A method and system for industrial equipment fault diagnosis based on graph structure joint optimization

Abstract

The present invention relates to the field of fault diagnosis technology, and in particular to an industrial equipment fault diagnosis method and system based on graph structure joint optimization. The method includes obtaining an original equipment data set; constructing an original graph structure based on the original equipment data set; extracting two basic views based on the original graph structure, using GCN to calculate the graph node embedding of the basic view, and recalculating the probability of the edge in the graph structure based on the graph node embedding; performing view fusion based on the probability of the edge in the graph structure to obtain a preliminary optimized view; processing the fused view through the GAT network to obtain an enhanced view; the present invention optimizes the problems of low prediction accuracy and poor robustness encountered in traditional fault diagnosis, thereby significantly improving the stability of industrial Internet equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to an industrial equipment fault diagnosis method and system based on graph structure joint optimization. Background Art

[0002] With the rapid development of industrial Internet equipment fault diagnosis, the system has become an important tool for equipment maintenance and management. By analyzing the historical data, operating status and fault records of the equipment, the fault diagnosis system can provide operators with timely warnings and repair suggestions, thereby improving the stability and efficiency of equipment operation.

[0003] At present, traditional fault diagnosis methods, such as rule-based methods, perform poorly when dealing with complex patterns and changes in equipment failures, while statistical models are often not comprehensive enough in capturing the deep operating status and failure modes of equipment. In addition, existing diagnostic systems face performance bottlenecks in terms of accuracy, real-time performance, and processing large-scale data.

[0004] As an advanced technology for processing graph-structured data, graph neural networks (GNNs) have demonstrated their potential in equipment fault diagnosis. However, when GNNs are directly applied to fault diagnosis, there are significant problems such as sparse graph structures, data noise, and computing resource limitations. The sparsity of the graph structure makes it difficult for GNNs to capture the potential relationship between equipment and faults, while data noise may cause the model to learn inaccurate features. In order to improve the effectiveness of the fault diagnosis system, how to effectively use GNNs for graph data modeling and solve these problems has become the focus of current research and application.

[0005] Existing GNN algorithms usually directly input the original graph structure into the network for training, but there are two main problems: on the one hand, it is challenging to extract effective graph structures from raw data and fuse them. Graph data after real-world data processing usually has problems such as noise and incomplete node features, which makes it complicated to build accurate graph structures and extract useful features from them; on the other hand, GNN algorithms are based on message passing mechanisms, which requires the model to continuously aggregate information from the multi-hop neighborhood of nodes during training. This process requires processing a large amount of node data, resulting in slow GNN models during training and inference, and because GNN models are usually regarded as black boxes, their internal feature conversion and decision-making processes lack transparency. Therefore, in order to solve these problems of GNN in practical applications, improvements are needed in terms of efficiency, interpretability, and model complexity. Summary of the invention

[0006] In order to solve the above-mentioned problems, the present invention provides an industrial equipment fault diagnosis method and system based on graph structure joint optimization.

[0007] In a first aspect, the present invention provides an industrial equipment fault diagnosis method based on graph structure joint optimization, which adopts the following technical solution:

[0008] A method for industrial equipment fault diagnosis based on graph structure joint optimization, comprising:

[0009] Get the original device data set;

[0010] Build the original graph structure based on the original device data set;

[0011] Extract two basic views based on the original graph structure, use GCN to calculate the graph node embedding of the basic views, and recalculate the probability of edges in the graph structure based on the graph node embedding;

[0012] Perform view fusion based on the probability of edges in the graph structure to obtain a preliminary optimized view;

[0013] The fused view is processed by the GAT network to obtain an enhanced view;

[0014] The random block model is used as the generative model, and the Bayesian inference and expectation maximization algorithms are used to iteratively optimize the enhanced views and reconstruct the graph structure.

[0015] Furthermore, the method of constructing the original graph structure based on the original device data set includes making the collected original device data into a graph data set suitable for network training, with devices as nodes, connections between devices as edges, and various data and attributes of the devices as node features.

[0016] Furthermore, the method extracts two basic views based on the original graph structure and uses GCN to calculate the graph node embedding of the basic views, including extracting two basic views from the original graph structure: an adjacency matrix and a transfer matrix, performing preliminary processing on the two selected basic views, and using the graph convolutional network GCN to obtain view embedding:

[0017]

[0018]

[0019] in, It is a nonlinear activation function that introduces nonlinear transformation to improve the network's ability to express and learn complex data.

[0020] Furthermore, the probability of edges in the graph structure is recalculated based on graph node embeddings, including, for a target node, connecting its embedding with the embedding of another node, and then normalizing the weights of the nodes to obtain the probability of edge existence between node pairs, combining the overall probabilities to obtain a probability matrix, and combining the probability matrix with the original graph structure to obtain a processed basic view.

[0021] Furthermore, the view fusion is performed based on the probability of edges in the graph structure to obtain a preliminary optimized view, including processing the fused view through the graph attention network GAT, applying the self-attention mechanism to enhance the interaction and feature expression between nodes, and dynamically evaluating and weighting the features of aggregated neighbor nodes, thereby generating a more informative enhanced view.

[0022] Furthermore, the self-attention mechanism is applied to enhance the interaction and feature expression between nodes, including calculating the attention coefficient between node pairs:

[0023]

[0024] in, is a shared parameter, represents the node feature, and d is a real parameter used to project the high-dimensional features of the connection into the real field.

[0025] Furthermore, the random block model is used as the generation model, including using the random block model SBM as the generation model to simulate the optimal graph structure, and the process probability of generating the simulated optimal graph structure 𝐺 is formalized as:

[0026]

[0027]

[0028] Among them, SBM divides the entire network into several communities. Representation Node The community you belong to, are the parameters of SBM, Represents any node in community c With nodes in community c The probability that there is an edge between them; Indicates label information;

[0029] Represented in the simulation graph structure The probability that there is an edge between node i and node j, It is the node representation obtained The prediction obtained after softmax processing, Representation Node The prediction results, express The true label of Represents all nodes in the training set. If the node Has a tag value, its The value is the label value If not, the predicted value is taken. .

[0030] In the second aspect, an industrial equipment fault diagnosis system based on graph structure joint optimization includes:

[0031] The data acquisition module is configured to acquire an original device data set and construct an original graph structure according to the original device data set;

[0032] The embedding module is configured to extract two basic views based on the original graph structure, calculate the graph node embeddings of the basic views using GCN, and recalculate the probability of edges in the graph structure based on the graph node embeddings;

[0033] The fusion module is configured to perform view fusion based on the probability of edges in the graph structure to obtain a preliminary optimized view;

[0034] The enhancement module is configured to process the fused view through the GAT network to obtain an enhanced view;

[0035] The optimization module is configured to adopt the random block model as the generation model, use the Bayesian inference and expectation maximization algorithm to iteratively optimize the enhanced view, reconstruct the graph structure, and obtain the final graph structure.

[0036] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for diagnosing industrial equipment faults based on graph structure joint optimization.

[0037] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the industrial equipment fault diagnosis method based on graph structure joint optimization.

[0038] In summary, the present invention has the following beneficial technical effects:

[0039] (1) In the present invention, artificial intelligence and deep learning algorithms are introduced, and the recommendation system is combined with the graph neural network to utilize a series of technologies such as graph structure learning algorithm, basic view processing, graph attention mechanism, and probabilistic inversion structure to realize fault diagnosis of industrial Internet equipment.

[0040] (2) The present invention proposes an advanced fault diagnosis solution for industrial Internet equipment. The system optimizes the problems of low prediction accuracy and poor robustness encountered in traditional fault diagnosis by combining basic view fusion processing with attention mechanism and Bayesian inverse structure, thereby significantly improving the stability of industrial Internet equipment.

[0041] (3) The problem of low adaptability between the dataset and the training model was solved. The basic view was extracted based on the source data, and the GCN was used to encode, reconstruct and fuse the preliminary view. The GAT network was used for secondary optimization to enhance the expressiveness of node information. The Bayesian idea was introduced to generate the optimal graph framework based on SBM, and the final graph structure was reconstructed using node mapping, which improved the accuracy of downstream tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of an industrial equipment fault diagnosis method based on graph structure joint optimization according to Example 1 of the present invention. DETAILED DESCRIPTION

[0043] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0044] Example 1

[0045] Reference Figure 1 , an industrial equipment fault diagnosis method based on graph structure joint optimization of this embodiment includes the following steps:

[0046] S1 converts the collected original device data set into a graph data set suitable for network training, with devices (machines, sensors, control systems, etc.) as nodes, the connections between devices as edges, and various data and attributes of the devices as node features.

[0047] In the initial stage, S2 extracts and preprocesses two basic views according to the original graph structure, uses GCN to calculate graph node embeddings, and recalculates edge connection probabilities based on node embeddings.

[0048] S3 uses the GCN network to obtain the prediction results of the basic view, integrates the information through the view fusion algorithm, and generates a preliminary optimized view.

[0049] S4 processes the fused view through the graph attention network (GAT), applies the self-attention mechanism to enhance the interaction and feature expression between nodes, and dynamically evaluates and weights the features of aggregated neighbor nodes to generate a more informative enhanced view.

[0050] S5 constructs a multi-level observation set, adopts the stochastic block model (SBM) as the generation model, and uses Bayesian inference and expectation maximization (EM) algorithms to iteratively optimize the graph structure, ultimately generating a high-quality graph structure to enhance the performance and robustness of graph neural networks in downstream tasks.

[0051] The step S1 specifically includes:

[0052] S1.1: Preprocess the acquired device data and establish a network structure diagram G=(V, A, X, Y) based on the preprocessed data set, where V is a node set consisting of basic information of the device, A is the adjacency matrix of the node, X is a node feature set consisting of attribute information of the device, and Y is the device label.

[0053] The step S2 specifically includes:

[0054] S2.1: Extract two basic views from the original graph structure: adjacency matrix, transfer matrix, perform preliminary processing on the selected two basic views, and use graph convolutional network (GCN) to obtain view embedding:

[0055]

[0056]

[0057] in, It is a nonlinear activation function that introduces nonlinear transformation to improve the network's ability to express and learn complex data.

[0058] S2.2: Recalculate the probability of edges in the graph using node embeddings. For a target node, concatenate its embedding with the embedding of another node: Then normalize the node weights to obtain the connection probability between node pairs. This process can be expressed as:

[0059]

[0060]

[0061] in, Representation Node and The connection weights, represents the node weight, is the bias vector.

[0062] S2.3: According to the probability of the edge appearing on each node pair, a probability matrix p1 is obtained. This matrix is ​​combined with the original structure to obtain the processed basic view. The process is expressed as follows:

[0063]

[0064] in, is the fusion coefficient, which is related to the type of data set.

[0065] The step S2 specifically includes:

[0066] S3.1: The method of assigning view weights based on prediction credibility is adopted to fuse views. First, a graph convolutional network (GCN) is used to obtain the prediction results of two basic views.

[0067]

[0068]

[0069] in, is a non-linear activation function.

[0070] The view weight reflects the uncertainty based on the predicted structural distribution. Specifically, when the maximum values ​​are the same, the larger the difference between the maximum value and other values, the smaller the uncertainty, so the view with the larger difference has a higher weight. For example, between [0.7, 0.1, 0.2] and [0.4, 0.35, 0.3], the former has a large difference from the maximum value and should be given a larger weight. Similarly, when the maximum values ​​are the same, the view with a larger gap between the maximum value and the second largest value should have a higher weight. For example, between [0.8, 0.7, 0.3] and [0.8, 0.5, 0.1], the latter should be given a larger weight.

[0071] S3.2: Based on the above principles, the view weight denoted as 𝐻 in the fusion process can be calculated as follows:

[0072]

[0073] in, and represents the largest and second largest predicted distributions, and is a hyperparameter.

[0074] S3.3: Normalized weights:

[0075]

[0076] Based on the calculated weights, generate the final view of the node:

[0077] .

[0078] The step S4 specifically includes:

[0079] S4.1: Calculate the attention coefficient between node pairs:

[0080]

[0081] in, is a shared parameter, represents the node feature, and d is a real number parameter used to project the high-dimensional features of the connection into the real field. The result needs to be normalized, and the selection As activation parameter:

[0082]

[0083] in, Represents all neighbor nodes of a node.

[0084] S4.2: Multiply the attention coefficient of each node pair by the original node feature and add the weight to obtain the new node feature:

[0085]

[0086] The above processing is performed on each node of the three views (two basic views and one fusion view), and the node features in the graph structure are adaptively updated so that the node features can aggregate the information hidden behind the relationships between nodes.

[0087] The step S5 specifically includes:

[0088] S5.1: Use the GCN network to construct the observation set. The aggregation rule of the k-th layer of GCN is:

[0089]

[0090] Among them, Ã represents the normalized adjacency matrix, D is a diagonal matrix, is a nonlinear activation function, and Represents the node features of the k-1th layer and the kth layer, W(k) is the weight matrix associated with layer k, which promotes linear transformation of nodes.

[0091] After each aggregation, the current node representation captures the structural information of this order, and the extracted node representation constructs a kNN subgraph, into which the original image is placed to form a multi-level observation set:

[0092] .

[0093] S5.2: Using the stochastic block model (SBM) as the generative model to simulate the optimal graph structure, the process of generating the simulated optimal graph structure G can be formalized probabilistically as follows:

[0094]

[0095]

[0096] in, are the parameters of SBM, represents the probability that there is an edge between any node v in community c and node vj in community cj, the variable Z represents the feature set inherent to the node in the original dataset, and YL corresponds to the node label set.

[0097] S5.3, now the simulated optimal graph structure generated by SBM is mapped to the observation set, and the rule is: compare the observed graph with the high-quality graph. If an edge exists in both graphs, then the probability of this situation occurring is called the true good rate p; if an edge in the observed graph does not exist in the high-quality graph, it is called the false good rate q; otherwise, the true error rate is 1-p and the false error rate is 1-q.

[0098] According to this rule, the probability of mapping the high-quality graph G to the observation graph O can be expressed as:

[0099]

[0100] in, represents the probability that there is an edge between nodes i and j, Represents the case where there are no edges in the graph structure.

[0101] S5.4: Each mapping between the SBM graph and the observation set produces a different possible value for the high-quality graph. Adding these values ​​together gives us the parameterized posterior probability:

[0102]

[0103] By maximizing the posterior probability, the adjacency matrix Q of the high-quality graph is calculated, which is then used to construct the deterministic high-quality graph:

[0104]

[0105] This structure is input into downstream tasks to obtain the fault detection results of the equipment.

[0106] Example 2

[0107] This embodiment provides an industrial equipment fault diagnosis system based on graph structure joint optimization, including:

[0108] The data acquisition module is configured to acquire an original device data set and construct an original graph structure according to the original device data set;

[0109] The embedding module is configured to extract two basic views based on the original graph structure, calculate the graph node embeddings of the basic views using GCN, and recalculate the probability of edges in the graph structure based on the graph node embeddings;

[0110] The fusion module is configured to perform view fusion based on the probability of edges in the graph structure to obtain a preliminary optimized view;

[0111] The enhancement module is configured to process the fused view through the GAT network to obtain an enhanced view;

[0112] The optimization module is configured to adopt the random block model as the generation model and utilize the Bayesian inference and expectation maximization algorithm to iteratively optimize the graph structure to obtain the final graph structure.

[0113] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for example, an industrial equipment fault diagnosis method based on graph structure joint optimization.

[0114] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to perform an industrial equipment fault diagnosis method based on graph structure joint optimization.

[0115] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for industrial equipment fault diagnosis based on graph structure joint optimization, characterized in that: include: Get the original device data set; Build the original graph structure based on the original device data set; Extract two basic views based on the original graph structure, use GCN to calculate the graph node embedding of the basic views, and recalculate the probability of edges in the graph structure based on the graph node embedding; Perform view fusion based on the probability of edges in the graph structure to obtain a preliminary optimized view; The fused view is processed by the GAT network to obtain an enhanced view; The random block model is used as the generation model, and the Bayesian inference and expectation maximization algorithm are used to iteratively optimize the enhanced view to obtain the final graph structure; The constructing of the original graph structure according to the original device data set includes making the collected original device data into a graph data set suitable for network training, with the devices as nodes, the connections between the devices as edges, and various data and attributes of the devices as node features; The method extracts two basic views based on the original graph structure and uses GCN to calculate the graph node embedding of the basic views, including extracting two basic views from the original graph structure: adjacency matrix, transfer matrix, performing preliminary processing on the two selected basic views, and using graph convolution network GCN to obtain view embedding: in, For nonlinear activation functions, nonlinear transformation is introduced to improve the network's ability to express and learn complex data; The recalculating the probability of edges in the graph structure based on the graph node embedding includes, for a target node, connecting its embedding with the embedding of another node, then normalizing the weights of the nodes, calculating the probability of the existence of edges for the node pairs, combining all probabilities to obtain a probability matrix, combining the probability matrix with the original graph structure, and obtaining a processed basic view; The view fusion is performed based on the probability of the edge in the graph structure to obtain a preliminary optimized view, including processing the fused view through the graph attention network GAT, applying the self-attention mechanism to enhance the interaction and feature expression between nodes, and dynamically evaluating and weighting the features of the aggregated neighbor nodes, so as to generate a more informative enhanced view; The self-attention mechanism is applied to enhance the interaction and feature expression between nodes, including calculating the attention coefficient between node pairs: in, is a shared parameter, represents the node feature, d is a real number parameter used to project the high-dimensional features of the connection into the real field; The random block model is used as the generation model, including using the random block model SBM as the generation model to simulate the optimal graph structure, and the process probability of generating the simulated optimal graph structure 𝐺 is formalized as: in, are the parameters of SBM, Denotes the probability that there is an edge between any node 𝑣 in community 𝑐 and node 𝑣𝑗 in community 𝑐𝑗, variable 𝑍 represents the feature set inherent to the node in the original dataset, and 𝑌𝐿 corresponds to the node label set.

2. An industrial equipment fault diagnosis system based on graph structure joint optimization, executing the method according to claim 1, characterized in that: include: The data acquisition module is configured to acquire an original device data set and construct an original graph structure according to the original device data set; The embedding module is configured to extract two basic views based on the original graph structure, calculate the graph node embeddings of the basic views using GCN, and recalculate the probability of edges in the graph structure based on the graph node embeddings; The fusion module is configured to perform view fusion based on the probability of edges in the graph structure to obtain a preliminary optimized view; The enhancement module is configured to process the fused view through the GAT network to obtain an enhanced view; The optimization module,is configured to adopt the random block model as the generation model,,and iteratively optimize the enhanced views using the Bayesian,inference and expectation maximization algorithms to obtain the final graph,structure.

3. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

4. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method as claimed in claim 1 .

Images