A spatio-temporal fusion-based graph convolution network industrial process fault diagnosis method

Abstract

The application relates to an industrial process fault diagnosis method, and discloses an industrial process fault diagnosis method based on a graph convolution network of space-time fusion, which comprises the following specific process: collecting fault data in an industrial production process, then performing z-score standardization processing and sliding window interception on the fault data to obtain a node feature matrix, and inputting the node feature matrix into a ST-Res-GCN model which has been trained and tested to obtain a fault classification result; the application proposes an industrial process fault diagnosis model based on a graph convolution network of space-time fusion, which can automatically acquire and deeply mine space-time fusion features in data, and effectively improves the fault diagnosis rate.

Description

Technical Field

[0001] This invention relates to a method for diagnosing industrial process faults, specifically a method for diagnosing industrial process faults based on spatiotemporal fusion graph convolutional networks. Background Technology

[0002] With the advent of Industry 4.0, manufacturing enterprises have been continuously exploring and developing technologies to upgrade their production equipment with intelligent systems. This has led to a surge in the complexity of modern industrial production systems, making process monitoring increasingly challenging. As we all know, the safety of industrial processes is paramount, because accidents can result in product defects and, in severe cases, serious casualties. Therefore, ensuring the stable operation of industrial processes has become a pressing issue, making the development of safe and efficient fault detection and diagnosis technologies particularly important.

[0003] Based on different modeling principles, fault diagnosis methods can be divided into three main categories: mechanism modeling, knowledge-driven modeling, and data-driven modeling. In the context of industrial upgrading, mechanism modeling and knowledge-driven modeling, which rely solely on experts' prior knowledge and practical experience, are increasingly insufficient for detecting and diagnosing faults in increasingly complex chemical production processes. The large amount of measured data monitored in the production process provides a prerequisite for the implementation of data-driven modeling methods. With the rapid development of artificial intelligence technology, data-driven industrial process fault diagnosis has gained significant favor among experts and scholars. Data-driven modeling methods, based on historical process data for model construction and adaptive parameter tuning, exhibit good adaptability and robustness. Commonly used data-driven modeling methods include principal component analysis, support vector machines, and artificial neural networks. While shallow data-driven modeling methods have achieved good results in fault diagnosis, their effectiveness in feature extraction from high-dimensional, nonlinear, and non-stationary data is limited. In recent years, deep learning methods have demonstrated superior performance in image classification, natural language processing, and video classification; therefore, more and more experts and scholars are applying deep learning methods to fault diagnosis in industrial processes. Deep learning is a deep machine learning method that uses stacked layers of nonlinear function mappings to uncover the relationships between data, thereby achieving a higher-level and more abstract representation of the data. Introducing deep learning into fault diagnosis can effectively overcome the limitations of previous methods in learning complex deep structures, allowing for more efficient use of data information and improving the model's predictive accuracy.

[0004] Today, industrial processes are becoming increasingly large-scale and complex, resulting in complex spatiotemporal characteristics for industrial process data. In terms of temporal sequence, different types of faults exhibit varying rates of occurrence, displaying non-stationary and nonlinear characteristics. In terms of feature space, the key features affecting different faults differ, leading to complex coupling relationships between features. Furthermore, from a spatiotemporal correlation perspective, there are static and dynamic coupling relationships between temporal and spatial features. Therefore, when modeling fault diagnosis for industrial processes with complex spatiotemporal data characteristics, it is necessary to consider not only temporal continuity and periodicity but also spatial correlations between different regions, and these spatial correlations change over time. Traditional spatiotemporal sequence research methods are primarily based on statistical principles. These methods treat spatiotemporal sequences as multiple time series, making it difficult to capture spatial correlations and extract nonlinear spatiotemporal information. The rise of deep learning has further advanced the research on spatiotemporal sequences. Existing spatiotemporal sequence algorithms based on deep learning methods are analyzed from two perspectives: sequence generation models and spatial relationship modeling. Sequence generation models are based on recurrent neural networks (RNNs), and based on RNNs, Long Short-Term Memory (LSTM) units, gated recurrent units (GRUs), and other derivative methods have been proposed. Subsequently, given the good performance of the seq2seq framework in natural language processing, sequence generation models combining RNNs and the seq2seq framework have also achieved good results. Spatial relation modeling is based on convolutional neural networks (CNNs), which can effectively extract spatial features. However, CNNs cannot effectively operate on non-Euclidean graph structures, thus Graph Neural Networks (GNNs) emerged. GNNs are a class of deep learning-based methods for processing graph information. Research on non-Euclidean graph data structures has greatly promoted the development of a series of graph analysis techniques such as Graph Convolutional Networks (GCNs).

[0005] Combining sequence models and spatial models to construct spatiotemporal models can be well applied to spatiotemporal sequence problems. Most existing methods extract spatiotemporal features only from the perspective of single graph structure data or single matrix structure data. However, industrial process data are mostly highly nonlinear, and the spatiotemporal features contained therein often cannot be fully extracted from a single perspective. Therefore, they cannot maximize or fully exploit the spatiotemporal feature information in industrial process fault diagnosis. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for fault diagnosis of industrial processes based on spatiotemporal fusion graph convolutional networks, which can be used to automatically classify fault types of fault data in industrial processes.

[0007] To address the aforementioned technical problems, this invention provides a graph convolutional network-based fault diagnosis method for industrial processes based on spatiotemporal fusion. The specific process includes: collecting fault data from the industrial production process, then performing z-score normalization on the fault data and using a sliding window to obtain the node feature matrix X. online , the node feature matrix X online The results are input into the trained and tested ST-Res-GCN model to obtain fault classification results;

[0008] The ST-Res-GCN model includes, in sequence, an input layer, a first batch normalization layer (BN Layer1), a first graph convolutional layer (GCN Layer1), three residual graph convolutional modules (Res-GCN), a fifth batch normalization layer (BN Layer5), a first one-dimensional convolutional layer (CNN Layer1), two spatiotemporal feature extraction modules (ST), a global max pooling layer (GMP Layer), a dropout layer, and an output layer.

[0009] The residual graph convolutional module Res-GCN includes a batch normalization layer (BN Layer), a graph convolutional layer (GCN Layer), and an adder layer.

[0010] The spatiotemporal feature extraction module ST includes a batch normalization layer (BN), a long short-term memory network (LSTM), a one-dimensional convolutional layer (CNN), and an additive layer (Add Layer).

[0011] As an improvement to the spatiotemporal fusion-based graph convolutional network method for industrial process fault diagnosis of the present invention:

[0012] The training and testing process of the ST-Res-GCN model is as follows:

[0013] (1) The simulation experimental data is divided into two parts: training set and test set. Each part contains normal data and fault data.

[0014] (2) Calculate the MIC values ​​between different variables in the normal data of the training set as the adjacency matrix A. After z-score standardization of the fault data in the training set, use a sliding window to extract the node feature matrix X of the training set. Then input the adjacency matrix A and the node feature matrix X of the training set into the ST-Res-GCN model for training. Use cross-entropy as the loss function, Adam algorithm as the optimization algorithm, and add an early stopping mechanism to find the best model and obtain the trained ST-Res-GCN model.

[0015] (3) After z-score standardization of the fault data in the test set, the node feature matrix X of the training set is obtained by using a sliding window. The node feature matrix X of the test set is input into the trained ST-Res-GCN model. The evaluation index is the error rate and F1 score. The parameters of the ST-Res-GCN model are adjusted until the evaluation index meets the preset evaluation criteria to obtain the trained and tested ST-Res-GCN model. The adjacency matrix A used during testing is the same as the adjacency matrix A used during training.

[0016] As a further improvement to the spatiotemporal fusion-based graph convolutional network method for industrial process fault diagnosis of the present invention:

[0017] The process of obtaining the adjacency matrix A is as follows:

[0018] (1) Data interception

[0019] For the normal data The dataset is obtained by performing z-score normalization and then selecting time-series data of length l. in, N is the number of time series data entries in dataset S;

[0020] (2) MIC value calculation

[0021] Select two time series data points s from dataset S in sequence i ,s j ∈S, calculate the MIC value between the two data points:

[0022]

[0023]

[0024] Where p(s) i ,s j ) for s i ,s j The joint distribution of p(s) i ), p(s j ) are respectively s i and s j Marginal distribution, I(s) i ,s j ) for s i ,s j Mutual information, where B is the upper limit of the grid, usually taken as B = n 0.6 n is the number of samples, and a and b are row a and column b of an a×b grid;

[0025] (3) Adjacency matrix calculation

[0026] right The maximum information coefficient (MIC) is calculated between all N time series data points according to equation (1) to obtain a symmetric N×N matrix. A threshold λ is set, and MIC values ​​less than λ in the N×N matrix are replaced with 0, while the remaining values ​​remain unchanged. The resulting MIC matrix is ​​used as the adjacency matrix.

[0027] As a further improvement to the spatiotemporal fusion-based graph convolutional network method for industrial process fault diagnosis of the present invention:

[0028] The process of capturing the sliding window is as follows:

[0029] The z-score-standardized fault data is truncated using a window length of t and a step size of 1, resulting in a dataset X = {x1, x2, ..., x...} of N×t groups, consisting of T-t+1 groups. T-t+1} represents the node feature matrix X, where T is the time series length of the fault data, i = 1, 2, ..., T-t+1.

[0030] As a further improvement to the spatiotemporal fusion-based graph convolutional network method for industrial process fault diagnosis of the present invention:

[0031] The input layer includes the node feature matrix X and the adjacency matrix A;

[0032] The residual graph convolution module Res-GCN includes a first residual graph convolution module Res-GCN 1, a second residual graph convolution module Res-GCN 2, and a third residual graph convolution module Res-GCN 3;

[0033] The spatiotemporal feature extraction module ST includes a first spatiotemporal feature extraction module ST1 and a second spatiotemporal feature extraction module ST2.

[0034] As a further improvement to the spatiotemporal fusion-based graph convolutional network method for industrial process fault diagnosis of the present invention:

[0035] The calculation process of the first convolutional layer (GCN Layer 1) is as follows:

[0036]

[0037] in, I N It is the identity matrix; yes The angle matrix, for The normalized matrix, W1 is the weight, σ is the activation function; f BN X is the output of the batch normalization layer calculation, and X is the input.

[0038] As a further improvement to the spatiotemporal fusion-based graph convolutional network method for industrial process fault diagnosis of the present invention:

[0039] The first residual graph convolutional module Res-GCN 1 includes: a second batch normalization layer (BN Layer2), a second graph convolutional layer (GCN Layer2), and a first additive layer (Add Layer1). The calculation process is as follows:

[0040] O1 = H1 + H2, (10)

[0041] Where O1 and H2 represent the outputs of the first residual graph convolutional module Res-GCN 1 and the second graph convolutional layer GCNLayer2, respectively.

[0042]

[0043] The second residual graph convolutional module Res-GCN 2 includes: a third batch normalization layer (BN Layer3), a third graph convolutional layer (GCN Layer3), and a second additive layer (Add Layer2). The calculation process is as follows:

[0044] O2=O1+H3(11)

[0045] Where O2 and H3 represent the outputs of the second residual graph convolutional module Res-GCN 2 and the third graph convolutional layer GCNLayer3, respectively.

[0046]

[0047] The third residual graph convolutional module Res-GCN3 includes: a fourth batch normalization layer (BN Layer4), a fourth graph convolutional layer (GCN Layer4), and a third additive layer (Add Layer3). The calculation process is as follows:

[0048] O3=O2+H4(12)

[0049] Where O3 and H4 represent the outputs of the third residual graph convolutional module Res-GCN3 and the fourth graph convolutional layer GCNLayer4, respectively.

[0050]

[0051] Among them, W2, W3 and W4 are all weights.

[0052] As a further improvement to the spatiotemporal fusion-based graph convolutional network method for industrial process fault diagnosis of the present invention:

[0053] The first spatiotemporal feature extraction module ST 1 includes: a sixth batch normalization layer (BN Layer 6), a first long short-term memory (LSTM) network layer 1, a seventh batch normalization layer (BN Layer 7), a second one-dimensional convolutional layer (CNN Layer 2), and a fourth additive layer (Add Layer 4). The calculation process is as follows:

[0054] STO1=R1+R2(22)

[0055] Wherein, STO1, R1 and R2 represent the output of the first spatiotemporal feature extraction module ST1, the output of the first one-dimensional convolutional layer CNNLayer1, and the output of the second one-dimensional convolutional layer CNNLayer2, respectively;

[0056] R1=f Conv (f BN (O3))(14)

[0057] R2 = ReLU(f Conv (f BN (f LSTM (f BN (R1)))))(21)

[0058] Among them, f Conv f represents the computation of a one-dimensional convolutional layer. LSTM Represents the computation of long short-term memory networks;

[0059] The second spatiotemporal feature extraction module ST 2 includes: an eighth batch normalization layer (BN Layer 8), a second long short-term memory (LSTM) network layer 2, a ninth batch normalization layer (BN Layer 9), a third one-dimensional convolutional layer (CNN Layer 3), and a fifth addition layer (Add Layer 5). The calculation process is as follows:

[0060] R3 = ReLU(f Conv (f BN (f LSTM (f BN (STO1)))))(23)

[0061] STO2=STO1+R3(24)

[0062] Wherein, STO2, STO1 and R3 represent the outputs of the second spatiotemporal feature extraction module ST2, the first spatiotemporal feature extraction module ST1 and the third one-dimensional convolutional layer CNN Layer3, respectively.

[0063] The beneficial effects of this invention are mainly reflected in:

[0064] 1. This invention proposes a fault diagnosis model for industrial processes based on spatiotemporal fusion graph convolutional networks, which can automatically extract and deeply mine spatiotemporal fusion features in data, effectively improving the fault diagnosis rate;

[0065] 2. This invention proposes a spatiotemporal feature extraction module based on a one-dimensional convolutional network and a long short-term memory network, realizing the extraction and fusion of temporal and spatial information. Furthermore, it employs a single GCN layer and three residual graph convolutional modules to extract spatiotemporal features from graph-structured data, effectively uncovering the essential features of the data. Combined with two spatiotemporal feature extraction modules consisting of a one-dimensional convolutional layer and an LSTM layer, it extracts spatiotemporal features from more angles, improving the accuracy of classification results. Simultaneously, it uses a global max pooling layer to downsample and reduce the dimensionality of feature data, and adds a dropout layer to avoid overfitting, enhancing the model's generalization ability and improving diagnostic rate and performance in practical applications.

[0066] 3. This invention uses the Maximum Information Coefficient (MIC) method to calculate the MIC value matrix between data variables as the adjacency matrix of GCN, which can solve the problem that deeper correlations between data variables cannot be obtained in conventional industrial processes. Attached Figure Description

[0067] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0068] Figure 1 This is a schematic diagram of the structure of the spatiotemporal fusion-based graph convolutional network ST-Res-GCN of the present invention;

[0069] Figure 2 for Figure 1 A schematic diagram of the Res-GCN residual graph convolution module;

[0070] Figure 3 for Figure 1 A schematic diagram of the spatiotemporal feature extraction module ST in the diagram;

[0071] Figure 4 This is a schematic diagram of a method for diagnosing industrial process faults based on spatiotemporal fusion graph convolutional networks according to the present invention. Detailed Implementation

[0072] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:

[0073] Example 1: A method for fault diagnosis in industrial processes based on spatiotemporal fusion graph convolutional networks, such as... Figure 4 As shown, the specific method of using this invention is as follows:

[0074] Step 1: Acquire normal and fault data of the industrial process;

[0075] In industrial production sites, the values ​​of variables such as material parameters, equipment parameters, and status parameters are collected periodically through technologies such as sensors and industrial computers during the industrial production process. This allows for the acquisition of time-series data of each variable under normal operating conditions (i.e., normal data) and time-series data of each variable under different fault conditions (i.e., fault data).

[0076] Step 2: Calculate the maximum information coefficient (MIC) among all different variables in the normal data obtained in Step 1 to obtain a matrix of MIC values ​​among all variables;

[0077] The correlation between variables in the normal data obtained in step 1 is represented based on the Maximum Information Coefficient (MIC) method. The calculation process of the Maximum Information Coefficient (MIC) method is as follows:

[0078] Step 2.1, Data Extraction

[0079] Given normal data The data is then subjected to z-score normalization to balance the scale. The formula for z-score normalization is:

[0080]

[0081] Where μ is the mean of all sample data and σ is the standard deviation of all sample data.

[0082] Then, time series data of length l are selected from the standardized dataset to reduce computational complexity, resulting in the dataset. Where S = {s1, s2, ..., s} i ,...,s N}, N represents the number of time series data entries in the dataset. The length l can be adjusted according to different datasets. Choose the length that best suits the dataset. For example, when using the TE dataset, 35 is the most suitable.

[0083] Step 2.2, MIC value calculation

[0084] Select two time series data points s from dataset S in sequence i ,s j ∈S, calculate the MIC value between two data points using the following formula:

[0085]

[0086]

[0087] Where p(s) i ,sj ) for s i ,s j The joint distribution of p(s) i ), p(s j ) are respectively s i and s j Marginal distribution, I(s) i ,s j ) for s i ,s j Mutual information, where B is the upper limit of the grid, usually taken as B = n 0.6 , where n is the number of samples.

[0088] Two time series data (s) i ,s j The values ​​of ) are sequentially divided into a and b cells, resulting in an a×b grid of row a and column b, such that the values ​​fall into the (s)th cell. i ,s j The frequency of data points in each of the ) cells is used as p(s) i ,s j Estimate (s), then calculate (s) i ,s j The mutual information of the two time series data points is determined by taking the maximum mutual information among different partitioning methods. i ,s j The maximum information coefficient.

[0089] Step 2.3, Adjacency Matrix Calculation

[0090] For all N time-series data obtained in step 2.1, the maximum information coefficient (MIC) is calculated in step 2.2 to obtain a symmetric N×N matrix. Based on this, to reduce potential interference from low-correlation variables, a threshold λ is manually set, replacing MIC values ​​less than 1 in the N×N matrix with 0, while leaving the remaining values ​​unchanged. In summary, the MIC matrix obtained through the above steps serves as the adjacency matrix for subsequent graph convolutional network layers. use.

[0091] Step 3: Standardize the fault data obtained in Step 1, and then use a sliding window to capture it;

[0092] Given fault data Where N is the number of feature variables, T is the time series length of the fault data, and z-score standardization is applied to it:

[0093] Data * =(Data2-μ) / σ (4)

[0094] Where μ and σ are the mean and variance of each feature variable, respectively;

[0095] Then, using a window length of t, and moving and truncating with a step size of 1, we can obtain a dataset X = {x1, x2, ..., x} consisting of N×t groups, with a window length of t and a step size of 1. T-t+1} represents the node feature matrix X, where i = 1, 2, ..., T-t+1;

[0096] Step 4: Construct a deep learning neural network ST-Res-GCN model (hereinafter referred to as the ST-Res-GCN model) based on graph convolution and spatiotemporal feature extraction modules, as follows: Figure 1 As shown, the ST-Res-GCN model is initialized;

[0097] The ST-Res-GCN model consists of the following layers in sequence: input layer, first batch normalization layer (BN) Layer1, first graph convolutional layer (GCNLayer1), first residual graph convolutional module (Res-GCN 1), second residual graph convolutional module (Res-GCN 2), third residual graph convolutional module (Res-GCN 3), fifth batch normalization layer (BN) Layer5, first one-dimensional convolutional layer (CNN) Layer1, first spatiotemporal feature extraction module (ST 1), second spatiotemporal feature extraction module (ST 2), global max pooling layer (GMP Layer), dropout layer, and output layer.

[0098] The input layer includes a node feature matrix X and an adjacency matrix A. Then, a graph convolutional layer (GCNLayer1) and three residual graph convolutional modules (Res-GCN) (Res-GCN 1, Res-GCN 2, and Res-GCN 3) are used to extract the spatiotemporal features of the graph structure data, with ReLU selected as the activation function. The features extracted by the third residual graph convolutional module (Res-GCN 3) are transposed and then convolved by a fifth batch normalization layer (BN Layer5) and a first one-dimensional convolutional layer (CNN Layer1). These features are then input into the first spatiotemporal feature extraction module (ST 1) and the second spatiotemporal feature extraction module (ST 2) for further feature extraction from different perspectives. Finally, the features are input into a global max pooling layer (GMP). The model performs downsampling and dimensionality reduction using a layer (Dense), and classification is performed using a fully connected layer (Dense). A dropout layer is added between the global max pooling layer and the fully connected layer to reduce overfitting of the model, ultimately obtaining accurate fault classification results.

[0099] Each Residual Graph Convolutional Module (Res-GCN) includes a Batch Normalization Layer (BN Layer), a Graph Convolutional Layer (GCNLayer), and an Add Layer.

[0100] (1) Calculation of graph convolutional layers

[0101] Except for the first graph convolutional layer (GCN Layer 1), the residual graph convolutional module (Res-GCN) in the ST-Res-GCN model includes graph convolutional layers (GCN Layers). Each graph convolutional layer (GCN) can be represented by a nonlinear function:

[0102] O=f(I,A) (5)

[0103] Where I and O represent the input and output of the graph convolutional layer, respectively;

[0104] Graph Convolutional Networks (GCNs) mimic the convolution operation on an image. They then add a self-loop to the adjacency matrix A to obtain a new adjacency matrix A, and normalize this new matrix A. All GCN layers use ReLU as the activation function, leading to the commonly used calculation formula for GCNs:

[0105]

[0106] Where H represents the output of the graph convolutional layer (GCN layer), This indicates adding a self-connect to the adjacency matrix A, I N It is the identity matrix; yes The diagonal matrix, with diagonal elements as All other elements are 0; for The normalized matrix, W is the weight of this layer, and X is the input;

[0107] The calculation process for the batch normalization layers (including BN Layer 1 and BN Layer 5) is as follows:

[0108]

[0109] Therefore, the calculation process of the first graph convolutional layer (GCN Layer 1) of the ST-Res-GCN model can be obtained as follows:

[0110]

[0111] The calculation process of the second convolutional layer (GCN Layer 2) is as follows:

[0112]

[0113] The calculation process of the third convolutional layer (GCN Layer 3) is as follows:

[0114]

[0115] Where O1 is the output of the first residual map convolution module Res-GCN 1;

[0116] The calculation process of the fourth convolutional layer (GCN Layer 4) is as follows:

[0117]

[0118] Where O2 is the output of the second residual map convolution module Res-GCN 2;

[0119] (2) Calculation of the residual map convolution module

[0120] The Residual Graph Convolutional Module (Res-GCN) includes: Batch Normalization Layer (BN Layer), Graph Convolutional Layer (GCNLayer), and Addition Layer, such as... Figure 2 As shown, the calculation process is as follows:

[0121] O (L) =I (L) +H (L+1) (9)

[0122] Among them, O (L) I (L) and H (L+1) These represent the output and input of the Lth residual graph convolutional module (Res-GCN) and the output of the graph convolutional layer (GCN Layer) of the (L+1)th residual graph convolutional module, respectively.

[0123] The first residual graph convolutional module Res-GCN 1 includes: a second batch normalization layer (BN Layer2), a second graph convolutional layer (GCNLayer2), and a first addition layer (Add Layer1). Therefore, the calculation process of the first residual graph convolutional module (Res-GCN 1) of the ST-Res-GCN model can be obtained as follows:

[0124] O1 = H1 + H2 (10)

[0125] Wherein, O1, H1 and H2 represent the output of the first residual graph convolution module Res-GCN 1, the output of the first graph convolution layer GCN Layer1, and the output of the second graph convolution layer GCN Layer2, respectively;

[0126] The second residual graph convolutional module, Res-GCN 2, includes: a third batch normalization layer (BN Layer3), a third graph convolutional layer (GCNLayer3), and a second additive layer (Add Layer2). Its calculation process is as follows:

[0127]

[0128] Where O2, O1 and H3 represent the outputs of the second residual graph convolution module Res-GCN 2, the first residual graph convolution module Res-GCN 1 and the third graph convolution layer GCN Layer3, respectively;

[0129] The third residual graph convolutional module, Res-GCN3, includes: the fourth batch normalization layer (BN Layer4), the fourth graph convolutional layer (GCNLayer4), and the third additive layer (Add Layer3). Its calculation process is as follows:

[0130] O3=O2+H4(12)

[0131] Where O3, O2 and H4 represent the outputs of the third residual graph convolution module Res-GCN3, the second residual graph convolution module Res-GCN2, and the fourth graph convolution layer GCN Layer4, respectively.

[0132] (3) Calculation of the spatiotemporal feature extraction module

[0133] The calculation process of the convolutional layer is as follows:

[0134]

[0135] Where x is the input data, y Conv For the output data, ω is the weight of the convolution kernel, a and β are the number of rows and columns of the convolution kernel, respectively, and the superscript i,j represents the element in the i-th row and j-th column of the data.

[0136] Therefore, the calculation process of the first one-dimensional convolutional layer (CNN Layer 1) of the ST-Res-GCN model can be obtained as follows:

[0137]

[0138] The computation process of Long Short-Term Memory (LSTM) layer f LSTM (x t )for:

[0139] Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i (15)

[0140] Forgotten Gate: f t =σ(W f ·[h t-1 ,x t ]+b f (16)

[0141] Output gate: ot =σ(W o ·[h t-1 ,x t ]+b o (17)

[0142] Candidate memory units:

[0143] Long memory:

[0144] Short memory (hidden state): h t =o t *tanh(C t (20)

[0145] W i W represents the weights of the input gate. f For the weight of the forget gate, W o h represents the weights of the output gate. t-1 Let x be the value of the short memory (hidden state) of the (t-1)th LSTM layer, and let x be the input x of the t-th LSTM layer. t b i Here is the bias of the input gate, and here is the bias of the forget gate. f b o For the bias of the output gate, W C b C These represent the weights and biases of the candidate memory units, respectively.

[0146] The network structure of the spatiotemporal feature extraction module is as follows: Figure 3 As shown, the input feature map is sequentially processed through a batch normalization layer (BN), a long short-term memory (LSTM) network, another batch normalization layer (BN), a one-dimensional convolutional layer (CNN), and a ReLU function before being added to the input feature map. Specifically, the first spatiotemporal feature extraction module ST1 includes: a sixth batch normalization layer (BN) Layer 6, a first long short-term memory (LSTM) network Layer 1, a seventh batch normalization layer (BN) Layer 7, a second one-dimensional convolutional layer (CNN) Layer 2, and a fourth addition layer (Add Layer 4). Its calculation process is as follows:

[0147] R2 = ReLU(f Conv (f BN (f LSTM (f BN (R1))))) (21

[0148] Among them, f Conv f represents the computation of a one-dimensional convolutional layer. LSTM Represents the computation of long short-term memory networks;

[0149] STO1=R1+R2 (22)

[0150] Wherein, STO1, R1 and R2 represent the output of the first spatiotemporal feature extraction module ST1, the output of the first one-dimensional convolutional layer CNNLayer1, and the output of the second one-dimensional convolutional layer CNNLayer2, respectively;

[0151] The second spatiotemporal feature extraction module ST 2 includes: the eighth batch normalization layer (BN Layer 8), the second long short-term memory network (LSTM Layer 2), the ninth batch normalization layer (BN Layer 9), the third one-dimensional convolutional layer (CNN Layer 3), and the fifth additive layer (AddLayer 5). Its calculation process is as follows:

[0152] R3 = ReLU(f Conv (f BN (f LSTM (f BN (STO1)))))(23)

[0153] STO2=STO1+R3(24)

[0154] Wherein, STO2, STO1 and R3 represent the outputs of the second spatiotemporal feature extraction module ST2, the first spatiotemporal feature extraction module ST1 and the third one-dimensional convolutional layer CNN Layer3, respectively.

[0155] In summary, using the Maximum Information Coefficient (MIC) method to calculate the MIC value matrix between data variables as the adjacency matrix of the GCN can solve the problem of not being able to obtain deeper correlations between data variables in conventional industrial processes. Furthermore, using a single GCN layer and three residual graph convolutional modules to extract spatiotemporal features from graph-structured data can effectively uncover the essential features of the data. Combining this with two spatiotemporal feature extraction modules consisting of one-dimensional convolutional layers and LSTM layers allows for the extraction of spatiotemporal features from more angles, improving the accuracy of classification results. At the same time, using a global max pooling layer to downsample and reduce the dimensionality of feature data, and adding a dropout layer to avoid model overfitting, enhances the model's generalization ability and can better improve the diagnostic rate and performance in practical applications.

[0156] Step 5: Train and test the ST-Res-GCN model

[0157] The simulation experimental data used in this invention is the Tennessee-Eastman (TE) process dataset, which is divided into a training set and a test set. Each set contains normal sample data and 21 types of fault sample data. Each sample data has 52 variables. The training set has 500 data points for each variable, and the test set has 800 data points for each variable. A 52×52 MIC matrix is ​​obtained by calculating the 52 variables in the normal sample data in step 2. For each type of fault sample data, standardization and sliding window truncation are performed in step 3, with a window width of 35, thus providing 466 sets of training data (i.e., 466 nodes). The test set yields 766 test data sets (i.e., 766 node feature matrices X). The 52×52 MIC matrix is ​​used as the adjacency matrix A in the input layer of the ST-Res-GCN model established in step 4. The 466 training data sets (i.e., 466 node feature matrices X) are used as the feature matrix X in the input layer of the ST-Res-GCN model. The ST-Res-GCN model is trained using cross-entropy as the loss function, the Adam algorithm as the optimization algorithm, and an early stopping mechanism to find the optimal model. Thus, the trained ST-Res-GCN model is obtained.

[0158] The 766 sets of test data (i.e., the feature matrix X of 766 nodes) are input into the trained ST-Res-GCN model. The adjacency matrix A used during testing is the same as that used during training, both obtained by calculating the MIC values ​​between different variables in the normal data of the training set. The ST-Res-GCN model is evaluated using the error rate and F1 score as preset metrics; a lower error rate and a higher F1 score are better. If the evaluation metrics of the ST-Res-GCN model do not meet the preset evaluation criteria, the parameters of the network structure of the ST-Res-GCN model are adjusted until the evaluation metrics meet the preset evaluation criteria, thus obtaining a trained and tested ST-Res-GCN model suitable for online industrial process fault diagnosis.

[0159] Error rate

[0160] error = 1 - accuracy(25)

[0161] F1 score

[0162]

[0163] Here, accuracy is the precision rate, precision is the accuracy rate, and recall is the recall rate.

[0164] Step 6: Using the trained ST-Res-GCN model obtained in Step 5, online industrial process fault diagnosis can be performed.

[0165] The fault data from the real-time industrial production site, collected using the same technical methods as in step 1, is then processed through standardization and sliding window capture in step 3 to obtain the node feature matrix X of the real-time industrial production site data. online X online Input the trained and tested ST-Res-GCN model from step 5 to obtain the fault classification and diagnosis results of the real-time data.

[0166] Experiment 1:

[0167] The Tennessee-Eastman (TE) process dataset was used as the simulation data for this experiment. The experimental hardware parameters are shown in Table 1 below:

[0168] Table 1 Experimental hardware parameters:

[0169]

[0170] To evaluate the prediction results of the network model, this experiment uses error rate and F1 score as evaluation metrics.

[0171] The ST-Res-GCN model from Example 1 was compared with the CNN (2D), CNN (1D), LSTM (Long Short-Term Memory), and SVM (Support Vector Machine) models in binary classification experiments for each type of fault under the same input data. The results are shown in Table 2.

[0172] Table 2: Results of the binary classification experiment

[0173]

[0174] The ST-Res-GCN model from Example 1 was compared with several other models, including Convolutional Neural Network (CNN) (2D), Convolutional Neural Network (CNN) (1D), Long Short-Term Memory Network (LSTM), and Support Vector Machine (SVM), using the same input data. The results are shown in Table 3.

[0175] Table 3: Results of the multi-class comparison experiment

[0176]

[0177]

[0178] The results in Tables 2 and 3 show that the ST-Res-GCN model has a better average F1 score than conventional deep learning and machine learning algorithms, and the fault diagnosis rate is significantly improved. This indicates that the ST-Res-GCN model can extract the spatiotemporal features of the data more effectively than conventional algorithms.

[0179] Experiments demonstrate that the ST-Res-GCN model, using the MIC matrix as the adjacency matrix in its graph convolutional structure, effectively extracts the correlations between different variables and the temporal variation characteristics of the variables themselves. Furthermore, the combination of the residual graph convolution module and the spatiotemporal feature extraction module effectively mines the spatiotemporal relationship of the data, thereby significantly improving the accuracy and performance of fault diagnosis. Prediction results show that, in both binary and multi-class classification tests, the average F1 score of the ST-Res-GCN model is 9% and 21% higher than that of the CNN (2D) model, respectively; 11% and 24% higher than that of the CNN (1D) model, respectively; 10% and 23% higher than that of the LSTMs model, respectively; and 22% and 46% higher than that of the SVM model, respectively.

[0180] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of this invention should be considered within the scope of protection of this invention.

Claims

1. A method for fault diagnosis in industrial processes based on spatiotemporal fusion graph convolutional networks, characterized in that: Fault data from industrial production processes are collected, and then the fault data is normalized using z-score and extracted using a sliding window to obtain the node feature matrix. , node feature matrix The results are input into the trained and tested ST-Res-GCN model to obtain fault classification results; The ST-Res-GCN model comprises, in sequence, an input layer, a first batch normalization layer (BN Layer1), a first graph convolutional layer (GCN Layer1), three residual graph convolutional modules (Res-GCN), a fifth batch normalization layer (BN Layer5), a first one-dimensional convolutional layer (CNN Layer1), two spatiotemporal feature extraction modules (ST), a global max pooling layer (GMP Layer), a dropout layer (Dropout Layer), and an output layer. The residual graph convolutional module Res-GCN includes a batch normalization layer (BN Layer), a graph convolutional layer (GCN Layer), and an adder layer. The spatiotemporal feature extraction module ST includes a batch normalization layer (BN), a long short-term memory network (LSTM), a one-dimensional convolutional layer (CNN), and an additive layer (Add Layer).

2. The method for fault diagnosis of industrial processes based on spatiotemporal fusion graph convolutional networks according to claim 1, characterized in that: The training and testing process of the ST-Res-GCN model is as follows: (1) The simulation experimental data is divided into two parts: training set and test set. Each part contains normal data and fault data. (2) Calculate the MIC values ​​between different variables in the normal data of the training set as the adjacency matrix. After z-score standardization of the fault data in the training set, a sliding window is used to obtain the node feature matrix of the training set. Then the adjacency matrix of the training set and node feature matrix The ST-Res-GCN model is input for training, with cross-entropy as the loss function, Adam algorithm as the optimization algorithm, and an early stopping mechanism is added to find the best model, thus obtaining a trained ST-Res-GCN model. (3) After z-score standardization of the fault data in the test set, the node feature matrix of the training set is obtained by using a sliding window. The node feature matrix of the test set Input the trained ST-Res-GCN model, evaluate it using the error rate and F1 score, and adjust the parameters of the ST-Res-GCN model until the evaluation metrics meet the preset evaluation criteria, thus obtaining the trained and tested ST-Res-GCN model, where the adjacency matrix during testing is... Adjacency matrix used during training same.

3. The method for fault diagnosis of industrial processes based on spatiotemporal fusion graph convolutional networks according to claim 2, characterized in that: The adjacency matrix The acquisition process is as follows: (1) Data extraction For the normal data Perform z-score normalization, then select a length of... Time series data to obtain dataset ,in, , N is the dataset The number of records in the time series data; (2) MIC value calculation Select datasets in sequence Two time series data points Calculate the MIC value between the two data points: (1) (2) in, for The joint distribution , They are respectively and marginal distribution, for mutual information, The upper limit of the grid. , For the sample size, and for Grid lines and List; (3) Adjacency matrix calculation right All The maximum information coefficient (MIC) is calculated between the time series data according to equation (1) to obtain a symmetrical result. Matrix, set a threshold ,Will Less than in the matrix The MIC values ​​are replaced with 0, while the other values ​​remain unchanged, to obtain the MIC matrix as the adjacency matrix. .

4. The method for fault diagnosis of industrial processes based on spatiotemporal fusion graph convolutional networks according to claim 3, characterized in that: The process of capturing the sliding window is as follows: The fault data after z-score normalization is used Given the window length, move and truncate the window with a step size of 1 to obtain... Group dataset As the node feature matrix X, where , The time series length of the fault data. .

5. The method for fault diagnosis of industrial processes based on spatiotemporal fusion graph convolutional networks according to claim 4, characterized in that: The input layer includes the node feature matrix. and the adjacency matrix ; The residual graph convolution module Res-GCN includes a first residual graph convolution module Res-GCN 1, a second residual graph convolution module Res-GCN 2, and a third residual graph convolution module Res-GCN 3; The spatiotemporal feature extraction module ST includes a first spatiotemporal feature extraction module ST1 and a second spatiotemporal feature extraction module ST2.

6. The method for fault diagnosis of industrial processes based on spatiotemporal fusion graph convolutional networks according to claim 5, characterized in that: The calculation process of the first convolutional layer GCN Layer 1 is as follows: (8-1) in, , It is the identity matrix; yes The angle matrix, for The normalized matrix, As weight, For activation functions; The calculation output for the batch normalization layer, For input.

7. The method for fault diagnosis of industrial processes based on spatiotemporal fusion graph convolutional networks according to claim 6, characterized in that: The first residual graph convolutional module Res-GCN 1 includes: a second batch normalization layer (BN Layer2), a second graph convolutional layer (GCN Layer2), and a first additive layer (Add Layer1). The calculation process is as follows: ，(10) in, and These represent the outputs of the first residual graph convolutional module Res-GCN 1 and the second graph convolutional layer GCNlayer2, respectively. (8-2) The second residual graph convolutional module Res-GCN 2 includes: a third batch normalization layer (BN Layer3), a third graph convolutional layer (GCN Layer3), and a second additive layer (Add Layer2). The calculation process is as follows: (11) in, and These represent the outputs of the second residual graph convolutional module Res-GCN 2 and the third graph convolutional layer GCLRayer3, respectively. (8-3) The third residual graph convolutional module Res-GCN3 includes: a fourth batch normalization layer (BN Layer4), a fourth graph convolutional layer (GCN Layer4), and a third additive layer (Add Layer3). The calculation process is as follows: (12) in, and These represent the outputs of the third residual graph convolutional module Res-GCN3 and the fourth graph convolutional layer GCLRayer4, respectively. (8-4) in, , and All are weights.

8. The method for fault diagnosis of industrial processes based on spatiotemporal fusion graph convolutional networks according to claim 7, characterized in that: The first spatiotemporal feature extraction module ST 1 includes: a sixth batch normalization layer (BN Layer 6), a first long short-term memory (LSTM) network layer 1, a seventh batch normalization layer (BN Layer 7), a second one-dimensional convolutional layer (CNN Layer 2), and a fourth additive layer (Add Layer 4). The calculation process is as follows: (22) in, , and These represent the outputs of the first spatiotemporal feature extraction module ST 1, the first one-dimensional convolutional layer CNNLayer1, and the second one-dimensional convolutional layer CNNLayer2, respectively. (14) (21) in, This represents the computation of a one-dimensional convolutional layer. Represents the computation of long short-term memory networks; The second spatiotemporal feature extraction module ST 2 includes: an eighth batch normalization layer (BN Layer 8), a second long short-term memory (LSTM) network layer 2, a ninth batch normalization layer (BN Layer 9), a third one-dimensional convolutional layer (CNN Layer 3), and a fifth addition layer (Add Layer 5). The calculation process is as follows: (23) (24) in, , and These represent the outputs of the second spatiotemporal feature extraction module ST2, the first spatiotemporal feature extraction module ST1, and the third one-dimensional convolutional layer CNN Layer3, respectively.

Images