Method and system for diagnosing failures of industrial equipment based on integrated optimization of graph structures

Abstract

This invention provides a fault diagnosis method for industrial equipment, a fault diagnosis system for industrial equipment, a storage medium, and terminal equipment that optimize problems such as low prediction accuracy and poor robustness in fault diagnosis, thereby significantly improving the stability of industrial internet equipment. [Solution] A fault diagnosis method for industrial equipment based on integrated optimization of a graph structure includes the steps of: obtaining an original equipment dataset; constructing an original graph structure based on the original equipment dataset; extracting two base views based on the original graph structure; calculating graph node embeddings of the base views using a GCN; recalculating edge probabilities in the graph structure based on the graph node embeddings; performing view fusion based on edge probabilities in the graph structure to obtain a view after initial optimization; and processing the fused view with a graph attention network to obtain an enhanced view.

Description

[Technical Field]

[0001] The present invention relates to the field of fault diagnosis technology, and more particularly to a fault diagnosis method and system for industrial equipment based on integrated optimization of graph structures. [Background technology]

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

[0003] Currently, conventional fault diagnosis methods, such as rule-based methods, are insufficient in handling the complex modes and variations of equipment failures, and statistical models often lack comprehensiveness in capturing the deep operating states and failure modes of equipment. Furthermore, conventional diagnostic systems face performance bottlenecks in terms of accuracy, real-time capabilities, and large-scale data processing.

[0004] Graph neural networks (GNNs), as an advanced technique for processing graph-structured data, represent a potential capability in equipment fault diagnosis. However, when directly applying GNNs to fault diagnosis, significant problems exist, including sparse graph structures, data noise, and computational resource constraints. Due to the sparsity of the graph structure, GNNs have difficulty capturing the latent relationships between equipment and faults, while data noise can cause the model to learn inaccurate features. How to effectively utilize GNNs for graph data modeling to overcome these problems and improve the effectiveness of fault diagnosis systems is a current focus of research and application.

[0005] Traditional GNN algorithms typically train by directly inputting the original graph structure into the network, but this presents two main problems. On the one hand, extracting and fusing a valid graph structure from the original data is a challenge. Graph data after processing real-world data commonly contains problems such as noise and incomplete node features, making it complex to construct an accurate graph structure and extract useful features from it. On the other hand, GNN algorithms are based on a message-passing mechanism, which requires the model to constantly superimpose information from the multi-hop neighborhood of nodes during training. This process requires processing a large amount of node data, resulting in slow training and inference speeds for GNN models. Furthermore, because GNN models are generally considered black boxes, their internal feature transformation and decision-making processes lack transparency. Therefore, improvements in efficiency, interpretability, and model complexity are necessary to solve these problems in practical applications of GNNs. [Overview of the project] [Problems that the invention aims to solve]

[0006] To solve the problems mentioned above, the present invention provides a fault diagnosis method and system for industrial equipment based on integrated optimization of graph structures. [Means for solving the problem]

[0007] In the first aspect, the fault diagnosis method for industrial equipment based on the integrated optimization of graph structures provided by the present invention employs the following technical proposal. A fault diagnosis method for industrial equipment based on integrated optimization of graph structures is: Steps to obtain the original equipment dataset, Steps include constructing the original graph structure based on the original equipment dataset, Extract two base views based on the original graph structure, calculate the graph node embeddings of the base views using GCN, and recalculate the probabilities of the edges in the graph structure based on the graph node embeddings; Perform view fusion based on the probabilities of the edges in the graph structure to obtain a preliminarily optimized view; Process the fused view by a GAT network to obtain an enhanced view; Adopt a probabilistic block model as a generative model, and iteratively optimize the enhanced view using Bayesian inference and the expectation-maximization algorithm to reconstruct the graph structure.

[0008] Furthermore, the step of constructing the original graph structure based on the original device dataset includes creating a graph dataset for applying the collected original device data to network training, using devices as nodes, the connection forms between devices as edges, and various data and attributes of the devices as node features.

[0009] Furthermore, the step of extracting two base views based on the original graph structure and calculating the graph node embeddings of the base views using GCN includes extracting two base views, namely the adjacency matrix and the transfer matrix, from the original graph structure, performing preliminary processing on the selected two base views, and obtaining view embeddings using the graph convolutional network GCN. Q<> 1 =σ(GCN(V1,X)) Q<> 2 =σ(GCN(V2,X)) Here, σ is a non-linear activation function, which introduces non-linear transformation and improves the network representation ability and the ability to learn complex data.

[0010] Furthermore, the step of recomputing the probability of an edge in the graph structure based on the graph node embedding includes, for a target node, connecting it to other node embeddings, subsequently normalizing the weights of the nodes to obtain the probability of the existence of an edge between node pairs, combining it with the overall probability to obtain a probability matrix, combining the probability matrix with the original graph structure, and obtaining a processed base view.

[0011] Furthermore, the step of performing view fusion based on the probability of an edge in the graph structure to obtain a preliminarily optimized view includes processing the fused graph by a graph attention network GAT, using a self-attention mechanism to enhance the interaction and feature representation between nodes, dynamically evaluating and weighting the features of neighboring nodes and aggregating them, thereby generating an enhanced view with more information.

[0012] Furthermore, the step of using a self-attention mechanism to enhance the interaction and feature representation between nodes includes calculating the attention coefficient between node pairs. e ij =d([Mh i ||Mh j ),j∈N i Here, M is a shared parameter, h represents node features, and d is a real-valued parameter for projecting the connected high-dimensional features into the real number field.

[0013] Furthermore, the step of adopting a stochastic block model as the generative model includes using a stochastic block model SBM as the generative model and simulating the optimal graph structure. The process of generating the simulated optimal graph structure G is formalized in probability as follows. JPEG2026062570000002.jpg17128JPEG2026062570000003.jpg144 Here, SBM divides the entire network into multiple communities, and c i represents the community to which node i belongs, Ω is a parameter of SBM, and Ω cicj is an arbitrary node v in community ci represents the probability that there is an edge between node v in community c j and Y, L represents tag information, G ij represents the probability that there is an edge between node i and node j in the simulation graph structure G, Z is the prediction obtained after performing softmax processing on the obtained node with respect to the node represented by H, and z i represents the prediction result of node v i and y i represents v i 's true tag, and v L represents all nodes in the training set. If there is a tagged value for node v, then the value of c i is the tagged value y i , otherwise the value is the predicted value z i . i

[0014] In the second aspect, an industrial equipment fault diagnosis system based on integrated optimization of a graph structure is configured with a data acquisition module that acquires an original equipment dataset and constructs an original graph structure based on the original equipment dataset, configured with an embedding module that extracts two base views based on the original graph structure, calculates graph node embeddings of the base views using a GCN, and recalculates the probability of edges in the graph structure based on the graph node embeddings, configured with a fusion module that performs view fusion based on the probability of edges in the graph structure to obtain a preliminarily optimized view, configured with an enhancement module that processes the fused view by a GAT network to obtain an enhanced view, and includes an optimization module that adopts a probabilistic block model as a generative model, iteratively optimizes the enhanced view using Bayesian estimation and the expectation maximization algorithm, reconstructs the graph structure, and obtains a final graph structure.

[0015] ​In a third aspect, the present invention provides a computer-readable storage medium in which a plurality of instructions are stored, the instructions being loaded by a processor of a terminal device and suitable for executing a fault diagnosis method for industrial equipment based on the integrated optimization of the graph structure.

[0016] In a fourth aspect, the present invention provides a terminal device comprising a processor used to implement each instruction and a computer-readable storage medium used to store a plurality of instructions, wherein the instructions are loaded by the processor and the terminal device is suitable for performing a fault diagnosis method for industrial equipment based on the integrated optimization of the graph structure. [Effects of the Invention]

[0017] In short, the present invention has the following beneficial effects. (1) In this invention, artificial intelligence and deep learning algorithms are introduced, and a recommendation system and a graph neural network are combined to perform fault diagnosis on industrial internet equipment using a series of technologies such as graph structure learning algorithms, base view processing, graph attention mechanisms, and inverse estimation of structure by probability. (2) The present invention provides an advanced fault diagnosis method for industrial internet equipment, which combines base view fusion processing and an attention mechanism, and also combines Bayesian inverse estimation of structure, thereby optimizing problems such as low prediction accuracy and poor robustness in conventional fault diagnosis, and thereby significantly improving the stability of industrial internet equipment. (3) To solve the problem of low goodness of fit between the dataset and the training model, a base view is extracted based on the source data, a preliminary view is obtained by performing coding, reconstruction, and fusion using GCN, quadratic optimization is performed using a GAT network to enhance node information representation power, and then Bayesian principles are introduced to generate the optimal graph frame based on SBM, the final graph structure is reconstructed using node mapping, and the accuracy of downstream tasks is improved. [Brief explanation of the drawing]

[0018] [Figure 1] This is a schematic diagram of a fault diagnosis method for industrial equipment based on integrated optimization of a graph structure according to Embodiment 1 of the present invention. [Modes for carrying out the invention]

[0019] The present invention will be described in more detail below in conjunction with the drawings. (Example 1)

[0020] Referring to Figure 1, the industrial equipment failure diagnosis method based on the integrated optimization of the graph structure according to this embodiment is: The collected original device dataset is transformed into a graph dataset for network training, with devices (machines, sensors, control systems, etc.) as nodes, the communication patterns between devices as edges, and various device data and attributes as node features (S1). In the initial stage, S2 extracts and preprocesses two base views based on the original graph structure, computes graph node embeddings using GCN, and recalculates edge connection probabilities based on node embeddings. S3 uses a GCN network to obtain prediction results for the base view, integrates the information using a view fusion algorithm, and generates a view after initial optimization. S4 processes the fused graph using a graph attention network (GAT), enhances the interactions and feature representations between nodes using a self-attention mechanism, dynamically evaluates and weights the features of neighboring nodes, and thereby generates an enhanced view with more information. This includes S5, which constructs a multi-layered set of observations, uses a probabilistic block model (SBM) as a generative model, iteratively optimizes the graph structure using Bayesian estimation and expectation maximization (EM) algorithms, and ultimately generates a high-quality graph structure, improving the performance and robustness of the graph neural network in downstream tasks.

[0021] The aforementioned step S1 specifically includes S1.1.

[0022] In S1.1, the acquired device data is preprocessed, and a network structure graph G=(V,A,X,Y) is established based on the preprocessed dataset, where V is the node set and consists of basic device information, A is the node adjacency matrix, X is the node feature set and consists of device attribute information, and Y is the device tag.

[0023] The aforementioned step S2 specifically includes S2.1 to S2.3.

[0024] In S2.1, two base views, the adjacency matrix and the transfer matrix, are extracted from the original graph structure. Initial processing is performed on the two selected base views, and view embeddings are obtained using the graph convolutional network GCN. Q 1 =σ(GCN(V1,X)) Q 2 =σ(GCN(V2,X)) Here, σ is a nonlinear activation function, introducing a nonlinear transformation to improve the network's representation capabilities and its ability to learn complex data.

[0025] In S2.2, the probability of edges in the graph structure is recalculated using node embeddings. For the target node, it is connected to other node embeddings, and then the node weights are normalized to obtain the connection probability between node pairs. This process can be represented as follows: W ij =W·[q i ||q j ]+a JPEG2026062570000004.jpg1752 Here, W represents the connection weight between nodes i and j, q represents the node weight, and a is the offset vector.

[0026] In S2.3, a probability matrix p1 is obtained based on the probability of an edge appearing in each node pair. This matrix is ​​then combined with the original structure to obtain the processed base view. This process can be represented as follows: JPEG2026062570000005.jpg841 Here, λ 1 This is the fusion coefficient, and this value is related to the type of dataset.

[0027] The aforementioned step S3 specifically includes S3.1 to S3.3.

[0028] In S3.1, views are merged using a method that assigns view weights based on prediction confidence, and first, prediction results for the two base views are obtained using a graph convolutional network (GCN). JPEG2026062570000006.jpg13128JPEG2026062570000007.jpg12128Here, σ is the nonlinear activation function.

[0029] The weights of a view reflect the uncertainty based on the predictive structure distribution. Specifically, when the maximum values ​​are the same, a larger difference between the maximum value and other values ​​indicates less uncertainty, and therefore a larger difference results in a higher view weight. For example, between [0.7,0.1,0.2] and [0.4,0.35,0.3], the former has a larger difference from the maximum value and should be given a higher weight. Similarly, when the maximum values ​​are the same, a view with a larger difference 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 higher weight.

[0030] In S3.2, based on the above principle, the weights of the views represented by H in the fusion process are calculated as follows: H vi =e β (γlogn 1 +(1-γ)logn 1 -n 2 ) Here, n1 and n 2 β and γ represent the largest and second largest predictive distributions, and β and γ are hyperparameters.

[0031] In S3.3, the weights are normalized. Based on the calculated weights, the final view of the nodes is generated. JPEG2026062570000009.jpg853

[0032] The aforementioned step S4 specifically includes S4.1 to S4.2.

[0033] In S4.1, the attention coefficient between node pairs is calculated. e ij =d([Mh i ||Mh j ]), j∈N i Here, M is a shared parameter, h represents a node feature, and d is a real parameter for projecting the connected high-dimensional features into a real field. The result requires normalization, and LeakyReLU is selected as the activation parameter. JPEG2026062570000010.jpg1885 Here, k represents all neighboring nodes of a given node.

[0034] In S4.2, the attention coefficient of each node pair is multiplied by the original node feature, and the weights are added to obtain a new node feature. The above process is performed on each node of the three views (two base views and one fused view) to self-adaptively update the node features in the graph structure, enabling the node features to aggregate information hidden in the relationships between nodes.

[0035] The aforementioned step S5 specifically includes S5.1 to S5.4.

[0036] In S5.1, an observation set is constructed using a GCN network structure, and the aggregation rule for the k-th layer of the GCN is as follows: JPEG2026062570000012.jpg1284JPEG2026062570000013.jpg627 Represents the normalized adjacency matrix, where D is a diagonal matrix, σ is a nonlinear activation function, and H (k-1) and H (k) represents the node features of the (k-1)th and k-th layers, and W(k) is the weight matrix for layer k, which facilitates linear transformations of the nodes.

[0037] After aggregation each time, the current node represents capturing the structural information in this order, and the extraction node represents constructing a kNN subgraph, into which the original image is placed, forming the following multi-layered set of observations. JPEG2026062570000014.jpg856

[0038] In S5.2, the process of generating the optimal graph structure G by simulating a generative model using a probabilistic block model (SBM) is expressed in probabilistic form as follows: JPEG2026062570000015.jpg17128JPEG2026062570000016.jpg1449 Here, Ω is a parameter of SBM, Ω cicj represents the probability that an edge exists between any node v in community c and node vj in community cj, the variable Z represents the node-specific feature set in the original dataset, and YL corresponds to the node tag set.

[0039] In S5.3, the simulation-optimal graph structure generated by SBM is mapped to the observation set. The rule compares the observed graph with the high-quality graph. If both graphs have one edge, the probability of this situation occurring is called the true high-quality rate p. If one edge in the observed graph does not exist in the high-quality graph, it is called the false high-quality rate q. Conversely, the true error rate is 1-p, and the false error rate is 1-q.

[0040] According to this rule, the probability that a high-quality graph G is mapped to an observed graph O is expressed as follows: JPEG2026062570000017.jpg52151

[0041] In S5.4, each mapping between the SBM graph and the set of observations generates different possible values ​​for the high-quality graph. Adding these values ​​together yields the following parameterized posterior probabilities: JPEG2026062570000018.jpg13128 By maximizing the posterior probability, the adjacency matrix Q of the high-quality graph is calculated, and this matrix is ​​then used to construct the determined high-quality graph as follows: The structure shown in JPEG2026062570000019.jpg1347 is input to a downstream task to obtain the equipment failure detection result. (Example 2)

[0042] This embodiment provides an industrial equipment failure diagnosis system based on integrated optimization of graph structures, and the system is A data acquisition module configured to retrieve the original equipment dataset and build the original graph structure based on the original equipment dataset, An embedding module configured to extract two base views based on the original graph structure, compute the graph node embeddings of the base views using GCN, and recalculate the edge probabilities in the graph structure based on the graph node embeddings, A fusion module configured to perform view fusion based on edge probabilities in a graph structure and obtain a view after initial optimization, An enhancement module configured to process a fused view using a GAT network and obtain an enhanced view, It includes an optimization module configured to adopt a probabilistic block model as a generative model, iteratively optimize the reinforced view using Bayesian estimation and expectation maximization algorithms, and obtain the final graph structure.

[0043] A computer-readable storage medium storing multiple instructions, wherein the instructions are loaded by a processor of a terminal device and are suitable for executing a fault diagnosis method for industrial equipment based on the integrated optimization of the graph structure.

[0044] A terminal device comprising a processor used to implement each instruction and a computer-readable storage medium used to store multiple instructions, wherein the instructions are loaded by the processor and the device is suitable for performing a fault diagnosis method for industrial equipment based on the integrated optimization of the graph structure.

[0045] The above are all preferred embodiments of the present invention and do not limit the scope of protection of the present invention. Therefore, equivalent modifications based on the structure, shape, and principle of the present invention should all be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing failures of industrial equipment based on the integrated optimization of graph structures, Steps to obtain the original equipment dataset, Steps include constructing the original graph structure based on the original equipment dataset, The process involves extracting two base views based on the original graph structure, calculating graph node embeddings for the base views using GCN, and recalculating edge probabilities in the graph structure based on the graph node embeddings. The first step is to perform view fusion based on the probability of edges in the graph structure to obtain a view after initial optimization, and the second step is to perform view fusion based on the probability of edges in the graph structure. The steps include processing the fused view using a GAT network to obtain an enhanced view, This includes the steps of employing a probabilistic block model as a generative model, iteratively optimizing the reinforcement view using Bayesian estimation and an expectation maximization algorithm to obtain the final graph structure, The step of constructing the original graph structure based on the original device dataset includes creating a graph dataset for applying the collected original device data to network training, with devices as nodes, the communication patterns between devices as edges, and various device data and attributes as node features. The step of extracting two base views based on the original graph structure and calculating graph node embeddings of the base views using a GCN includes extracting two base views, an adjacency matrix and a transfer matrix, from the original graph structure, performing preliminary processing on the two selected base views, and obtaining view embeddings using a graph convolutional network (GCN). Q １ =σ(GCN(V １ ,X)) Q ２ =σ(GCN(V ２ ,X)) Here, σ is a nonlinear activation function, introducing a nonlinear transformation to improve the network's representation capabilities and its ability to learn complex data. The step of recalculating the probability of edges in the graph structure based on graph node embeddings includes, for a target node, connecting it to other node embeddings, then normalizing the node weights to calculate the probability that an edge exists in a node pair, combining it with the overall probability to obtain a probability matrix, combining the probability matrix with the original graph structure to obtain a processed base view, The step of performing view fusion based on edge probabilities in the graph structure to obtain a preliminary optimized view includes processing the fused graph with a graph attention network (GAT), enhancing the interactions and feature representations between nodes using a self-attention mechanism, dynamically evaluating and weighting the features of neighboring nodes, and thereby generating an enhanced view with more information. The step of enhancing internode interactions and feature representations using a self-attention mechanism includes calculating attention coefficients between node pairs, s ｉｊ J(N) ｉ ||() ｊ N),E∈9 ｉ Here, M is a shared parameter, h represents a node feature, and d is a real parameter for projecting the connected high-dimensional features onto a real field. The step of adopting a probabilistic block model to create a generative model includes using a probabilistic block model (SBM) to create a generative model and simulating the optimal graph structure. The process of generating the simulated optimal graph structure G is expressed in probabilistic form as follows: Here, Ω is a parameter of the SBM, and Ω ｃｉｃｊ represents the probability that an edge exists between an arbitrary node v in community c and a node vj in community cj. The variable Z represents the set of node-specific features in the original dataset, and YL is characterized by corresponding to the node tag set. A fault diagnosis method for industrial equipment based on the integrated optimization of a graph structure

2. A fault diagnosis system for industrial equipment based on integrated optimization of a graph structure, which performs the method described in claim 1, A data acquisition module configured to retrieve the original equipment dataset and build the original graph structure based on the original equipment dataset, An embedding module configured to extract two base views based on the original graph structure, compute graph node embeddings of the base views using GCN, and recalculate edge probabilities in the graph structure based on the graph node embeddings, A fusion module configured to perform view fusion based on edge probabilities in a graph structure and obtain a view after initial optimization, An enhancement module configured to process a fused view using a GAT network and obtain an enhanced view, A fault diagnosis system for industrial equipment based on integrated optimization of graph structures, comprising: an optimization module configured to adopt a probabilistic block model as a generative model, iteratively optimize the reinforced view using Bayesian estimation and an expectation maximization algorithm to obtain a final graph structure; and

3. A computer-readable storage medium having multiple instructions stored therein, wherein the instructions are loaded by a processor of a terminal device and are suitable for performing the method described in claim 1.

4. A terminal device comprising a processor used to implement each instruction and a computer-readable storage medium used to store a plurality of instructions, wherein the instructions are loaded by the processor and the terminal device is suitable for performing the method described in claim 1.

Images