A chemical process fault diagnosis method based on deep graph network

Abstract

The present invention discloses a chemical process fault diagnosis method based on a deep graph network, comprising the following steps: (1) data acquisition and integration; (2) data selection and sorting; (3) establishment of a causal directed graph; (4) modeling training; and (5) model testing. The present invention utilizes a chemical process fault diagnosis method based on a deep graph network, improves fault identification, and accurately locates fault variables, thus having universality and versatility.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical process fault diagnosis, and in particular to a chemical process fault diagnosis method based on a deep graph network. Background Art

[0002] In modern process industries, sensor technology and information systems are widely used for data collection and storage. This increasing complexity of chemical processes poses significant challenges for fault diagnosis. Although fault diagnosis has been developed for decades, it remains a difficult task due to the widespread propagation of faults across diverse process units. A large-scale industrial process typically consists of a variety of interconnected units, such as chemical reactors, heat exchangers, and distillation columns. Consequently, faults can easily propagate from one unit to another along information and material flows. The use of feedback control loops further complicates the analysis of propagation mechanisms. Fault diagnosis primarily relies on three approaches: knowledge-based, model-based, and data-based. However, due to the inherent drawback of knowledge-based and model-based approaches, which rely heavily on expert knowledge, they are not widely applicable to diverse chemical processes. Therefore, data-based approaches have emerged as a promising approach.

[0003] Data-based methods can be categorized as statistical, shallow learning, and deep learning. Statistical methods include principal component analysis, independent component analysis, partial least squares, Fisher discriminant analysis, qualitative trend analysis, and their derivatives. Shallow learning methods include support vector machines, artificial immune systems, K-nearest neighbor, Gaussian mixture models, and artificial neural networks. While both types of data-driven methods offer advantages, they are still far from widespread practical application because they typically require extensive domain knowledge to determine fault characteristics in spatial and temporal domains and their fault diagnosis rates are often insufficient.

[0004] In the past few years, deep learning has become an outstanding technology, showing better performance than the above methods in many fields. However, although deep learning methods have achieved high accuracy, they are still far from practical applications because they cannot explain the excellent performance of the models in combination with mechanistic processes.

[0005] Graph neural networks (GNNs), as a branch of deep learning, are becoming increasingly popular and have demonstrated powerful learning capabilities in multiple fields. Because their data is defined in a non-Euclidean space, GNNs possess more powerful visualization capabilities. Graph convolutional networks (GCNs) are a variant of traditional GNNs and share all of their advantages. GCNs achieve powerful representation capabilities by performing convolution operations on structured data. However, almost all graph networks lack causal relationships between nodes. This means that previous work assumed that different variables were identical and could not use correlation to model data from a causal analysis perspective. Therefore, this method introduces causal analysis into GCNs. In recent years, causal analysis has received increasing attention in the field of fault diagnosis research because causal graphs can intuitively represent fault propagation paths and reveal the root causes of faults. Summary of the Invention

[0006] To address the shortcomings of the existing technology, this paper proposes a causal-based deep spatiotemporal causal graph convolution network (DSTCGCN). Unlike many previous fault diagnosis models, DSTCGCN introduces a causal relationship graph, treating each variable in the graph as a graph node. It models and analyzes the time series data of process variables and the causal relationships between them. Since process data of different time lengths contain different dynamics, this method uses a fixed sampling module to enable process data of different lengths to be used to train the same model. In DSTCGCN, there are two types of edges: spatial edges that conform to the natural connectivity of process variables, and temporal edges that connect the same process variable across consecutive time steps. This means that the information propagation process is based on both spatial and temporal dimensions. By stacking identical layers on this basis, DSTCGCN enables the extraction of more abstract, high-dimensional data. Furthermore, by incorporating a graph collapse mechanism into the model, DSTCGCN simplifies the system into several independent subsystems, which provides universal expressive power and achieves higher performance.

[0007] The technical solutions of the present invention are as follows:

[0008] A chemical process fault diagnosis method based on a deep graph network, the method comprising the following steps:

[0009] 1) Data acquisition and integration:

[0010] Set step faults through the TE simulation process, and divide, collect and integrate the data generated by each fault;

[0011] 2) Data selection and sorting:

[0012] Data selection is performed based on the Euclidean distance metric to remove redundant and useless data; data is sorted based on the maximum mutual information criterion;

[0013] 3) Establish a causal directed graph:

[0014] According to the TE flow chart, a causal directed graph is constructed between its variables;

[0015] 4) Modeling training:

[0016] A DSTCGCN model based on deep spatiotemporal graph convolutional network is constructed. The input of the model is the data retained after data selection. The DSTCGCN model is then used to learn and train the integrated data.

[0017] 5) Model testing:

[0018] The trained DSTCGCN model is used for fault diagnosis of TE process.

[0019] Further, the process of step 1) is:

[0020] Step 1.1: Set the Tennessee Eastman (TE) process faults and obtain the output data for each fault.

[0021] Step 1.2: Divide the data into training and test sets according to the ratio;

[0022] Step 1.3: Since the data differences between different feature variables are large, it is necessary to standardize the data to facilitate model processing and calculation. The specific formula is as follows:

[0023]

[0024] Among them, X' is the data after standardization, X is the original data without standardization, μ is the mean of the data, and γ is the standard deviation of the data.

[0025] Further, the process of step 2) is:

[0026] Step 2.1: Use the Euclidean distance to measure the similarity between the data of each variable and the normal working condition r z , which represents the similarity between the fault variable number z and the normal variable number z. The greater the similarity, the smaller the contribution of the variable to the fault. Similarity r z The formula is as follows:

[0027]

[0028] Z is the total number of fault categories, ε is to prevent the denominator from being an infinitesimal amount of 0, x zand x0 represent the fault variable number z and the normal variable number z respectively;

[0029] Step 2.2: Exclude variables with a similarity of 0.98-1 to complete the feature selection process;

[0030] Step 2.3: Randomly select a variable and calculate the maximum mutual information of the remaining variables with respect to the selected variable. The formula is as follows:

[0031]

[0032] Where p(x,y) is the joint probability distribution between random variables x and y, B = N 0.55～0.6 , N is the size of the sample set;

[0033] Step 2.4: Sort the features based on maximum mutual information.

[0034] Further, the process of step 3) is:

[0035] Step 3.1: Decompose the process into several equipment-level units along the process flow;

[0036] Step 3.2: Mark the connection points of the equipment units with numbers;

[0037] Step 3.3: Create a causal directed graph for each equipment unit;

[0038] Step 3.4: Using the causal directed graph of each equipment unit, construct a causal directed graph of the system based on the numbers marked at the process flow and equipment unit connections.

[0039] Further, the process of step 4) is:

[0040] Step 4.1: DSTCGCN is a hierarchical model built on a fixed sampling layer, a graph pooling layer, and a spatiotemporal graph convolution layer;

[0041] The fixed sampling layer mainly divides the input data of different time lengths into a fixed number of divisions and then takes the average to obtain data of fixed time length;

[0042] The graph pooling layer is an algorithm that can simplify the graph structure. For the graph G, the cluster allocation matrix is ​​defined. Where C represents the number of nodes in the simplified graph, and finally the simplified adjacency matrix A is obtained coar :

[0043] A coar =S T AS (3)

[0044] Where A is the adjacency matrix of the original causal directed graph with Z nodes. According to the simplified number of nodes R, its Laplacian matrix is ​​used to define its pooling operator to simplify the number of variables:

[0045]

[0046] where Θ l is the adjacency matrix A coar The eigenvector of

[0047] The spatiotemporal convolution layer is mainly composed of a graph convolutional network and a convolutional long short-term memory network. Graph convolution is an algorithm that convolves data in non-Euclidean space and can be expressed as:

[0048]

[0049] Where H is the current layer feature vector, l represents the number of layers, * represents the convolution operation, Γ is the convolution kernel, σ is the activation function, A represents the input adjacency matrix, D = ∑ j A j Degree matrix representing the adjacency matrix;

[0050] The convolutional long short-term memory network is an algorithm that determines the future state of a cell in the grid by its current input and past state, which can be expressed as:

[0051]

[0052] Where, о represents the Hadamard product, I t represents the result of the input gate, F t Represents the result of the forget gate, C t Represents the result of the memory unit at the current moment, O t Represents the result of the output gate, H t Represents the module output result at the current moment, W represents the trainable parameter, and B represents the trainable bias;

[0053] Step 4.2: Stack the layers in 4.1 in order and combine them with the feature selection and sorting parts to form the required DSTCGCN model. Fault identification of the TE process is performed using the DSTCGCN model.

[0054] Step 4.3: Use the training data to train DSTCGCN and finally obtain the trained DSTCGCN model.

[0055] Furthermore, the process of step (5) is as follows:

[0056] Step 5.1: Input the test data into the trained DSTCGCN model. Evaluate the model using FDR, where FDR is the ratio of the number of correct predictions to the actual number of classes:

[0057]

[0058] Where p is the number of correctly predicted classes, and b is the number of incorrectly predicted classes.

[0059] The beneficial effects of this invention are primarily manifested in the following: It proposes a chemical process fault diagnosis method based on a deep graph network, which improves the accuracy of defect identification and location. This method utilizes the powerful visualization capabilities of a causal directed graph and graph networks to identify and extract system fault information. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a flow chart of the TE process of the present invention as an example;

[0061] Figure 2 is a causal directed graph of the variables selected by the present invention;

[0062] Figure 3 It is a schematic diagram of the fixed sampling layer of the present invention;

[0063] Figure 4 It is the spatiotemporal convolution layer structure of the present invention;

[0064] Figure 5 It is the framework of the method of the present invention;

[0065] Figure 6 It is a simplified causal directed graph of the variables selected by the present invention. DETAILED DESCRIPTION

[0066] The present invention will be further described below with reference to the accompanying drawings.

[0067] Reference Figures 1 to 6 , a chemical process fault diagnosis method based on deep graph network, the method comprising the following steps:

[0068] 1) Data acquisition and integration

[0069] Seven step faults were set through the Tennessee Eastman (TE) simulation process. Each fault generated 160 sets of data which were divided, collected and integrated.

[0070] The process of step 1) is:

[0071] Step 1.1: Set 7 faults in the TE process and obtain 160 sets of 52 variables of output data for each fault. Table 1 shows the 7 faults, and Table 2 shows the 52 variables and their descriptions. Figure 1 For the TE process flow.

[0072] Table 1 Different fault descriptions

[0073]

[0074] Table 2 Description of different variables

[0075]

[0076] Step 1.2: Divide the data into training and test sets in a 5:5 ratio.

[0077] Step 1.3: Since the data differences between different feature variables are large, it is necessary to standardize the data to facilitate model processing and calculation. The specific formula is as follows:

[0078]

[0079] Among them, X' is the data after standardization, X is the original data without standardization, μ is the mean of the data, and γ is the standard deviation of the data.

[0080] 2) Data selection and sorting

[0081] Data selection is performed based on the Euclidean distance metric to remove redundant and useless data. Data is sorted based on the maximum mutual information criterion.

[0082] The process of step 2) is:

[0083] Step 2.1: Use the Euclidean distance to measure the similarity between the data of each variable and the normal working condition r z , which represents the similarity between the fault variable number z and the normal variable number z. The greater the similarity, the smaller the contribution of the variable to the fault. Similarity r z The formula is as follows:

[0084]

[0085] In this case, the total number of fault categories is Z = 7, and ε is used to prevent the denominator from being an infinitesimal number of 0. z and x0 represent the fault variable number z and the normal variable number z respectively.

[0086] Step 2.2: Eliminate variables with too high similarity (0.98-1) to complete the feature selection process.

[0087] Step 2.3: Randomly select a variable and calculate the maximum mutual information of the remaining variables with respect to the selected variable. The formula is:

[0088]

[0089] Where p(x,y) is the joint probability distribution between random variables x and y, B = N 0.55～0.6, N is the size of the sample set.

[0090] Step 2.4: Sort the features based on maximum mutual information.

[0091] 3) Establish a causal directed graph

[0092] According to the TE flow chart, a causal directed graph is constructed between its variables.

[0093] The process of step 3) is:

[0094] Step 3.1: Decompose the process into several equipment-level units along the process flow.

[0095] Step 3.2: Mark the connections of the device units with numbers.

[0096] Step 3.3: Create a causal directed graph for each equipment unit.

[0097] Step 3.4: Using the causal directed graph of each equipment unit, construct a causal directed graph of the system based on the numbers marked at the process flow and equipment unit connections. Figure 2 is the causal directed graph of the selected variables.

[0098] 4) Modeling training

[0099] An algorithm model DSTCGCN for chemical process fault diagnosis based on deep graph network is constructed. The input of the model is the selected retained data; then the DSTCGCN model is used to learn and train the integrated data.

[0100] The process of step 4) is:

[0101] Step 4.1: DSTCGCN is a hierarchical model mainly based on fixed sampling layer, graph pooling layer, and spatiotemporal graph convolution layer. The fixed sampling layer mainly divides the input data of different time lengths into a fixed number of divisions and then takes the average to obtain data of fixed time length, such as Figure 3 shown.

[0102] The graph pooling layer is an algorithm that can simplify the graph structure. For the graph G, define the cluster assignment matrix Where C represents the number of nodes in the simplified graph. Finally, the simplified adjacency matrix A is obtained coar :

[0103] A coar =S T AS (4)

[0104] Where A is the adjacency matrix of the original causal directed graph with Z nodes. Based on the simplified number of nodes R, its Laplacian matrix is ​​used to define its pooling operator to simplify the number of variables:

[0105]

[0106] where Θ l is the adjacency matrix A coar The eigenvector of .

[0107] The spatiotemporal convolution layer is mainly composed of graph convolutional networks and convolutional long short-term memory networks, such as Figure 4 As shown. Graph convolution is an algorithm that convolves data in non-Euclidean space and can be expressed as:

[0108]

[0109] Where H is the current layer feature vector, l represents the number of layers, * represents the convolution operation, Γ is the convolution kernel, σ is the activation function, A represents the input adjacency matrix, D = ∑ j A j Degree matrix representing the adjacency matrix.

[0110] The convolutional long short-term memory network is an algorithm that determines the future state of a cell in the grid by its current input and past state, which can be expressed as:

[0111]

[0112] Where, о represents the Hadamard product, I t represents the result of the input gate, F t Represents the result of the forget gate, C t Represents the result of the memory unit at the current moment, O t Represents the result of the output gate, H t Represents the module output result at the current moment, W represents the trainable parameter, and B represents the trainable bias.

[0113] Step 4.2: Stack the layers in 4.1 in order, and combine them with the feature selection and sorting parts to form the required DSTCGCN model, and use DSTCGCN to identify faults in the TE process. Figure 5 It is the DSTCGCN framework.

[0114] Step 4.3: Use the training data to train DSTCGCN and finally obtain the trained DSTCGCNM model. Figure 6 To simplify the causal directed graph.

[0115] 5) Model testing

[0116] The trained DSTCGCN model is used for fault diagnosis of TE process.

[0117] The process of step 5) is:

[0118] Step 5.1: Input the test data into the trained DSTCGCN model. Evaluate the model using FDR, where FDR is the ratio of the number of correct predictions to the actual number of classes:

[0119]

[0120] Where p is the number of correctly predicted classes, and b is the number of incorrectly predicted classes.

[0121] Example:

[0122] (1) Obtain the TE process data set. The process is as follows:

[0123] Step 1.1: Set up the TE process, measure variables every 3 minutes, set 7 different faults, and repeat the simulation process 160 times. Finally, obtain 160 × 52 × 400 data with 52 variables and 400 time points.

[0124] Step 1.2: Divide the data into two parts of 80 × 52 × 400 and 80 × 52 × 400 according to the partitioning criteria, which are the training set and the test set respectively;

[0125] Step 1.3: Process each data according to the standardization formula.

[0126] (2) Variable selection and sorting for the TE process is as follows:

[0127] Step 2.1: Calculate each variable according to the similarity formula.

[0128] Step 2.2: Delete the variables with a similarity of 1, and finally retain the following 35 variables: S11, S18, MV7, S17, S4, MV11, S8, MV2, MV1, S21, S27, S26, S23, MV10, S10, MV6, MV3, S1, MV8, S9, S2, S3, MV4, S16, S7, S13, S22, S25, S24, S12, S5, S6, S15, S14, S19.

[0129] Step 2.3: Select variable S11 and calculate the mutual information of the remaining variables.

[0130] Step 2.4: Sort the variables according to mutual information as follows: S11, S18, MV7, MV2, MV1, S21, MV10, S10, MV6, MV3, S1, MV4, S16, S7, S13, S22, S25, S27, S26, S23, S24, MV8, S9, S2, S3, S12, S5, S6, S15, S14, S19, S17, S4, MV11, S8.

[0131] (3) Draw the causal directed graph of the screening variables. The process is as follows:

[0132] Step 3.1: Draw a device-level causal directed graph.

[0133] Step 3.2: Based on the device-level causal directed graph, draw the causal directed graph of all variables.

[0134] (4) Perform DSTCGCN model training. The process is as follows:

[0135] Step 4.1: According to the causal directed graph, set the number of nodes of the simplified causal directed graph to 10 and simplify the causal directed graph. Table 3 shows the subordinate relationship between the nodes and the simplified nodes.

[0136] Step 4.2: Based on the causal directed graph and the simplified causal directed graph, the training set is input into the model to obtain the first recognition result of the model.

[0137] Step 4.3: According to the recognition results and the true results, adjust the model parameters so that the difference between the recognition results and the true results is reduced.

[0138] Step 4.4: Repeat steps 4.2 to 4.3 until the recognition error of the model is within the allowable error.

[0139] Table 3 Subordinate relationships between nodes

[0140]

[0141] (5) Use the test data to test the model. The process is as follows:

[0142] Step 5.1: Use the support vector machine (SVM) as a comparison model. SVM is a classic classification and recognition model. Input the test data into the trained DSTCGCN and SVM, and obtain the recognition results of both on the test set.

[0143] Step 5.2: Calculate the FDR of DSTCGCN and SVM on the test set and evaluate their performance. The evaluation results are shown in Table 4:

[0144] Table 4 Comparison between DSTCGCN and SVM

[0145]

[0146] From the above comparison results, it can be seen that the fault identification of the TE process by the present invention is much better than that of the traditional classification and identification model.

[0147] The method of the present invention adopts a chemical process fault diagnosis method based on a deep graph network, which improves the fault identification effect and accurately locates the fault variables, and has universality and versatility.

[0148] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The scope of protection of the present invention should not be regarded as limited to the specific forms described in the embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A chemical process fault diagnosis method based on deep graph network, characterized in that: The method comprises the following steps: 1) Data acquisition and integration: Set step faults through the Tennessee-Eastman TE simulation process and divide, collect and integrate the data generated by each fault; 2) Data selection and sorting: Data selection is performed based on the Euclidean distance metric to remove redundant and useless data; data is sorted based on the maximum mutual information criterion; 3) Establish a causal directed graph: According to the TE flow chart, a causal directed graph is constructed between its variables; The process of step 3) is: Step 3.1: Decompose the process into several equipment-level units along the process flow; Step 3.2: Mark the connection points of the equipment units with numbers; Step 3.3: Create a causal directed graph for each equipment unit; Step 3.4: Using the causal directed graphs of each equipment unit, construct a causal directed graph of the system based on the numbers marked at the process flow and equipment unit connections; 4) Modeling training: A DSTCGCN model based on deep spatiotemporal graph convolutional network is constructed. The input of the model is the data retained after data selection. The DSTCGCN model is then used to learn and train the integrated data. The process of step 4) is: Step 4.1: DSTCGCN is a hierarchical model built on a fixed sampling layer, a graph pooling layer, and a spatiotemporal graph convolution layer; The fixed sampling layer mainly divides the input data of different time lengths into a fixed number of divisions and then takes the average to obtain data of fixed time length; The graph pooling layer is an algorithm that can simplify the graph structure. For the graph G, the cluster allocation matrix is ​​defined. Where C represents the number of nodes in the simplified graph, and finally the simplified adjacency matrix A is obtained coar : A coar =S T AS (3) Where A is the adjacency matrix of the original causal directed graph with Z nodes. According to the simplified number of nodes R, its Laplacian matrix is ​​used to define its pooling operator to simplify the number of variables: where Θ l is the adjacency matrix A coar The eigenvector of The spatiotemporal convolution layer is mainly composed of a graph convolutional network and a convolutional long short-term memory network. Graph convolution is an algorithm that convolves data in non-Euclidean space and can be expressed as: Where H is the current layer feature vector, l represents the number of layers, * represents the convolution operation, Γ is the convolution kernel, σ is the activation function, A represents the input adjacency matrix, D = ∑ j A j Degree matrix representing the adjacency matrix; The convolutional long short-term memory network is an algorithm that determines the future state of a cell in the grid by its current input and past state, which can be expressed as: in, represents the Hadamard product, I t represents the result of the input gate, F t Represents the result of the forget gate, C t Represents the result of the memory unit at the current moment, O t Represents the result of the output gate, H t Represents the module output result at the current moment, W represents the trainable parameter, B represents the trainable bias, X t represents input data, tanh(.) represents the hyperbolic tangent function, W xi ,W xf ,W xc ,W xo Represents the input data X in the input gate, forget gate, memory unit and output gate respectively t The trainable weights W hi ,W hf ,W hc ,W ho Represents the module output result H at the previous moment in the input gate, forget gate, memory unit and output gate respectively t-1 The trainable weights of B i ,B f ,B c ,B o Represent the trainable bias in the input gate, forget gate, memory unit and output gate respectively; Step 4.2: Stack the layers in 4.1 in order and combine them with the feature selection and sorting parts to form the required DSTCGCN model. Fault identification of the TE process is performed using the DSTCGCN model. Step 4.3: Use the training data to train DSTCGCN and finally obtain the trained DSTCGCN model; 5) Model testing: The trained DSTCGCN model is used for fault diagnosis of TE process.

2. A chemical process fault diagnosis method based on deep graph network according to claim 1, characterized in that: The process of step 1) is: Step 1.1: Set TE process faults and obtain output data for each fault; Step 1.2: Divide the data into training and test sets according to the ratio; Step 1.3: Since the data differences between different feature variables are large, it is necessary to standardize the data to facilitate model processing and calculation. The specific formula is as follows: Among them, X' is the data after standardization, X is the original data without standardization, μ is the mean of the data, and γ is the standard deviation of the data.

3. A chemical process fault diagnosis method based on deep graph network according to claim 1, characterized in that: The process of step 2) is: Step 2.1: Use the Euclidean distance to measure the similarity between the data of each variable and the normal working condition r z , which represents the similarity between the fault variable number z and the normal variable number z. The greater the similarity, the smaller the contribution of the variable to the fault. Similarity r z The formula is as follows: Z is the total number of fault categories, ε is to prevent the denominator from being an infinitesimal amount of 0, x z and x0 represent the fault variable number z and the normal variable number z respectively; Step 2.2: Exclude variables with a similarity of 0.98-1 to complete the feature selection process; Step 2.3: Randomly select a variable and calculate the maximum mutual information of the remaining variables with respect to the selected variable. The formula is as follows: Where p(x,y) is the joint probability distribution between random variables x and y, B = N 0.55～0.6 , N is the size of the sample set; Step 2.4: Sort the features based on maximum mutual information.

Images