diagonal element whose sum of values ​​from the diagonal to the end of the row is 1 as the partitioning element, and find the k-th layer diagonal partitioning element set PT k , recorded as

[0074]

[0075] Will The elements between any two adjacent split elements (except The k-th layer fault type set corresponding to (does not contain the previous split element, and contains the next split element) is regarded as a k-th layer fault cluster, and also represents a k+1-th layer parent class fault;

[0076] The fault type of the upper node is assigned and the upper attribute vector is calculated as follows:

[0077] According to the k-th layer fault clustering obtained by the matrix rearrangement and partitioning strategy, the seen and unseen fault types of the k+1th layer will be reallocated respectively, denoted as

[0078]

[0079]

[0080] in, and Represents the k-th layer fault obtained by clustering and The parent class fault of is obtained by the aforementioned k-th layer fault clustering. and Represents the k+1th layer fault respectively and At the same time, the attributes of the k+1th layer are obtained by averaging the attributes of the subclass faults, which is recorded as

[0081]

[0082] in represents the attribute vector of fault j at layer k, and γ j Representative satisfaction The number of faults j;

[0083] Step 3: Construct a multi-granularity fault diagnosis model based on the category tree constructed in step 2; Figure 3As shown, the multi-granularity fault diagnosis model includes a branch shared feature extractor and n multi-granularity fault diagnosis branches corresponding one-to-one to the category tree layers, each multi-granularity fault diagnosis branch includes a branch feature extractor and a branch attribute predictor; the branch shared feature extractor is used to extract shared features of industrial process data, the branch feature extractor is used to extract branch features of the corresponding category tree layer based on the shared features, and the branch attribute predictor is used to output a prediction attribute vector of the corresponding category tree layer based on the branch features;

[0084] Step 4: Use the industrial process data in step 1 as the input of the multi-granularity fault diagnosis model to train the multi-granularity fault diagnosis model;

[0085] For the training sample set X s Industrial process data sample x in s , first through the branch shared feature extractor x s Perform feature extraction to obtain shared features Record

[0086]

[0087] Among them, f FC () represents the function corresponding to the shared feature extractor;

[0088] Input to n multi-granularity fault diagnosis branches (n is the number of fault category tree layers), where the kth branch first contains a branch feature extractor and a branch attribute predictor composed of a multi-layer fully connected neural network; input shared features Extract branch features through branch feature extractor Record

[0089]

[0090] in, Represents the function corresponding to the k-th branch feature extractor;

[0091] Based on the extracted branch features Obtain the k-th layer prediction attribute vector through the branch attribute predictor Record

[0092]

[0093] in, represents the function corresponding to the k-th branch attribute predictor;

[0094] Calculate the attention vector m of the kth branch k , recorded as

[0095]

[0096]

[0097] where m k By all the total k+1 indivual Take the average to obtain; represents the k+1th level parent node with two or more kth level unseen or other child nodes, η k+1 represent The total number of By the same The standard deviation V(Q) is obtained by calculating the fault attribute set Q of all k-th layer that is not seen or other nodes. V(Q) represents the standard deviation of the same position of each attribute vector in the vector set Q, which is recorded as

[0098] V(Q)=[σ1,σ2,…,σ r ,…,σ M ]

[0099]

[0100] where ε i,r represents the rth attribute value of the i-th fault in the set Q, μ r represents the mean of the rth attribute values ​​of all faults in set Q;

[0101] The attribute vector of the k-th branch based on the prediction and the attention vector m of the kth branch k , calculate the classification probability of the i-th fault (need to be a seen or other fault) Record

[0102]

[0103] in represents the Hadamard product, exp() represents the exponential function with e as the base;

[0104] Based on classification probability Calculate the loss function of the fault diagnosis task of the k-th branch, such as cross entropy, mean square error, etc. In this embodiment, cross entropy loss is used, which is recorded as

[0105]

[0106] in Represent the learnable parameters of the shared feature extractor, branch feature extractor, and branch attribute predictor, respectively. Indicates that in the kth layer, x s Is it fault i? If yes, otherwise

[0107] The loss function of each branch is weighted to obtain the total loss function. Through backpropagation, the parameters of the model are optimized until the total number of training rounds is reached or the loss function converges, and then the trained multi-granularity fault diagnosis model can be obtained.

[0108] Furthermore, the present invention also designs a layer-by-layer knowledge transfer strategy, which is specifically as follows:

[0109] From the loss function L of the k-th branch (k = 1, 2... n) k , the total loss function is calculated in a weighted manner, denoted as

[0110] L M [e]=λ1[e]L1+…+λ n-1 [e]L n-1 +L n

[0111] where λ k [e] is the loss weight of the k-th branch, which is calculated based on the discrete unit step function and denoted as

[0112] λ k [e]=unit[e - e k

[0113] where unit[] represents the discrete unit step function, e represents a certain training round, and e k is the equally spaced division point of the total number of training rounds, denoted as

[0114]

[0115] where ξ is the total number of training rounds;

[0116] When the training round reaches e k (1≤k < n), the parameters of the k + 1-th layer branch feature extractor are used to initialize the k-th layer branch feature extractor. That is, when e = e[[ID=;49]] k , use to initialize layer-by-layer knowledge transfer, and set the learning rates of the shared feature extractor and the branch feature extractors and branch attribute predictors from the k + 1-th layer to the n-th layer to a lower value to improve the detection accuracy of the trained multi-granularity fault diagnosis model.

[0117] Based on the trained multi-granularity fault diagnosis model in step 2, hierarchical zero-shot fault diagnosis can be achieved. The samples applicable to the model of the present invention only contain the faults not seen in the training dataset. In the actual industrial process, it can be pre-judged by expert knowledge that it may be one of the faults not seen in advance, and then the model of the present invention is used for diagnosis.

[0118] ​Specifically, the present invention provides a hierarchical zero-sample diagnosis method for thermal power equipment faults, such as Figure 1 As shown in the test phase on the far right, firstly, the industrial process data without any faults belonging to the training data set is obtained and input into the multi-granularity fault diagnosis model. u The test data x in u , obtain the predicted attribute vector of branch k through the shared feature extractor, the feature extractor of branch k and the attribute predictor A total of n prediction attribute vectors are obtained; a probability fusion strategy is designed to pass the upper layer prediction results to the lower layer to obtain x u Classification probability of the i-th fault on the k-th layer (needs to be unseen or other faults) Record

[0119]

[0120] in represents the set of fault types corresponding to the unseen nodes and other nodes in the kth layer, Represents the predicted probability of the parent fault of the k-th layer fault i (needs to be an unseen or other fault) predicted by the k+1-th branch;

[0121] Finally, the classification probability of the first branch on the unseen fault (no other faults on the first layer) is calculated The fault corresponding to the highest value is taken as the diagnosis result, recorded as

[0122]

[0123] Where R(x u ) is the unseen fault data x u Final fault diagnosis results.

[0124] Based on the five classification results in Table 2, the 12 faults in the data set are considered as observed faults, and the test results of the three unobserved faults are used to illustrate the effectiveness of the present invention. The effectiveness of the present invention is measured by the diagnostic accuracy. Here, direct attribute prediction DAP (Lampert CH, Nickisch H, Harmeling S. Learning to detect unseenobject classes by between-class attribute transfer [C] / / 2009IEEE conference oncomputer vision and pattern recognition. IEEE, 2009:951-958.), IAP (Lampert CH, Nickisch H, Harmeling S. Attribute-based classification for zero-shot visual object categorization[J].IEEE transactions on pattern analysis and machineintelligence,2013,36(3):453-465.), ESZSL(Romera-Paredes B,Torr P.Anembarrassingly simple approach to zero-shot learning[C] / / Internationalconference on machine learning.PMLR,2015:2152-2161.), Devise(Frome A,Corrado GS,Shlens J,et al.Devise:A deep visual-semantic embedding model[J].Advances inneural information processing systems,2013,26.),ALE(Akata Z,Perronnin F,Harchaoui Z,et al.Label-embedding for image classification[J].IEEEtransactions on pattern analysis and machine intelligence,2015,38(7):1425-1438.),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.) was used for experimental comparison. In the DAP and IAP methods, replacing the base learner can achieve different results. Here, Random Forest (RF), Probabilistic Naive Bayes (Naive Bayes), and Radial Basis Function (RBF) are used as the base learners, respectively. The results are shown in Table 3.

[0125] Experimental results show that the fault diagnosis accuracy of the proposed method ranges from 0.7823 to 0.9325 depending on the training / testing split. The average accuracy of the proposed method across five groups is 0.8621, achieving the best performance among all compared methods. Furthermore, the diagnostic accuracy for each group is at least 9% higher than that of other methods. This demonstrates the feasibility and effectiveness of the proposed method.

[0126] Table 3 Comparison of fault diagnosis accuracy of 9 groups of fault classification using different methods (%)

[0127]

[0128] 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 method for constructing a category tree and multi-granularity fault diagnosis model for thermal power equipment faults, characterized in that: The following steps are involved: Step 1: Collect various types of fault data of thermal power equipment to construct a training dataset. The training dataset includes fault data of both observed faults and unobserved faults. The observed fault data includes the fault type and its corresponding industrial process data and attribute vector. The fault data of unseen faults include the fault type and its corresponding attribute vector; Step 2: Construct a category tree for thermal power equipment faults based on the training data set constructed in step 1; the height of the category tree is n, each node of the category tree corresponds to a fault type, the nodes of the first layer of the category tree are composed of all the fault types in the training data set, and the nodes of the 2nd to nth layers of the category tree are respectively composed of the clustering results of the fault data corresponding to the nodes of the previous layer; wherein, the nodes of the kth layer correspond one-to-one to the cluster categories of the fault data corresponding to the nodes of the k-1th layer, the fault type corresponding to the node is the cluster category label, and the attribute vector is the mean of the attribute vectors of all fault types in the cluster category; at the same time, mark the node type, for the nodes of the first layer, mark according to whether the fault type corresponding to the node in the training stage contains industrial process data, if not, mark it as unseen, otherwise mark it as seen; for the nodes of the 2nd to nth layers of the category tree, if all the child nodes constituting the node of the kth layer are unseen nodes, then the node is marked as unseen, if all the child nodes constituting the node of the kth layer are seen nodes, then the node is marked as seen, otherwise it is marked as other; Step 3: Construct a multi-granularity fault diagnosis model based on the category tree constructed in step 2; the multi-granularity fault diagnosis model includes a branch shared feature extractor and n multi-granularity fault diagnosis branches corresponding one-to-one to the category tree layers, each multi-granularity fault diagnosis branch includes a branch feature extractor and a branch attribute predictor; the branch shared feature extractor is used to extract shared features of industrial process data, the branch feature extractor is used to extract branch features of the corresponding category tree layer based on the shared features, and the branch attribute predictor is used to output a predicted attribute vector of the corresponding category tree layer based on the branch features; Step 4: Use the industrial process data in step 1 as the input of the multi-granularity fault diagnosis model, and obtain the classification probability of each fault type based on the attention vector of each branch and the predicted attribute vector of the corresponding category tree layer output by each branch attribute predictor; construct a loss function based on the classification probability for training to obtain a trained multi-granularity fault diagnosis model.

2. The method according to claim 1, characterized in that The clustering result of the fault data corresponding to the upper layer nodes is obtained by the following method: Extract features from the industrial process data corresponding to each seen node and other nodes in the previous layer k and calculate the centroid c of the fault type features corresponding to each seen node and other nodes i : in, represents the process data set of the fault type corresponding to the i-th node in the k-th layer, represents the industrial process data and its label of the fault type corresponding to the i-th node in the k-th layer. The i-th node must be a seen or other node. |*| represents the number of samples in the set *; f(*) represents the output of the feature extractor; C k represents the centroid set of the kth layer; Represents the set of fault types corresponding to the k-th layer’s seen nodes and other nodes; Calculate the k-th layer data similarity matrix by Euclidean distance: If the i-th and j-th nodes are both seen nodes and other nodes, and the Euclidean distance between the centroids of their features is less than the first threshold Then the data similarity matrix The element value of row i and column j is set to 1 otherwise it is 0; Calculate attribute similarity: Calculate the k-th layer attribute similarity matrix through cosine distance and Euclidean distance: If the i-th or j-th node is an unseen node, and the cosine distance of the attribute vectors corresponding to the i-th and j-th nodes is greater than the set threshold Or if the i-th and j-th nodes are both seen nodes and other nodes, and the cosine distance of the corresponding attribute vectors is greater than the set threshold The Euclidean distance between the centroid of the feature is less than the second threshold Then the attribute similarity matrix The element value of row i and column j is set to 1, otherwise it is 0; Calculate the fusion similarity matrix Fusion similarity matrix The element value of the i-th row and j-th column is determined by the data similarity matrix and attribute similarity matrix The corresponding position element value is determined when and The element values ​​of row i and column j are all 0. The element value of row i and column j is 0, otherwise it is 1; The fusion similarity matrix is ​​transformed into Rearrange to Then, traverse along the diagonal The first diagonal element and the diagonal element whose sum of values ​​from the diagonal to the end of the row is 1 are used as the partitioning elements. The set of k-th layer fault types corresponding to the elements between any two adjacent split elements is regarded as a k-th layer fault cluster, and also represents a k+1-th layer parent type fault.

3. The method according to claim 1, characterized in that The classification probability of each seen or other fault type is obtained based on the attention vector of each branch and the prediction attribute vector of the corresponding category tree layer output by each branch attribute predictor, as follows: The attention vector m of each branch k Obtained through the following methods: where η k+1 Indicates the total number of parent nodes in the k+1th layer that have two or more unseen or other child nodes in the kth layer, each is the standard deviation of the fault attribute vectors of all k-th layer unseen or other nodes belonging to the same parent node, and M is the dimension of the attribute vector; The attribute vector of the k-th branch based on the prediction and the attention vector m of the kth branch k , calculate the classification probability of the i-th fault The i-th fault must be a seen or other fault, recorded as in represents the set of fault types corresponding to the k-th layer’s seen nodes and other nodes, represents the Hadamard product.

4. The method according to claim 1, wherein The loss function is composed of the weighted cross entropy loss of each branch.

5. The method according to claim 4, characterized in that The loss weight of the k-th branch is expressed as: λ k [e]=unit[e k ] Among them, unit[] represents the discrete unit step function, e represents a certain training round, e k is the equidistant division point of the total number of training rounds, and ξ is the total number of training rounds.

6. The method according to claim 5, characterized in that After the number of training rounds reaches e k When , the parameters of the k+1th layer branch feature extractor are used to initialize the kth layer branch feature extractor, and the learning rates of the shared feature extractor and the k+1th to nth layer branch feature extractors and branch attribute predictors are reset. The reset learning rate is smaller than the original learning rate.

7. A hierarchical zero-sample diagnosis method for thermal power equipment faults, characterized in that: Specifically: Obtaining industrial process data of unseen faults and inputting it into a multi-granularity fault diagnosis model constructed by the method according to any one of claims 1 to 6 to obtain n prediction attribute vectors; calculating the classification probability of each unseen or other fault type based on the similarity between each prediction attribute vector and the attribute vector corresponding to each unseen or other node in the corresponding classification tree in represents the set of fault types corresponding to the unseen nodes and other nodes in the kth layer, represents the predicted probability of the parent fault of fault i at layer k predicted by the k+1th branch. Fault i must be an unseen or other fault; Take the first branch to calculate the classification probability of unseen faults The fault corresponding to the highest value is the diagnosis result.

Images