A fault diagnosis method based on graph isomorphism autoencoder

Abstract

A fault diagnosis method based on a graph isomorphism autoencoder includes the following steps: 1) data acquisition and integration: using fault information data from a TE simulation process, and collecting, dividing, and processing the data; 2) modeling and training: constructing a GIAE fault diagnosis model and using the GIAE model to learn and train the integrated data; 3) model prediction and comparison: applying the trained GIAE model to fault prediction, providing prediction accuracy, and verifying the superiority of GIAE performance. The method utilizes a large amount of data to guide the encoder to mine the sample's discriminative intrinsic information and guides the model to map this intrinsic information to a fault category space. Furthermore, GIAE uses variable importance and score vectors to quantify the contribution of variables and variable relationships to sample categories, better reflecting the subtle differences between different industrial process faults, effectively improving fault prediction, and demonstrating universality and versatility.

Description

Technical Field

[0001] The present invention belongs to the field of industrial process fault diagnosis, and in particular relates to a fault diagnosis method based on graph isomorphism autoencoder. 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-driven methods can be categorized into statistical methods, shallow learning methods, and deep learning methods. Statistical methods include principal component analysis, partial least squares, qualitative trend analysis, and their derivatives. Shallow learning methods include support vector machines, Gaussian mixture models, and artificial neural networks. Although both types of data-driven methods offer advantages, they are still far from being widely adopted in practice because they typically require extensive domain knowledge to determine fault characteristics in both spatial and temporal domains and their fault diagnosis rates are often low.

[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] Recently, graph neural networks (GNNs), a branch of deep learning, have demonstrated superior performance in many fields. As the core of GNNs, topological graphs contain sufficient process knowledge and mechanisms. Combining neural networks with process mechanisms can simultaneously leverage the advantages of both knowledge and data to model and explore fault information. Furthermore, GNNs effectively propagate variable information through topological graphs, which helps model the propagation of faults. Generally speaking, topological graphs under different faults are considered a set of isomorphic graphs, meaning that the graph structure remains unchanged. Therefore, the training process of GNNs is dominated by variables, meaning that the topological graph only guides the GNN in message transmission and does not provide additional fault information for discriminative learning. GNN-based base models need to consider how to provide differentiated graph representations for isomorphic graphs based on faults to reveal the complex interactions between variables. Summary of the Invention

[0006] To address the aforementioned issues in the prior art, the present invention aims to provide a fault diagnosis method based on a graph isomorphic auto-encoder (GIAE). Starting from process knowledge, it learns the discriminative intrinsic information of the data and provides data classification judgment. First, a topological map of the variables is constructed using a flowchart, which ensures that the model contains sufficient process knowledge. In addition, the contribution of variables and variable relationships to different faults is quantified using variable importance and score vectors, respectively, which helps to increase the interpretability of the model. Subsequently, an encoder containing a graph isomorphic embedding (GIE) and a mapping head is used as a base model to model the fault propagation process, followed by a classification head for prediction output. This method has been applied to fault diagnosis in the Tennessee Eastman (TE) simulation process. It has been verified that it effectively improves fault identification and accurately locates the fault variable, showing its universality and versatility.

[0007] The technical solution adopted by the present invention to solve its technical problem is:

[0008] A fault diagnosis method based on graph isomorphism autoencoder includes the following steps:

[0009] 1) Data acquisition and integration

[0010] The fault information data of Tennessee Eastman TE simulation process is used, and the data is collected, divided and processed.

[0011] 2) Modeling training

[0012] A fault diagnosis method GIAE based on graph isomorphism autoencoder is constructed. The input of the model is the selected retained data; then the GIAE model is used to learn and train the integrated data.

[0013] 3) Model prediction and comparison

[0014] The trained GIAE model is used for fault prediction, and the prediction accuracy is given. To verify the superiority of GIAE's performance, a series of baseline methods are compared with it, including Graph Isomorphic Network (GIN) and Consistent Regularized Auto-Encoder (CRAE).

[0015] Further, the process of step 1) is:

[0016] Step 1.1) Set up the TE simulation process, set a step fault in the simulation process, and divide, collect and integrate the data generated by each fault.

[0017] Step 1.2) Since the data differences between different variables are large, it is necessary to standardize the data to facilitate model processing and calculation. The specific formula is as follows:

[0018]

[0019] In the formula, X' is the data after standardization, X is the original data without standardization, μ is the mean of the data, and σ is the standard deviation of the data.

[0020] Further, the process of step 2) is:

[0021] Step 2.1) Construction of the GIAE topology graph with scores.

[0022] Compared with scalars, vectors have stronger expressive power, so choosing vectors as scores helps provide differentiation. Multi-head Attention Mechanism (MAM) is used to adaptively learn score vectors. The process data is defined as Where K represents the number of samples, V represents the number of variables, and T represents the number of time steps. MAM combines variables pairwise and calculates the variable x ii →x ij The fractional vector Where h is the number of heads. MAM calculates s ii→ij The process is as follows:

[0023]

[0024] In the formula, g(·) represents a nonlinear mapping function, and the ReLU activation function is used here to transform x ii from Map to Where F represents the mapping dimension.

[0025] Then, the V×V score vector s of the i-th sample ii→ij The final score matrix Finally, use To represent X i The mixed score topological graph of Represents the Hadamard product, which represents the element-wise multiplication operation.

[0026] Step 2.2) Modeling the graph isomorphism embedding encoder of GIAE.

[0027] The graph isomorphism embedding encoder is proposed for graph isomorphism learning and graph classification tasks. It has strong expressive power and flexibility. It can fully utilize the structure and edge information of isomorphic graphs to model the propagation process of faults. The implementation steps of the graph isomorphism embedding encoder are as follows:

[0028] 1. Convert the score vector Map to

[0029] 2. For variable x ii Add a self-loop, that is, with the variable x ii Add an edge pointing to itself to the corresponding node to ensure the influence of the variable itself;

[0030] 3. Aggregate x using activation and update functions ii 、x ij and s ij→ii Get new feature representation.

[0031] The formula for the graph isomorphism embedding encoder is as follows:

[0032]

[0033] Where, Represents the variable x in the lth layer GIAE ii The embedding representation of g(·) represents a nonlinear mapping function. The graph isomorphism embedding encoder uses the same structure g(·) as MAM. ε represents a trainable parameter for x ii Add a self-loop with weight, Ω(i) represents x ii The set of adjacent variables in the topological graph, Represents a trainable linear mapping matrix used to change s ij→ii The dimension, W XUsed to convert X i from Map to

[0034] Then, we use a graph isomorphic embedding encoder and a mapping head consisting of two fully connected layers to construct the encoder f(·;θ). This method of using the mapping head to map the embedding to another latent space can improve the performance, where θ represents the encoder parameters, so the variable x ii The encoder output embedding for:

[0035]

[0036] Next, a classification head is connected after the encoder to embed the output z ii Mapping to available labels Y i , the calculation formula is:

[0037] Y i =h(z ii ) (5)

[0038] Where h(·) represents a nonlinear mapping function, and the Sigmoid activation function is used here.

[0039] Step 2.3) GIAE loss establishment and training

[0040] By adding z ii With Y i Do cross entropy loss and repeat training to reduce the loss. The calculation formula is:

[0041]

[0042] Where p(·) and q(·) represent probability distributions.

[0043] Further, the process of step 3) is:

[0044] Step 3.1) Input the test data into the trained GIN, CRAE, and GIAE models to obtain the predicted values.

[0045] Step 3.2) Use the Fault Diagnosis Rate (FDR) as the evaluation metric to calculate the FDR value of each model. The calculation formula is as follows:

[0046]

[0047] Where p represents the number of correctly predicted values ​​for that class, and b represents the number of incorrectly predicted values ​​for that class. Generally speaking, a larger FDR means the closer the model's predicted values ​​are to the true values, meaning the better the model's prediction performance.

[0048] The beneficial effects of the present invention are mainly manifested in:

[0049] This paper proposes a fault diagnosis method based on graph isomorphism autoencoder, which uses a large amount of data to guide the encoder to mine the discriminative intrinsic information of the sample and guide the model to map the intrinsic information to the fault category space; in addition, GIAE uses variable importance and score vectors to quantify the contribution of variables and variable relationships to sample categories, which allows GIAE to better reflect the subtle differences under different faults in industrial processes, effectively improve the fault prediction effect, and has universality and versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a structural flow chart of the model of the present invention; DETAILED DESCRIPTION

[0051] The present invention will be further described below with reference to the accompanying drawings.

[0052] Reference Figure 1 , a fault diagnosis method based on graph isomorphism autoencoder, the method comprising the following steps:

[0053] (1) Obtaining process fault information dataset

[0054] Step 1.1: Set up the TE simulation process. The dataset has a data sampling interval of 3 minutes and includes 52 variables and 21 fault types. Fault 0 represents normal operation, S1-41 are process variables, and MV42-52 are manipulated variables. The training and test sets contain 500 and 960 samples for each fault, respectively. The fault is introduced in the first hour of the training set and the eighth hour of the test set; therefore, the data from hours 0-1 and 0-8 are excluded from the training and test sets, which contain 480 and 800 samples, respectively. Furthermore, Fault 6 in the training and test sets causes the machine to shut down after 7 and 14 hours, respectively. Therefore, the training and test sets contain only 140 and 280 samples for Fault 6, respectively. The dataset used in this study is available from the Harvard database. Variables S28-41 are excluded because they are challenging to measure with sensors in real time. Furthermore, S20, MV5, and MV9 are also excluded as they do not contribute to the fault, as per relevant literature. Therefore, 35 variables are used as model inputs.

[0055] Step 1.2: Process each data according to the standardization formula.

[0056] (2) GIAE model training is performed as follows:

[0057] Step 2.1: Input the training set into the model to obtain the first prediction value of the model.

[0058] Step 2.2: Based on the difference between the predicted value and the output value, adjust the parameters of the model so that the difference between the predicted value and the true value is reduced.

[0059] Step 2.3: Repeat steps 2.1 to 2.2 until the difference between the model's predicted value and the true value is within the allowable error.

[0060] (3) Use the test data to test the model. The process is as follows:

[0061] Step 3.1: Input the test data into the trained GIN, CRAE, and GIAE to obtain the predicted values ​​of GIN, CRAE, and GIAE on the test set, respectively.

[0062] Step 3.2: Calculate the FDR value of the predicted value of the test set and evaluate the performance of GIN, CRAE, and GIAE. The evaluation results are shown in Table 1.

[0063] Table 1 FDR values ​​of GIN, CRAE and GIAE for each fault

[0064]

[0065]

[0066] From the above comparison results, it can be seen that GIAE outperforms the other two models in the prediction accuracy of most faults, thereby improving the effect of fault diagnosis.

[0067] The present invention adopts a fault diagnosis method based on graph isomorphism autoencoder, which effectively improves the prediction effect of faults in industrial processes and has universality and versatility.

[0068] 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 fault diagnosis method based on graph isomorphism autoencoder, characterized in that: The following steps are involved: 1) Data acquisition and processing Use the fault information data from the Tennessee Eastman TE simulation process and collect, divide and process the data; 2) Modeling training A fault diagnosis model based on a graph isomorphism autoencoder (GIAE) is constructed. The model input is the selected retained data. The GIAE model is used to train the integrated data. The specific process of step 2 is as follows: Step 2.1: Construction of the score topology of the GIAE model Step 2.1.

1. Select a vector as the score and use the multi-head attention mechanism MAM to adaptively learn the score vector; define the process data as: Where K represents the number of samples, V represents the number of variables, and T represents the number of time steps; for the i-th sample MAM combines variables pairwise and calculates the variable x ii →x ij The score vector Where h is the number of heads; MAM calculation s ii→ij The process is as follows: In the formula, g(·) represents a nonlinear mapping function, which is used to transform x ii from Map to Where F represents the mapping dimension; Step 2.1.2, V×V score vector s of the i-th sample ii→ij Constructing the score matrix Step 2.1.3, use To represent X i The mixed score topological graph of is the Hadamard product, which represents the element-wise multiplication operation; Step 2.2: Modeling the graph isomorphism embedding encoder of the GIAE model Step 2.2.1: Graph Isomorphism Embedding Encoder. The specific implementation steps are as follows: (1) The score vector s ij→ii from Map to (2) is the variable x ii Add a self-loop, that is, with the variable x ii Add an edge pointing to itself to the corresponding node to ensure the influence of the variable itself; (3) Aggregate x using activation and update functions ii 、x ij and s ij→ii Get a new feature representation; the formula of the graph isomorphism embedding encoder is as follows: Where, Represents the variable x in the lth layer GIAE ii The embedding representation of g(·) represents a nonlinear mapping function. The graph isomorphism embedding encoder uses the same structure g(·) as MAM. ε represents a trainable parameter for x ii Add a self-loop with weight, Ω(i) represents x ii The set of adjacent variables in the topological graph, Represents a trainable linear mapping matrix used to change s ij→ii The dimension, W X Used to convert X i from Map to Step 2.2.2, use the graph isomorphic embedding encoder and the mapping head containing two layers of fully connected layers to form the encoder f(·; θ), and use the mapping head to map the embedding to another latent space to improve the performance, where θ represents the encoder parameters; therefore, the variable x ii The encoder output embedding for: Step 2.2.

3. Connect the classification head after the encoder to embed the output z ii Mapping to available label Y i , the calculation formula is: Y i =h(z ii ) (5) Where h(·) represents the nonlinear mapping function; Step 2.3, Loss establishment and training of GIAE model By adding z ii With Y i Do cross entropy loss and repeat training to reduce the loss. The calculation formula is: Where p(·) and q(·) represent probability distributions; 3) Model prediction and comparison The trained GIAE model is used for fault prediction, the prediction accuracy is given, and the superiority of GIAE performance is verified.

2. A fault diagnosis method based on graph isomorphism autoencoder according to claim 1, characterized in that: The process of step 1) is as follows: Step 1.1: Set up the TE simulation process. Set step faults in the simulation process and collect, divide, and process the data generated by each fault. Step 1.2: Standardize the data to facilitate model processing and calculation. The specific formula is as follows: Where X' is the data after standardization, X is the original data without standardization, μ is the mean of the data, and σ is the standard deviation of the data.

3. A fault diagnosis method based on graph isomorphism autoencoder according to claim 1, characterized in that: The process of step 4) is as follows: Step 3.1: Input the test data into the trained GIN, CRAE, and GIAE models to obtain the predicted values. Step 3.2: Use the fault diagnosis rate (FDR) as the evaluation indicator to calculate the FDR value of each model. The calculation formula is as follows: In the formula, p represents the number of correctly predicted classes, and b represents the number of incorrectly predicted classes. The larger the FDR, the closer the predicted value of the model is to the true value, that is, the better the prediction effect of the model.

Images