Graphic automatic encoder fault case clustering method based on multi-feature fusion
The graph autoencoder-based fault case clustering method addresses the limitations of existing algorithms by integrating multi-feature fusion, optimizing graph embedding and clustering, resulting in accurate and efficient fault case recommendations.
Patent Information
- Application Number
- CN202411149290.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2024-12-31
AI Technical Summary
Existing graph clustering algorithms fail to fully utilize the characteristics and structural information of graph data, leading to inaccurate fault case recommendations due to their adaptation to Euclidean space rather than graph data.
A graph autoencoder-based fault case clustering method that integrates multi-feature fusion, utilizing an encoder, decoder, and self-training clustering unit to optimize graph embedding and clustering, enhancing the accuracy and efficiency of node division and clustering.
Improves the accuracy and efficiency of fault case clustering by embedding similar cases close together in a low-dimensional space, facilitating fault prediction, diagnosis, and repair through joint optimization of graph embedding and clustering.
Smart Images
Figure CN119226509A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of fault diagnosis, and in particular to a fault case clustering method of a graph automatic encoder based on multi-feature fusion. Background Art
[0002] Product failure case data contains a large amount of characteristic information, such as product model, product usage time, cause of failure, etc. It can provide data support for subsequent fault diagnosis.
[0003] In the prior art, the research on graph clustering algorithms is based on the application of classic clustering algorithms on the basis of graph embedding, such as k-means or spectral clustering algorithms. Although this two-step framework can solve the problem of graph clustering, it has certain limitations: graph embedding technology can map graph data into a low-dimensional vector space, but it cannot fully utilize the characteristics and structural information of graph data, and a single clustering algorithm is usually designed for data in Euclidean space, which may not be well adapted to graph data, affecting the final clustering results, resulting in inaccurate recommended fault cases.
[0004] Therefore, it is necessary to provide a graph autoencoder fault case clustering method based on multi-feature fusion to achieve the joint optimization of graph embedding and clustering process, so as to efficiently and accurately divide nodes in the feature space and improve the accuracy of node division and clustering efficiency. Summary of the invention
[0005] The present invention provides a fault case clustering method based on a multi-feature fusion-based graphic autoencoder, comprising: obtaining a plurality of fault cases; establishing a fault case graph structure based on the plurality of fault cases; determining a plurality of case clustering clusters and a cluster center of each of the case clustering clusters based on the fault case graph structure through a case clustering model, wherein the case clustering model at least includes an encoder, a decoder and a self-training clustering unit; obtaining a current fault description text; determining an embedded representation of the current fault description text; determining a target case clustering cluster based on the embedded representation of the current fault description text and the cluster center of each of the case clustering clusters; and searching for a target fault case from the target case clustering cluster based on the embedded representation of the current fault description text.
[0006] Furthermore, the case clustering model determines multiple case clustering clusters and the cluster center of each case clustering cluster based on the fault case graph structure, including: obtaining the embedded representation of each fault case node through the encoder; generating a reconstructed adjacency matrix between any two fault case nodes based on the embedded representation of each fault case node through the decoder; and determining multiple case clustering clusters and the cluster center of each case clustering cluster based on the embedded representation of each fault case node and the reconstructed adjacency matrix between two fault case nodes through the self-training clustering unit.
[0007] Furthermore, through the encoder, an embedded representation of each fault case node is obtained, including: for each of the fault case nodes, feature information of the second-order neighbor nodes of the fault case node is aggregating to generate an embedded representation of the first-order neighbor nodes of the fault case node, and embedding features of the first-order neighbor nodes are aggregating to generate a low-dimensional embedded representation of the fault case node; cluster distribution of the fault case nodes is learned based on the low-dimensional embedded representation of the fault case nodes, and the cluster information vector of the fault case nodes is determined through the attention layer; and an embedded representation of the fault case node is generated based on the low-dimensional embedded representation and cluster information vector of the fault case node.
[0008] Further, the cluster information vector of the fault case node is determined based on the following formula: in, is the cluster information vector learned for the i-th fault case node, α ij is the attention coefficient of the pooling vector between the i-th fault case node and the j-th cluster, W is the weight matrix, is the cluster center of the jth cluster, k is the number of clusters, δ is the nonlinear activation function, is the weight vector of the attention mechanism, || is the concatenation operation of vectors.
[0009] Furthermore, the target total loss function of the self-training clustering unit is: L = L G +αL KL , where L is the target total loss, L G To indicate loss, L KL is the clustering loss, and α is the coefficient used to balance the balance between representation loss and clustering loss.
[0010] Furthermore, the representation loss is: Where n is the total number of fault case nodes, a ij is the element in the i-th row and j-th column of the original adjacency matrix, is the element in the i-th row and j-th column of the reconstructed adjacency matrix.
[0011] Furthermore, the clustering loss is: Where KL(·) is the KL divergence, Q is the membership distribution, P is the target distribution, and Z i is the embedding representation of the i-th fault case node, m j is the jth cluster center, m j′ is the j′th cluster center, q ij is the similarity between the embedding representation of the i-th fault case node and the j-th cluster center, p ij is the target distribution.
[0012] Further, determining the cluster center of the case clustering cluster includes: calculating a weighted average of the embedded representations of all fault case nodes in the case clustering cluster to determine the cluster center of the case clustering cluster.
[0013] Furthermore, based on the embedded representation of the current fault description text and the cluster center of each of the case clustering clusters, a target case clustering cluster is determined, including: calculating the cosine similarity between the embedded representation of the current fault description text and the cluster center of each of the case clustering clusters; based on the cosine similarity between the embedded representation of the current fault description text and the cluster center of each of the case clustering clusters, determining the target case clustering cluster.
[0014] Furthermore, based on the embedded representation of the current fault description text, a target fault case is searched from the target case clustering cluster, including: based on the embedded representation of the current fault description text, calculating the similarity between the embedded representation of the current fault description text and each fault case node included in the target case clustering cluster; based on the similarity between the embedded representation of the current fault description text and each fault case node included in the target case clustering cluster, searching for the target fault case from the target case clustering cluster.
[0015] Compared with the prior art, the fault case clustering method of the graph automatic encoder based on multi-feature fusion provided by the present invention has at least the following beneficial effects:
[0016] 1. The method of combining graph embedding and clustering analysis is adopted. The objective function is defined through a joint optimization strategy, and the graph embedding learning is closely combined with the node clustering process. By jointly optimizing the objective function, the joint optimization of graph embedding and clustering processes is achieved to efficiently and accurately divide nodes in the feature space. Not only the accuracy of node division is improved, but also the clustering efficiency is improved. Cases with similar fault characteristics or performance patterns are embedded in a low-dimensional space through the clustering process, so that these similar fault cases are close in distance in the low-dimensional space, which facilitates the classification, analysis and identification of faults. This clustering method helps to extract valuable information from a large amount of fault data, providing strong support for fault prediction, diagnosis and repair.
[0017] 2. Through the results of fault case clustering and similarity calculation, the most similar target case clustering cluster is determined, and the target fault cases similar to the current fault description text are found in the target case clustering cluster. These target fault cases correspond to specific fault cases in the original data set. In order to restore the target fault case to a complete text case, the feature vector of the node is mapped back to the original data set to extract information including product model, product usage time, cause of fault, fault time, fault phenomenon, cause of fault and solution. Finally, this information is recombined into a complete text case to provide users with detailed fault diagnosis and solutions. This method effectively solves the task of finding similar cases based on the cluster center and provides corresponding references and solutions. This method not only takes into account the global clustering structure, but also ensures the accuracy and practicality of the search results. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:
[0019] Figure 1 is a flowchart of a method for clustering fault cases of a graphic autoencoder based on multi-feature fusion according to some embodiments of this specification;
[0020] Figure 2 is a schematic diagram of node sampling according to some embodiments of this specification;
[0021] Figure 3 is a schematic diagram of aggregate learning according to some embodiments of this specification;
[0022] Figure 4 It is a schematic diagram of the node embedding results shown in some embodiments of this specification. DETAILED DESCRIPTION
[0023] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.
[0024] Figure 1 is a flow chart of a method for clustering fault cases of a graphic automatic encoder based on multi-feature fusion according to some embodiments of this specification, such as Figure 1 As shown, the graph autoencoder fault case clustering method based on multi-feature fusion can include the following steps.
[0025] Step 110, obtaining multiple fault cases.
[0026] Among them, the failure case may include characteristic information such as product model, product usage time, and cause of the failure.
[0027] Step 120: Establish a fault case graph structure based on multiple fault cases.
[0028] As an example only, the fault case diagram structure can be established based on the following process:
[0029] S1. Fault case collection: Fault cases are collected from multiple channels such as repair stations, product manufacturers, customer feedback, warranty records, etc. Each fault case should include detailed information such as product model, product usage time, cause of the fault, fault description, fault phenomenon, and fault solution.
[0030] S2. Data cleaning: Clean the collected fault cases, remove duplicate, incomplete or erroneous data, standardize the data and ensure the consistency of data format.
[0031] S3. Feature extraction: Extract key feature information from the failure case, such as product model, product usage time, cause of failure, etc. The features can be further processed as needed, such as encoding, conversion, etc. The extracted feature information is organized into structured data for subsequent analysis and processing.
[0032] S4. Statistical analysis: Conduct statistical analysis on failure cases to understand the distribution of failures under different product models, product usage duration, and other conditions.
[0033] S5. Graph structure construction: In the graph, the edges between nodes are constructed according to the following rules: ① The product model is the same. If the product model of two nodes is the same, there is an edge between them; ② The product usage time is similar: If the product usage time of two nodes is similar within a certain range (for example, the difference in product usage time does not exceed 50 hours), there is an edge between them; ③ The cause of the failure is the same. If the cause of the failure of two nodes is the same, there is an edge between them; ④ The season is the same. If the failure time of two nodes corresponds to the same season, there is an edge between them; ⑤ The fault phenomenon, fault cause and solution are similar. The cosine similarity is used to calculate the text similarity of the fault phenomenon, fault cause and solution. In this study, the text similarity threshold is set to 80%. This setting is based on the specific characteristics of the fault phenomenon and solution for analysis. The description of the fault phenomenon and solution is already the standardized result after the case knowledge base is processed. It has a certain quality and consistency. The description is relatively brief and contains specific keywords. Setting the threshold to 80% can effectively distinguish and ensure that only similar fault phenomena will be identified as related, thereby building a more accurate graph structure. When the text similarity of two fault phenomena, fault causes or solutions exceeds 80%, they are considered to have sufficient similarity and there is an edge between the two nodes where they are located.
[0034] The graph is defined as G = (V, E, X), where V is the set of fault case nodes, and V represents the set of fault case nodes V = (v1, v2, ..., v n ), E is a set of edges representing a set of edges between fault case nodes. The topological structure of the graph G can be represented by the adjacency matrix A, and A∈R is defined. n×n , if (v i , v j )∈E, then A i，j =1, otherwise A i，j = 0. X = (x1, x2, ..., x n ) is the characteristic value of the fault case node, where x i ∈R m is related to the fault case node v i The main purpose of attribute graph embedding is to map the nodes of the graph to a low-dimensional space while maintaining the topological information and feature information of the original attribute graph, and then apply the vector to the node clustering task: for the graph G, the goal is to divide the nodes in G into k disjoint groups G = (G1, G2, ..., G k ), so that nodes in the same cluster are usually close to each other in terms of graph structure, far away from each other in other aspects, and may have more similar attribute values.
[0035] Step 130 , determining a plurality of case clusters and a cluster center of each case cluster based on the fault case graph structure through a case clustering model.
[0036] The case clustering model at least includes an encoder, a decoder and a self-training clustering unit.
[0037] In some embodiments, determining a plurality of case clusters and a cluster center of each case cluster based on the fault case graph structure through a case clustering model includes:
[0038] Obtain the embedded representation of each fault case node through the encoder;
[0039] The decoder generates a reconstructed adjacency matrix between any two fault case nodes based on the embedding representation of each fault case node:
[0040] A plurality of case clustering clusters and the cluster center of each case clustering cluster are determined based on the embedded representation of each fault case node and the reconstructed adjacency matrix between two fault case nodes through a self-training clustering unit.
[0041] In some embodiments, obtaining, by an encoder, an embedded representation of each fault case node includes:
[0042] For each fault case node, aggregate the feature information of the second-order neighbor nodes of the fault case node to generate an embedded representation of the first-order neighbor nodes of the fault case node, aggregate the embedded features of the first-order neighbor nodes to generate a low-dimensional embedded representation of the fault case node;
[0043] The cluster distribution of fault case nodes is learned based on the low-dimensional embedding representation of fault case nodes, and the cluster information vector of the fault case nodes is determined through the attention layer;
[0044] Based on the low-dimensional embedding representation of the fault case node and the cluster information vector, an embedding representation of the fault case node is generated.
[0045] Specifically, in the fault case graph structure, the number of neighbors of the fault case nodes may vary greatly. Some fault case nodes may have many neighbors, while some fault case nodes have very few neighbors. In order to effectively handle neighbor sets of different sizes, the encoder uses a sampling strategy. In order to reduce the complexity of calculation and improve the calculation efficiency, the encoder adopts local sampling. Each node only needs to consider its surrounding neighbors, which helps to reduce the complexity of calculation. Figure 2 is a schematic diagram of node sampling according to some embodiments of this specification, such as Figure 2 As shown in the figure, for each fault case node, the entire fault case graph structure is divided into K layers from the inside to the outside, and the fault case node is processed as the center. The number of hops K of each fault case node is set to 2. When K = 1, at the fault case node vi After sampling, the neighbor node information is aggregated to generate an embedded representation of the fault case node; Figure 3 is a schematic diagram of aggregate learning according to some embodiments of this specification, such as Figure 3 As shown, the fault case node v i By aggregating the feature information of the sampled second-order neighbor nodes, the embedded representation of the first-order neighbor nodes is generated, and then the embedded representation of the sampled first-order neighbor nodes is aggregated to generate the fault case node v i For example, the fault case node v i The characteristic information of the second-order neighbor nodes and the fault case node v i The feature information of the first-order neighbor nodes is merged to ensure that each fault case node contains not only the information of direct neighbors but also the information of indirect neighbors, so that the low-dimensional embedding representation is more comprehensive. Then, the feature information of these merged second-order neighbor nodes is aggregated with the feature information of the first-order neighbor nodes into the fault case node, further enriching the low-dimensional embedding representation of the fault case node. Finally, the low-dimensional embedding representation of the aggregated fault case node is processed to generate the final embedding representation of the fault case node.
[0046] In order to generate the feature representation of the core node, we first learn the pooling aggregation function. This process involves nonlinearly transforming the representation vectors of the neighbor nodes of the core node, and then performing a pooling operation on the transformed feature representation vectors. Next, we perform nonlinear transformations on the pooled node features and the feature representation of the core node. Finally, we superimpose the transformed structures of the two to obtain the feature representation of the core node at this layer.
[0047]
[0048]
[0049] in, represents the feature representation of all neighbor nodes of the fault case node v at the kth layer; k represents the number of layers of the fault case graph structure; σ represents the selected nonlinear activation function; W represents the weight matrix; represents the embedding of the neighbor node u of the fault case node v in layer k; b is the bias; represents the feature representation of the fault case node v at the kth layer; N(v) is defined as extracting a fixed number of neighbors from all neighbor nodes of the fault case node v. To represent the multi-layer representation of each node, L is the number of layers used, forming the final low-dimensional embedding representation Z of the fault case node v , which contains all the hierarchical information.
[0050] The process of cluster distribution is expressed as:
[0051]
[0052] Among them, the cluster distribution process is represented by C t , the number of clusters is k, n k is the case cluster The number of nodes in Z, m is v After learning the cluster distribution, an average pooling operation is performed on each case cluster to finally obtain a representative node:
[0053]
[0054] in, is a representative node in the case cluster, Concat is a concatenation operation, and MeanPo0ling is an average pooling operation.
[0055] An attention layer is added to further integrate the cluster information of the fault case nodes. The attention mechanism captures more important neighbor node features by assigning different weights to different neighbor nodes to measure the influence vector of each cluster on it.
[0056] In some embodiments, the cluster information vector of the fault case node is determined based on the following formula:
[0057]
[0058]
[0059] in, is the cluster information vector learned for the i-th fault case node, α ij is the attention coefficient of the pooling vector between the i-th fault case node and the j-th cluster, W is the weight matrix, is the cluster center of the jth case cluster, k is the total number of case clusters, δ is the nonlinear activation function, is the weight vector of the attention mechanism, is the low-dimensional embedding representation of the i-th fault case node, and || is the concatenation operation of the vectors.
[0060] Use the residual block to fuse the low-dimensional embedding representation of the fault case node and the cluster information vector to obtain the final embedding representation:
[0061] Z=Z v +Z c
[0062] Where Z is the embedded representation of the fault case node, is the size of the low-dimensional embedding representation of the i-th fault case node, is the cluster information vector.
[0063] The encoder can effectively fuse the topological information and feature information of the fault case nodes in the graph, and integrate the feature information of its neighboring nodes. By learning the cluster distribution of nodes, the cluster information is integrated into the embedding generation process, which improves the accuracy of the embedding. By introducing the attention mechanism, the influence of different clusters is integrated into the embedding, making the embeddings more similar in the same class and easier to distinguish in different classes.
[0064] Use the inner product decoder to reconstruct the graph:
[0065]
[0066] in, represents the value of the element in the i-th row and j-th column of the reconstructed adjacency matrix, z i and z j They represent the embedded representations of the i-th fault case node and the j-th fault case node, respectively, and use the sigmoid activation function. The inner product decoder predicts the connection relationship between fault case nodes by calculating the inner product of the embedded representations of the fault case nodes. The inner product can directly measure the similarity of the embedded representations of the fault case nodes and reflect the connection strength between the fault case nodes.
[0067] The decoder can also optimize node embedding through back propagation so that it can restore node features well.
[0068] In some embodiments, the objective total loss function of the self-training clustering unit is:
[0069] L=L G +αL KL
[0070] Among them, L is the target total loss, L G To represent the loss, the main purpose is to minimize the cross entropy between the reconstructed adjacency matrix and the original adjacency matrix, L KL is the clustering loss, and α is the coefficient used to balance the balance between representation loss and clustering loss.
[0071] In some embodiments, the loss is represented by:
[0072]
[0073] Where n is the total number of fault case nodes, a ij is the element in the i-th row and j-th column of the original adjacency matrix, is the element in the i-th row and j-th column of the reconstructed adjacency matrix.
[0074] In some embodiments, the clustering loss is:
[0075]
[0076]
[0077]
[0078] Where KL(·) is the KL divergence, Q is the membership distribution, P is the target distribution, and Z i is the embedding representation of the i-th fault case node, m j is the jth cluster center, m j′ is the j′th cluster center, q ij is the similarity between the embedding representation of the i-th fault case node and the j-th cluster center, p ij is the target distribution.
[0079] Figure 4 is a schematic diagram of node embedding results according to some embodiments of this specification, such as Figure 4 As shown, the clustering results are obtained from the final optimized Q, and the fault case node v i Assign label Label:
[0080]
[0081] Among them, u i is the label of the i-th fault case node, argmax j q ij The index for selecting the case cluster with the greatest similarity.
[0082] The self-trained clustering unit aggregates cases with high similarity into the same cluster, forming a subgraph with obvious internal connections.
[0083] In some embodiments, determining a cluster center of a case cluster includes:
[0084] Calculate the weighted average of the embedding representations of all fault case nodes in the case clustering cluster and determine the cluster center of the case clustering cluster. For example:
[0085]
[0086] Where |C| is the number of fault case nodes in the case cluster, f x is the embedding representation of the fault case x in the case cluster, c C is the cluster center of the case cluster.
[0087] Step 140, obtaining the current fault description text.
[0088] The current fault description text may include information such as fault characteristics.
[0089] Step 150, determining an embedded representation of the current fault description text.
[0090] The input current fault description text is processed by text vectorization, and the same embedding method as step 130 is required to convert it into an embedded representation. After inputting the current fault description text and other information (for example, product model, product usage time, etc.), a unique id is assigned to it as a sample point in the case space, and the current fault description text and other information are used as features to form graph data with the sample point. The BERT (Bidirectional Encoder Representations from Transformers) model is used to convert the input current fault description text into a high-dimensional vector representation y, and the encoder layer of GA-NNF (Graph Auto-encoder Based on Neighbor Node Features) is used to embed the high-dimensional vector representation y into the same low-dimensional vector space as the fault case node. The embedding process can be expressed as:
[0091] Z y =GANNF(A,Y)
[0092] Among them, Z y is the embedded representation of the current fault description text in the low-dimensional space, A is the adjacency matrix of the fault case graph structure, and Y is the feature matrix of the nodes in the graph, including the features of y and other nodes.
[0093] Step 160 , determining a target case cluster based on the embedded representation of the current fault description text and the cluster center of each case cluster.
[0094] Specifically include:
[0095] Calculate the cosine similarity between the embedding representation of the current fault description text and the cluster center of each case cluster;
[0096] Based on the cosine similarity between the embedding representation of the current fault description text and the cluster center of each case clustering cluster, the target case clustering cluster is determined. For example, the case clustering cluster where the cluster center with the highest cosine similarity is located can be used as the target case clustering cluster.
[0097] For example, the cosine similarity between the embedding representation of the current fault description text and the cluster center of each case cluster can be calculated based on the following formula:
[0098]
[0099] Among them, Similarity(Z y , Z C) is the cosine similarity between the embedding representation of the current fault description text and the embedding representation of the cluster center of each case cluster, where Z y is the embedding representation of the current fault description text, Z C Embedded representation of the cluster centers of case clusters.
[0100] In some embodiments, the target case cluster may also be determined based on the following process:
[0101] Sampling multiple fault cases from each case cluster as sample fault cases;
[0102] For each sample fault case, the cosine similarity between the embedding representation of the sample fault case and the cluster center of each case cluster is calculated, and based on the cosine similarity between the embedding representation of the sample fault case and the cluster center of each case cluster, a similarity sequence corresponding to the sample fault case is generated, wherein an element in the similarity sequence represents the cosine similarity between the embedding representation of the sample fault case and the cluster center of a case cluster;
[0103] Calculate the cosine similarity between the embedding representation of the current fault description text and the cluster center of each case cluster;
[0104] Based on the cosine similarity between the embedding representation of the current fault description text and the cluster center of each case cluster, the cosine similarity variance is calculated;
[0105] When the cosine similarity variance is greater than the variance threshold, the case clustering cluster where the cluster center with the highest cosine similarity is located can be used as the target case clustering cluster;
[0106] When the cosine similarity variance is less than or equal to the variance threshold, based on the cosine similarity between the embedding representation of the current fault description text and the cluster center of each case clustering cluster, generate a similarity sequence corresponding to the current fault description text, calculate the sequence similarity between the similarity sequence corresponding to the current fault description text and the similarity sequence corresponding to each sample fault case, and determine the matching score of each case clustering cluster based on the cosine similarity between the embedding representation of the current fault description text and the cluster center of each case clustering cluster and the sequence similarity between the similarity sequence corresponding to the current fault description text and the similarity sequence corresponding to each sample fault case. If there is at least one case clustering cluster whose matching score is greater than the matching score threshold, the case clustering cluster with the matching score greater than the matching score threshold is used as the target case clustering cluster. If there is no at least one case clustering cluster whose matching score is greater than the matching score threshold, the case clustering cluster with the largest matching score is used as the target case clustering cluster.
[0107] For example, the cosine similarity variance can be calculated based on the following formula:
[0108]
[0109] Among them, σ Similarity is the cosine similarity variance, Similarity(Z y , Z C(i) ) is the cosine similarity between the embedding representation of the current fault description text and the embedding representation of the cluster center of the i-th case cluster, and I is the total number of case clusters.
[0110] The sequence similarity can be calculated based on the following formula:
[0111]
[0112] Among them, S j is the sequence similarity between the similarity sequence corresponding to the current fault description text and the similarity sequence corresponding to the jth sample fault case, Similarity(Z j , Z C(i) ) is the cosine similarity between the embedding representation of the current fault description text and the embedding representation of the jth sample fault case.
[0113] The matching score of the case cluster can be calculated based on the following formula:
[0114]
[0115] Among them, M i is the matching score of the i-th case cluster, α1 and α2 are weights, and α1+α2=1, α1 and α2 are both greater than 0, and J is the total number of sample fault cases included in the i-th case cluster.
[0116] It can be understood that the above process not only relies on a single cosine similarity to determine the target case cluster, but also considers the sequence similarity of the similarity sequence. This multi-dimensional evaluation method can more comprehensively reflect the similarity relationship between the fault description text and each cluster, thereby improving the accuracy of cluster assignment. By introducing the variance threshold and the matching score threshold, the process can adaptively handle data distributions of different complexities. When the cosine similarity variance is small, the clusters are further distinguished by comparing the sequence similarity of the similarity sequence; when the cosine similarity variance is large, the cluster with the highest cosine similarity is directly selected as the target. When the cosine similarity variance is small, by calculating the sequence similarity of the similarity sequence, the similarity of the current fault description text and the sample fault case in similarity distribution can be more carefully compared, thereby reducing the misclassification caused by the similarity of a single similarity value. By setting the matching score threshold, only when the matching score of a cluster is significantly higher than that of other clusters, it is determined as the target case cluster. This mechanism helps to improve the robustness and stability of clustering and prevents fluctuations in results due to minor differences.
[0117] Step 170 , based on the embedded representation of the current fault description text, searching for a target fault case from the target case cluster.
[0118] Specifically include:
[0119] Based on the embedding representation of the current fault description text, calculate the similarity between the embedding representation of the current fault description text and each fault case node included in the target case clustering cluster;
[0120] Based on the similarity between the embedding representation of the current fault description text and each fault case node included in the target case clustering cluster, the target fault case is found from the target case clustering cluster.
[0121] For example, the similarity between the embedding representation of the current fault description text and each fault case node included in the target case cluster can be calculated based on the following formula:
[0122]
[0123] Among them, Z x is the embedding representation of the fault case node x included in the target case cluster, Similarity(Z y , Z x ) is the similarity between the embedding representation of the current fault description text and each fault case node included in the target case clustering cluster.
[0124] As an example only, the similarity threshold is set to 60%, and the fault case nodes exceeding the threshold are selected as target fault cases for recommendation. The decoder layer of GA-NNF is used to decode the target fault case, extract information including product model, product usage time, fault cause, fault time, fault phenomenon, fault cause and solution, and output it.
[0125] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.
Claims
1. A fault case clustering method based on a multi-feature fusion graph autoencoder, characterized in that: include: Get multiple failure cases; Based on the multiple fault cases, establish a fault case graph structure; Determine multiple case clusters and the cluster center of each case cluster based on the fault case graph structure through a case clustering model, wherein the case clustering model at least includes an encoder, a decoder and a self-training clustering unit; Get the current fault description text; Determining an embedded representation of the current fault description text; Determine a target case cluster based on the embedded representation of the current fault description text and the cluster center of each case cluster; Based on the embedded representation of the current fault description text, a target fault case is searched from the target case cluster.
2. The method for clustering fault cases of a graph autoencoder based on multi-feature fusion according to claim 1 is characterized in that: The method of determining a plurality of case clusters and a cluster center of each case cluster based on the fault case graph structure by using a case clustering model includes: Obtaining an embedding representation of each fault case node through the encoder; generating, by the decoder based on the embedded representation of each of the fault case nodes, a reconstructed adjacency matrix between any two fault case nodes; The self-training clustering unit determines a plurality of case clusters and a cluster center of each case cluster based on the embedded representation of each fault case node and the reconstructed adjacency matrix between two fault case nodes.
3. The method for clustering fault cases of a graph autoencoder based on multi-feature fusion according to claim 2 is characterized in that: Through the encoder, an embedded representation of each fault case node is obtained, including: For each of the fault case nodes, aggregate feature information of the second-order neighbor nodes of the fault case node to generate an embedded representation of the first-order neighbor nodes of the fault case node, aggregate the embedded features of the first-order neighbor nodes to generate a low-dimensional embedded representation of the fault case node; Learning the cluster distribution of the fault case nodes based on the low-dimensional embedding representation of the fault case nodes, and determining the cluster information vector of the fault case nodes through an attention layer; Based on the low-dimensional embedding representation of the fault case node and the cluster information vector, an embedding representation of the fault case node is generated.
4. The method for clustering fault cases of a graph autoencoder based on multi-feature fusion according to claim 3 is characterized in that: The cluster information vector of the fault case node is determined based on the following formula: in, is the cluster information vector learned for the i-th fault case node, α ij is the attention coefficient of the pooling vector between the i-th fault case node and the j-th cluster, W is the weight matrix, is the cluster center of the jth cluster, k is the number of clusters, δ is the nonlinear activation function, is the weight vector of the attention mechanism, is, ∥ is the concatenation operation of vectors.
5. The method for clustering fault cases of a graphics autoencoder based on multi-feature fusion according to any one of claims 1 to 4, characterized in that: The target total loss function of the self-training clustering unit is: L=L G +αL KL Among them, L is the target total loss, L G To express the loss, L KL is the clustering loss, and α is the coefficient used to balance the balance between representation loss and clustering loss.
6. The method for clustering fault cases of a graph autoencoder based on multi-feature fusion according to claim 5 is characterized in that: The loss is expressed as: Where n is the total number of fault case nodes, a ij is the element in the i-th row and j-th column of the original adjacency matrix, is the element in the i-th row and j-th column of the reconstructed adjacency matrix.
7. The method for clustering fault cases of a graph autoencoder based on multi-feature fusion according to claim 5, characterized in that: The clustering loss is: Where KL(·) is the KL divergence, Q is the membership distribution, P is the target distribution, and Z i is the embedding representation of the i-th fault case node, m j is the jth cluster center, m j′ is the j′th cluster center, q ij is the similarity between the embedding representation of the i-th fault case node and the j-th cluster center, p ij is the target distribution.
8. The method for clustering fault cases of a graphics autoencoder based on multi-feature fusion according to any one of claims 1 to 4, characterized in that: Determining the cluster center of the case cluster, including: The weighted average of the embedded representations of all fault case nodes in the case cluster is calculated to determine the cluster center of the case cluster.
9. The method for clustering fault cases of a graphics autoencoder based on multi-feature fusion according to any one of claims 1 to 4, characterized in that: Determining a target case cluster based on the embedded representation of the current fault description text and the cluster center of each case cluster, including: Calculate the cosine similarity between the embedding representation of the current fault description text and the cluster center of each of the case clustering clusters; The target case cluster is determined based on the cosine similarity between the embedding representation of the current fault description text and the cluster center of each of the case clusters.
10. The method for clustering fault cases of a graphics autoencoder based on multi-feature fusion according to any one of claims 1 to 4, characterized in that: Based on the embedded representation of the current fault description text, searching for a target fault case from the target case clustering cluster includes: Based on the embedded representation of the current fault description text, calculating the similarity between the embedded representation of the current fault description text and each fault case node included in the target case cluster; Based on the similarity between the embedding representation of the current fault description text and each fault case node included in the target case cluster, a target fault case is searched from the target case cluster.