Industrial process fault diagnosis method based on multi-scale collaborative graph convolutional network

Abstract

The invention relates to an industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network. According to the method, the data characteristics of the industrial process are fully considered, aiming at the multi-scale dynamic characteristics of the industrial process data, firstly, feature information of different scales is captured through the multi-scale feature extraction module, then the topological structure relation of the sensor network is learned by using the mixed graph convolutional network, and the feature association learned by the model is enhanced. Particularly, the hybrid graph convolutional network effectively solves the problem of excessive smoothness of a traditional graph convolutional network by introducing a historical information memory mechanism. Compared with an existing fault diagnosis model, multi-scale features can be efficiently extracted, and the topological structure relation of the complex industrial process can be represented. According to the method, high diagnosis accuracy can still be kept under the condition of small-batch data training, and compared with a traditional method, the fault diagnosis performance is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial process control, and relates to an industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network. Background Art

[0002] In the field of industrial process management, fault diagnosis is a crucial link to ensure operation efficiency, safety and cost-effectiveness, and its importance is self-evident. However, a major problem faced by the actual industrial process is the scarcity of labeled data. Given that the data labeling process usually relies on expert knowledge and consumes a large amount of resources, in the actual industrial environment, only a small amount of the collected data can be effectively labeled. This limiting factor seriously restricts the performance of traditional supervised learning (SL) fault diagnosis models, and thus weakens the accuracy and reliability of their fault diagnosis.

[0003] To solve the above problems, an effective strategy is to construct a fault diagnosis model based on a multi-scale feature extraction method. The core advantage of this model lies in its ability to simultaneously use limited labeled data to construct an excellent multi-scale feature extractor, and further realize the fusion of features at different scales, so as to improve the performance of fault diagnosis in the scenario of scarce labeled data. Specifically, this model aims to enhance the generalization ability of the model by combining the advantages of multi-scale feature extraction and feature fusion, learning the implicit information in different scale spaces from limited labeled data to make up for the deficiency of labeled data, and improving the fault diagnosis performance of the model.

[0004] However, the complex interactions of different variables and operating parameters in the industrial process at various time scales make it difficult for traditional static or single-scale fault diagnosis models to accurately represent the multi-scale relationships therein, and the small batch of labeled samples makes the model prone to underfitting during the training process, resulting in a significant decline in the fault diagnosis ability of the model.

[0005] At the same time, industrial sensors distributed at different positions construct a complex sensor network, which collects industrial process data at different positions. Traditional fault diagnosis models cannot effectively represent the topological structure changes of different fault categories in the sensor network and cannot assist industrial engineers in accurately locating the fault positions.

[0006] In summary, the existing solutions to the problem of degraded fault diagnosis performance in the scenario of scarce labeled data fail to fully consider the multi-scale characteristics of industrial process data and the changes in its topological structure. To solve the above problems, the present invention proposes an industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network, which constructs a fault diagnosis model by combining a multi-scale collaborative feature extraction method and a mixed graph convolutional network (MGCN). Through multi-scale feature extraction, features are collaboratively extracted from different scales to capture the dynamic characteristics of industrial process data at different time scales. At the same time, the computational overhead of multi-scale feature extraction is reduced through collaborative feature extraction; combined with the graph convolutional network, the internal variable relationships in the industrial topology are learned and characterized, and an industrial process fault diagnosis model supported by the topological structure is constructed. Summary of the Invention

[0007] The object of the present invention is to provide an industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional encoding and decoding network.

[0008] The present invention specifically includes the following steps:

[0009] Step 1: Collect the labeled data set in the industrial process. Randomly sample, and take a samples X l ∈R V×1 in the labeled data set as the training set X train , and the remaining labeled sample set as the test set X test . The input samples of the training set correspond to the fault category labels Y train ∈R n×1 , the input samples of the test set correspond to the fault category labels Y test ∈R n×1 , and save them to the storage medium, where V represents the number of variables and n represents the number of fault categories.

[0010] Step 2: Construct a scale collaborative graph convolutional network fault diagnosis model MsPaF-GCN:

[0011] The MsPaF-GCN model includes a multi-scale encoding network, a multi-scale decoding network, a multi-scale feature fusion network, and a classification network.

[0012] The multi-scale encoding network includes a multi-scale collaborative network, a multi-scale adaptive graph structure learning network, and a mixed graph convolutional network. First, use the multi-scale collaborative network MsPaF to collaboratively map the input samples of the training set to different scale spaces to obtain multi-scale features {X1,…,X K}. Then, use the multi-scale adaptive graph structure learning network MsGSL to construct corresponding topological structures for different scale features to sparsify the adjacency matrices {A1,…,A K} is represented. Finally, the multi-scale features and the corresponding adjacency matrices are mapped to the latent spaces of different scales through the encoding network using the Mixed Graph Convolutional Network (MGCN) to obtain the latent multi-scale features.

[0013] The multi-scale decoding network includes a multi-scale adaptive graph structure learning network and a mixed graph convolutional network. The multi-scale adaptive graph structure learning network and the mixed graph convolutional network are used to map the latent multi-scale features and the corresponding adjacency matrices {A1, …, A K} to the latent spaces of different scales to obtain the decoded multi-scale features {h1, …, h K}.

[0014] The multi-scale feature fusion network MsFF includes a convolutional neural network and a fully connected network. The multi-scale features {h1, …, h K} are input into the convolutional neural network to obtain the fused feature f, realizing the information interaction between different scales.

[0015] The classification network includes a fully connected layer and the activation function relu. The multi-scale fused feature f is input into this network to obtain the fault diagnosis categories of the labeled input data in the training set.

[0016] Step 3: Input the training sample X train into the MsPaF-GCN model built in Step 2 for training.

[0017] Step 4: Input the test set sample X test into the network trained in Step 3 to obtain the output diagnosis category. Compare the output diagnosis category with the output sample category Y test to verify the model.

[0018] Step 5: Input any variable in the industrial process into the MsPaF-GCN model verified in Step 4 to output the fault diagnosis classification result.

[0019] Furthermore, the specific content of Step 3 is as follows:

[0020] Select samples from the training sample set X train as the input in the classification training stage, and update the model parameters through classification training. Set the parameters of the encoding and decoding network as θ c , and the classification loss function adopts the cross-entropy loss: According to the loss function, the fault classification labels and fault category labels of the training samples, use the Adam optimizer to perform gradient iterative update on each network parameter.

[0021] Furthermore, the specific strategy for gradient iterative update of the network parameters is as follows: the parameters of the encoding and decoding network where μ0 is the learning rate for classification training.

[0022] Furthermore, in Step 4, the model evaluation metrics precision, recall, and F1-score are used to evaluate the prediction results of the test set, and the model is verified based on the evaluation results.

[0023] The present invention fully considers the data characteristics of industrial processes. Aiming at the multi-scale dynamic characteristics of industrial process data, it first captures feature information at different scales through a multi-scale feature extraction module, and then uses a hybrid graph convolutional network to learn the topological structure relationship of the sensor network, enhancing the feature correlation learned by the model. In particular, the hybrid graph convolutional network effectively solves the over-smoothing problem of traditional graph convolutional networks by introducing a historical information memory mechanism. Compared with existing fault diagnosis models, the present invention can efficiently extract multi-scale features and characterize the topological structure relationship of complex industrial processes. This method can still maintain a high diagnostic accuracy under the condition of small-batch data training, and significantly improves the fault diagnosis performance compared with traditional methods.

[0024] For industrial processes, data is collected from the same sensor network, and there is a time delay in the data collected by different sensors. These data reflect the internal correlations between variables at different scales. The present invention not only improves the utilization efficiency of data resources, but also significantly reduces the dependence on expert experience and manual annotation resources; secondly, multi-level collaborative feature extraction can effectively capture the complex correlation patterns and abnormal features hidden in industrial process faults, reducing the computational overhead in the training process while ensuring the model performance; in addition, the use of a hybrid graph convolutional network endows the model with the ability to learn the topological structure of complex industrial sensor networks. This method not only provides a high-performance and high-precision fault diagnosis modeling method for industrial process management, but also provides a more solid technical foundation for equipment fault monitoring and maintenance management under the background of Industry 4.0. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is the overall structure block diagram of the MsPaF-GCN model;

[0026] Figure 2 is the process flow diagram of the Tennessee Eastman (TE) process;

[0027] Figure 3 is the fault diagnosis result diagram of the MsPaF-GCN model for the TE process;

[0028] Figure 4 is the topological structure diagram generated by the MsPaF-GCN model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0030] Aiming at the problem of fault diagnosis modeling for complex dynamic industrial processes, the present invention extracts multi-scale features from data and adaptively learns the topological structure relationship, and proposes an industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network.

[0031] This method uses a multi-scale encoding network to extract features of different scales from data and map them to the latent space, decodes the latent multi-scale features through a multi-scale decoding network and learns the latent association between the multi-scale features and the fault labels, and uses the topological structure relationship to fix the variable relationship beneficial to fault classification. Combining multi-scale features with topological structures for classification modeling of fault diagnosis, diagnosing and predicting the faults hidden in industrial process data. It can automatically extract multi-scale features from the input data to be measured, thereby strengthening the association between variables by using the learned topological structure relationship. Under the condition of small batch data training, the classification accuracy of the fault diagnosis model for complex industrial process faults can be greatly improved.

[0032] An industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network specifically includes the following steps:

[0033] Step 1: Use sensors, industrial control systems, etc. to collect a labeled data set in the industrial process. Randomly sample, and use a samples X l ∈R V×1 in the labeled data set as the training set X train , and the remaining labeled sample set as the test set X test . The input samples of the training set correspond to the fault category labels Y train ∈R n×1 , the input samples of the test set correspond to the fault category labels Y test ∈R n×1 , and save them to the storage medium, where V represents the number of variables and n represents the number of fault categories.

[0034] Step 2: Build a multi-scale collaborative graph convolutional network fault diagnosis model (MsPaF-GCN model) as shown in Figure 1 on the open-source deep learning platform Torch; the MsPaF-GCN model includes a multi-scale encoding network, a multi-scale decoding network, a multi-scale feature fusion network, and a classification network. The input feature X t is collaboratively mapped to different scale spaces through the multi-scale encoding network to achieve parallel extraction of features of different scales, obtaining multi-scale features {X1,…,X K}, and through the multi-scale adaptive graph structure learning network and the hybrid graph convolutional network, the multi-scale features {X1,…,X K} is mapped to potential multi-scale features The multi-scale decoding network decodes the multi-scale latent features to obtain the decoded multi-scale features {h1, …, h K}, and finally the multi-scale fusion network fuses the multi-scale features and maps the fused features to fault classification labels through the classification network.

[0035] The multi-scale encoding network includes a multi-scale collaborative network, a multi-scale adaptive graph structure learning network, and a hybrid graph convolutional network. First, the multi-scale collaborative network (MsPaF) is used to collaboratively map the input samples of the training set to different scale spaces. This network uses a convolutional network containing several different convolutional kernels to extract features from the data, thereby achieving parallel extraction of features at different scales and obtaining multi-scale features {X1, …, X K}. Then, the multi-scale adaptive graph structure learning network (MsGSL) is used to construct corresponding topological structures for the features at different scales to sparsify the adjacency matrix {A1, …, A K} representation. Finally, the hybrid graph convolutional network (MGCN) maps the multi-scale features and the corresponding adjacency matrix to the latent space at different scales through the encoding network to obtain the potential multi-scale features

[0036] The traditional graph convolutional network takes features and topological structures as inputs and realizes topological fusion through node information transmission and message aggregation. However, it is limited by the over-smoothing problem caused by the increase in network depth. The present invention uses the hybrid graph convolutional network (MGCN) to effectively alleviate the over-smoothing phenomenon by dynamically integrating historical node information into the calculation process of the current graph network layer, enabling the model to construct a deeper network structure.

[0037] The multi-scale decoding network includes a multi-scale adaptive graph structure learning network and a hybrid graph convolutional network. The multi-scale adaptive graph structure learning network and the hybrid graph convolutional network are used to map the potential multi-scale features and the corresponding adjacency matrix {A1, …, A K} to the latent space at different scales to obtain the decoded multi-scale features {h1, …, h K}.

[0038] The multi-scale feature fusion network (MsFF) includes a convolutional neural network and a fully connected network. The multi-scale features {h1, …, h K} are input into the convolutional neural network to obtain the fused feature f, realizing information interaction between different scales.

[0039] The classification network includes a fully connected layer and the activation function relu. The multi-scale fused feature f is input into this network to obtain the fault diagnosis categories of the labeled input data in the training set

[0040] Step 3. Input the training sample X described in Step 1 train , into the MsPaF-GCN model built in Step 2 for training; Select samples from the training sample set X train as the input in the classification training stage, and update the model parameters through classification training. Set the encoding-decoding network parameters as θ c , and the classification loss function adopts cross-entropy (CE) loss: According to the loss function, the fault classification label and fault category label of the training sample, use the Adam optimizer to perform gradient iteration update on each network parameter. The specific strategy is as follows:

[0041] Encoding-decoding network parameters where μ0 is the learning rate of classification training.

[0042] Step 4. Input the test set sample X test into the network trained in Steps 1-3 to obtain the output diagnosis category Compare the output diagnosis category with the output sample category Y test for verification of the model.

[0043] To evaluate the industrial process fault diagnosis model based on the multi-scale collaborative graph convolutional encoding-decoding network (MsPaF-GCN), use the model evaluation indicators Precision, Recall, and F1 score to evaluate the prediction results of the test set:

[0044]

[0045] Step 5. Input any variable in the industrial process into the MsPaF-GCN model verified in Step 4 to output the fault diagnosis classification result.

[0046] The following demonstrates the effectiveness of the multi-scale collaborative graph convolutional encoding-decoding network model in combination with the Tennessee Eastman process (hereinafter referred to as the TE process), Figure 2 and gives the process flow diagram of the TE process. The TE process consists of five operating units: a reactor, a condenser, a separator, a compressor, and a stripper. The main products are gases A, C, D, E, inert gas B, liquids G, H, and by-product F. This process has a total of 41 measurement variables and 12 operating variables. The specific descriptions of the measurement variables are shown in Table 1. Among the 41 measurement variables, the first 22 are basic process variables, and the last 19 are composition variables. In this example, 31 variables are used as the input variables of the model.

[0047] Table 1: Description of TE process measurement variables

[0048] Number Description Number Description 1 Feed A 17 Bottom Flow of Stripping Tower 2 Feed D 18 Temperature of Stripping Tower 3 Feed E 19 Steam Flow of Stripping Tower 4 Total Feed 20 Compressor Power 5 Recirculation Flow 21 Outlet Temperature of Reactor Cooling Water 6 Feed Flow Rate of Reactor 22 Outlet Temperature of Separator Cooling Water 7 Reactor Pressure 23 Feed Flow Rate of D 8 Reactor Level 24 Feed Flow Rate of E 9 Reactor Temperature 25 Feed Flow Rate of A 10 Discharge Rate 26 Total Feed Flow Rate 11 Temperature of Product Separator 27 Discharge Valve 12 Level of Product Separator 28 Liquid Flow of Separator Tank 13 Pressure of Product Separator 29 Liquid Product Flow of Stripping Tower 14 Bottom Flow of Product Separator 30 Cooling Water Flow of Reactor 15 Level of Stripper 31 Cooling Water Flow of Condenser 16 Pressure of Stripping Tower

[0049] 1. The TE process introduces industrial process data of one working condition, including a total of 28 types of fault samples and one type of normal sample. The training set has 20,254 sampling points of labeled data, the validation set has 40,509 sampling points, and the test set has 20,254 sampling points.

[0050] 2. Build the MsPaF-GCN model on the deep learning platform Torch and use the test set to verify the fault diagnosis performance. The number of multi-scale collaborative layers is set to 4, the depth of the hybrid graph convolution is set to 1, the output channels of the multi-scale fusion network are set to 8, and the number of hidden layer nodes of the built encoding and decoding network is set to 31. Among them, the classifier for the fused features is fitted through a one-layer neural network with 29 nodes. During training, the learning rate μ0 = 10 -3 , the number of training iterations is 50, and the adaptive momentum estimation iterative algorithm is used to update the model parameters.

[0051] 3. The MsPaF-GCN model proposed in the present invention was compared with the long short-term memory (LSTM) and graph deviation network (GDN) models in terms of prediction results on the test set. The prediction effect of the method proposed in the present invention is as Figure 3 shown, and the multi-scale graph structure is as Figure 4 shown. The evaluation indicators Precision, Recall, and F1 score of the above three models are shown in Table 2.

[0052] Table 2: Prediction result table of four models

[0053] Index / Model LSTM GDN MsPaF - GCN Test Precision 0.6503 0.6994 0.9032 Test Recall 0.6983 0.7059 0.9000 Test F1 score 0.6617 0.6919 0.9002

[0054] It can be seen that the MsPaF-GCN model can effectively separate multi-scale features and generate topological relationships of different scales. Under the condition of small-batch sample training, it has good classification accuracy for fault diagnosis of complex industrial processes.

[0055] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network, characterized in that: Specifically, it includes the following steps: Step 1: Collect the labeled dataset in the industrial process; randomly sample, and use a samples X in the labeled dataset l ∈R V ×1 as the training set X train , and use the remaining labeled sample set as the test set X test ; in the training stage, input the sample corresponding to the fault category label Y train ∈R n×1 of the training set, input the sample corresponding to the fault category label Y test ∈R n×1 of the test set, and save them to the storage medium, where V represents the number of variables and n represents the number of fault categories; Step 2: Construct a multi-scale collaborative graph convolutional network fault diagnosis model MsPaF-GCN: The MsPaF-GCN model includes a multi-scale encoding network, a multi-scale decoding network, a multi-scale feature fusion network, and a classification network; The multi-scale encoding network includes a multi-scale collaborative network, a multi-scale adaptive graph structure learning network, and a hybrid graph convolutional network. First, the training set input samples are collaboratively mapped to different scale spaces using the multi-scale collaborative network MsPaF to obtain multi-scale features {X1, …, X K}; Then, the multi-scale adaptive graph structure learning network MsGSL is used to construct corresponding topological structures for different scale features, which are represented by sparsifying the adjacency matrices {A1, …, A K}; Finally, the hybrid graph convolutional network MGCN maps the multi-scale features and the corresponding adjacency matrices to the latent spaces of different scales through the encoding network to obtain the latent multi-scale features The multi-scale decoding network includes a multi-scale adaptive graph structure learning network and a hybrid graph convolutional network. The multi-scale adaptive graph structure learning network and the hybrid graph convolutional network are used to map the latent multi-scale features and the corresponding adjacency matrices {A1, …, A K} to latent spaces of different scales, obtaining the decoded multi-scale features {h1, …, h K}; The multi-scale feature fusion network MsFF includes a convolutional neural network and a fully connected network. The multi-scale features {h1,…,h K} are input into the convolutional neural network to obtain the fused feature f, realizing information interaction between different scales; The classification network includes a fully connected layer and an activation function relu. The multi-scale fusion feature f is input into this network to obtain the fault diagnosis categories of the labeled input data in the training set. Step 3: Input the training sample X in Step 1 train into the MsPaF-GCN model constructed in Step 2 for training; Step 4: Input the test set sample X test into the network trained in Step 3 to obtain the output diagnostic category Compare the output diagnostic category with the output sample category Y test to verify the model; Step 5: Input any variable in the industrial process into the MsPaF-GCN model verified in Step 4, and output the fault diagnosis classification result.

2. The industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network according to claim 1, wherein: Specifically, Step 3 is as follows: Select samples from the training sample set X train as the input for the classification training stage, and update the model parameters through classification training; set the encoding-decoding network parameters as θ c , and the classification loss function uses cross-entropy loss: According to the loss function, the fault classification labels and fault category labels of the training samples, use the Adam optimizer to perform gradient iterative updates on each network parameter.

3. The industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network according to claim 2, characterized in that: The specific strategy for gradient iterative update of the network parameters is as follows: the network parameters of the encoder-decoder where μ0 is the learning rate for classification training.

4. The industrial process fault diagnosis method based on a multi-scale collaborative graph convolutional network according to claim 1, characterized in that: In Step 4, the model evaluation metrics precision, recall, and F1 score are used to evaluate the prediction results of the test set, and the model is verified based on the evaluation results.

Images