Abnormity detection method and system based on graph neural network and auto-encoder

Abstract

The invention belongs to the technical field of artificial intelligence and industrial automation, and discloses an anomaly detection method and system based on a graph neural network and an auto-encoder, and the method comprises the steps: constructing a directed graph to represent a sensor network, carrying out the coding of a device state, fusing the coded device state with node features to reflect physical connection, extracting features, and carrying out the weighted aggregation of node information, the method comprises the following steps: automatically learning weight to highlight contribution of key equipment, reconstructing a sensor state by adopting an auto-encoder, carrying out consistency judgment, outputting abnormal early warning when an error exceeds a threshold value or a prediction deviation is large, and detecting abnormity based on a reconstruction error and a structure inconsistency probability by utilizing an unsupervised learning training model. According to the invention, the accuracy and robustness of the anomaly detection of the switch data of the nuclear power plant are improved, and an unsupervised efficient detection means is provided especially in the environment of complex dependency relationship among multiple data and diversified equipment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and industrial automation, and particularly relates to a device on-off data anomaly detection method and system fusing a graph neural network and an autoencoder. BACKGROUND

[0002] Large nuclear power units are generally equipped with digital instrument control systems (DCS) and plant information systems (PI). The PI system can store a large amount of unit operation data generated in the DCS system, and the stored operation data is divided into analog data and on-off data. The analog data generally includes temperature, pressure, flow, and other sensor measurement data, and the on-off data generally includes the on-off state data of remote control devices such as valves and pumps. These operation data provide an important data basis for the digitalization and intelligentization of nuclear power systems.

[0003] On-off data is a common binary state information in the operation of nuclear power units, mainly from remote control devices such as valves, pumps, and switches. The on-off state of these devices is crucial for the operation and safety of the unit. Compared with analog data, on-off data is usually represented as "on" or "off" two states, used to reflect the start-stop, working and non-working state of the device. Although on-off data is small in quantity and simple in change, it plays an important role in fault diagnosis, state monitoring and early warning systems of nuclear power units. For example, whether the valve is fully open or the pump is working normally, on-off data can provide operators with intuitive device state information. In order to improve the intelligent level of nuclear power unit operation, the accuracy and real-time performance of on-off data must be guaranteed, and it is often used together with analog data in analysis to form a more comprehensive and real-time monitoring of the unit.

[0004] However, random errors may occur in the process of generating and storing these data, resulting in abnormal data. Traditional anomaly detection methods based on rules or single-point features are difficult to capture the topological dependence between data, and require high labeling of abnormal samples, which is insufficient in accuracy and reliability of anomaly detection, and difficult to meet actual needs. SUMMARY

[0005] The purpose of the present application is to provide a device on-off data anomaly detection method and system fusing a graph neural network and an autoencoder, to improve the accuracy and robustness of on-off data anomaly detection in nuclear power plants, especially in complex environments with multiple data dependencies and diverse devices, to provide an unsupervised and efficient detection method.

[0006] In order to achieve the above purpose, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an anomaly detection method based on a graph neural network and an autoencoder, comprising:

[0008] Step 1: Construct a directed graph to represent the sensor network, and fuse the device state encoding with node features to reflect physical connections;

[0009] Step 2: Extract features, weighted aggregate node information, and automatically learn weights to highlight the contributions of key equipment;

[0010] Step 3: Reconstruct the sensor state using an autoencoder and perform consistency judgment. Output an abnormal warning when the error exceeds the threshold or the prediction deviation is large.

[0011] Step 4: Train the model using unsupervised learning and detect anomalies based on reconstruction error and structural inconsistency probability.

[0012] As an feasible approach, a node contains two input dimensions: the current value and the difference value.

[0013] As an feasible approach, the input data construction based on graph structure modeling can form a directed acyclic graph structure of various sensors in a nuclear power plant to reflect the physical connection relationship of the working fluid flow in the process flow.

[0014] As an feasible approach, a graph attention mechanism is introduced for feature extraction, a graph neural network encoder is constructed, and a multi-layer conditional graph attention module is used to achieve weighted aggregation of information between multiple nodes in the steam flow topology.

[0015] As an feasible approach, the attention mechanism is formulated as follows:

[0016]

[0017] In the formula, W is the linear transformation matrix, h i and h j These are the features of node i and node j, respectively, h k Let be the features of node k, a be the learning weight vector of the attention mechanism, N(i) represent the set of neighboring nodes of node i, and || represent vector concatenation.

[0018] As an feasible approach, the sensor state is reconstructed using an autoencoder structure, and a dedicated valve verification module is introduced to make a consistency judgment on the valve control state. If the sensor error is higher than the normal statistical threshold, or if there is a strong deviation between the valve prediction and the original state, an abnormal warning signal is output.

[0019] As an feasible approach, unsupervised learning is used to adapt to different working conditions. The model is trained on all normal data, eliminating the need for manual labeling of anomaly types or locations. Anomalies are judged by the reconstruction error and structural inconsistency probability of the embedding space.

[0020] As an implementable way, in anomaly detection, the autoencoder evaluates the degree of abnormality of the data points through reconstruction error, and if the reconstruction error exceeds the predetermined threshold, it is judged as abnormal, the formula is as follows:

[0021] Encoder: The encoder maps the input x to a latent space representation z through a nonlinear transformation:

[0022] z=f enc (x)=σ(W enc x+b enc )

[0023] In the formula, f enc is the encoding function, W enc is the weight matrix of the encoder, b enc is the bias term, and σ is the activation function.

[0024] Decoder: The decoder maps the latent representation z back to the original space:

[0025] x′=f dec (z)=σ(W dnc z+b dnc )

[0026] In the formula, f dnc is the decoding function, W dnc is the weight matrix of the decoder, b dnc is the bias term, and x ′ is the reconstructed input data.

[0027] In a second aspect, the application provides an anomaly detection system based on graph neural network and autoencoder, comprising:

[0028] The data processing module is used for separating sensor and valve data, calculating sensor differential value, and initializing and splicing into an input matrix;

[0029] The model construction and training module is used for encoding the input matrix through GAT, combining the decoder to reconstruct the sensor and valve data, calculating the loss optimization parameters after standardization, and training the reconstruction model;

[0030] The anomaly detection and judgment module is used for standardizing the test data and splicing, using the trained model to infer and calculate the reconstruction error, judging the anomaly according to the threshold, and outputting the visualization result file;

[0031] The optimization and multivariate modeling module is used for optimizing the model structure, adjusting the sensor and valve weight coefficients, combining GNN and autoencoder to reduce misjudgment, and improving the anomaly detection accuracy;

[0032] The model evaluation and result output module is used for evaluating the reconstruction error and accuracy of the test set, saving the result file for analysis.

[0033] As an feasible approach, the model evaluation and results output module is used to evaluate the performance of the model, with evaluation metrics including detection accuracy, recall, and F1 score.

[0034] Compared with existing technologies, the anomaly detection method and system based on graph neural networks and autoencoders provided in this application have the following advantages:

[0035] This application integrates graph neural networks and temporal autoencoders to achieve high-precision anomaly detection under complex working conditions.

[0036] The GAT-AE (autoencoder based on graph attention mechanism) model proposed in this application makes full use of the physical topological relationship between sensor nodes in the steam flow network. By introducing graph neural network (GAT) to model the information flow between nodes, it effectively improves the detection capability of local anomalies and chain anomalies under complex structural conditions.

[0037] This application combines the representational capabilities of deep learning, the structural modeling advantages of graph neural networks, and the operability of industrial deployment, significantly improving the accuracy of anomaly detection and system adaptability under complex operating conditions, and is applicable to data anomaly detection in nuclear power process systems.

[0038] This application supports multi-source heterogeneous data modeling, enhancing the model's ability to perceive multiple types of anomalies. The model simultaneously receives continuous sensor data such as temperature and pressure, as well as discrete valve state data. It constructs a Conditional Graph Attention (CondGAT) mechanism, which, by fusing valve control states as global conditional information of the graph, enhances the model's ability to identify two different types of anomalies: "abnormal valve control" and "sensor drift."

[0039] This application achieves unsupervised anomaly detection without relying on large-scale anomaly samples for training. During the training phase, an autoencoder model is built solely based on normal data, enabling the model to learn the system's high-dimensional feature reconstruction capabilities under normal conditions. In the testing phase, anomalies are jointly determined by reconstruction error and the probability of structural consistency anomalies (such as in the valve verification module), effectively reducing the dependence on anomaly labels and providing greater engineering adaptability and deployment convenience.

[0040] This application offers strong interpretability and clearly identifies the paths to abnormal nodes and events. By leveraging the node weight distribution output by the graph attention mechanism, the system can trace which key sensors or node regions each detected abnormal event occurred in, providing good visualization and expert interpretability, which helps on-duty personnel to quickly diagnose and respond.

[0041] The application has structural generalization capability, and is suitable for different topological working conditions or sensor number change scenes. Through the modular design of CondGAT and the general embedding layer, the application supports dynamic adjustment of the number of sensors, is suitable for nuclear power units with different working conditions, local topology changes or sensor addition and deletion, does not need to retrain the model, and only needs to load the trained model weight to quickly apply, has good portability and robustness.

[0042] The application has real-time deployment capability, low computing cost and fast response speed. The model inference process is realized based on graph convolution and attention mechanism optimization, and the overall computing amount is much lower than that of the traditional detection scheme based on sliding window clustering or high-order statistical method, which meets the online real-time deployment requirements and meets the requirements of nuclear power emergency response on detection timeliness and stability. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the application, the following will briefly introduce the drawings needed to be used in the technical description.

[0044] Figure 1 The flowchart of the abnormality detection method based on the graph neural network and the autoencoder provided by the application is shown in the figure.

[0045] Figure 2 The data processing and model input structure diagram provided by the application is shown in the figure.

[0046] Figure 3 The joint reconstruction flowchart of the sensor and the valve state in the GAT model provided by the application is shown in the figure.

[0047] Figure 4 The abnormality detection flowchart provided by the application shows the data input, inference, error calculation and abnormality judgment process.

[0048] Figure 5 The error visualization diagram provided by the application directly shows the parameter error of the abnormal sample after reconstruction, and provides visual explanation for the detection result.

[0049] Figure 6 The ROC (Receiver Operating Characteristic) curve diagram provided by the application is commonly used to evaluate the performance of a classification model. The blue curve is drawn by calculating the TPR and FPR under different thresholds. The closer the curve is to the upper left corner, the better the model performance is. The AUC is 0.9066, indicating that the model has good performance, especially at a high TPR, the false positive rate is low, which means that the model can better distinguish positive and negative examples.

[0050] Figure 7 The topological structure diagram in the implementation example of the application is shown in the figure.

[0051] Figure 8 Topology information extraction diagram provided for the present application;

[0052] Figure 9 Structure diagram of the anomaly detection system based on graph neural network and autoencoder provided for the present application. DETAILED DESCRIPTION

[0053] The specific embodiments are further described in detail below.

[0054] As Figures 1 to 6 shown, the present application provides an anomaly detection method based on graph neural network and autoencoder, comprising:

[0055] Step 1: Construct a directed graph to represent the sensor network (such as temperature, pressure), the node contains the current value and the differential value, and the device state is encoded and fused with the node feature to reflect the physical connection;

[0056] Step 2: Use the graph attention mechanism (GAT) to extract features, and use the multi-layer conditional graph attention module (CondGATConv) to weight and aggregate node information, and automatically learn the weight to highlight the key device contribution;

[0057] Step 3: Use the autoencoder to reconstruct the sensor state, and introduce a valve verification module to judge the consistency, and output an abnormal warning when the error exceeds the threshold or the prediction deviation is large;

[0058] Step 4: Train the model using unsupervised learning, only normal data is required, and detect anomalies based on reconstruction error and structural inconsistency probability, improve the generality and generalization ability.

[0059] In step 1, the input data construction based on graph structure modeling, the temperature, pressure and other sensors in the nuclear power plant are constructed into a directed graph acyclic graph structure to reflect the physical connection relationship of the working medium flow in the process flow. Each node contains two input dimensions, i.e. the current value + differential value of the sensor. The device state is encoded by a neural network and fused with the node feature as a global conditional variable.

[0060] In step 2, the graph attention mechanism (GAT) is introduced for feature extraction. A graph neural network encoder is constructed, and a multi-layer conditional graph attention module (CondGATConv) is used to realize the weighted aggregation of information among multiple nodes in the steam flow topology structure. The attention mechanism automatically learns the weight according to the influence degree of different nodes on the overall state, so as to highlight the contribution of key devices or sensors.

[0061] In step 3, a multi-branch decoding mechanism is used to realize anomaly restoration and structure verification. The sensor state is reconstructed using an autoencoder structure, and a dedicated valve validation module is introduced to judge the consistency of the valve control state. If the sensor error exceeds the normal statistical threshold, or if there is a strong deviation between the valve prediction and the original state, an anomaly warning signal is output.

[0062] In step 4, unsupervised learning is employed to flexibly adapt to different working conditions. The model is trained using all normal data, eliminating the need for manual annotation of anomaly types or locations. Anomalies are identified by utilizing the reconstruction error and structural inconsistency probability in the embedding space, demonstrating strong versatility and generalization ability.

[0063] The model trained in this application can be exported via ONNX and run using the ONNX runtime on industrial control platforms without GPUs or PyTorch environments. For different sensor numbers and topologies in various operating scenarios, the model supports dynamic adaptation of the input dimensions without requiring the redevelopment of core logic modules.

[0064] Graph Neural Networks (GNNs) can effectively process graph-structured data, capture the topological relationships between sensors, and extract meaningful features through information transfer and updates between nodes. In traditional machine learning methods, the dependencies between nodes are often ignored, while GNNs explicitly encode these relationships through graph structure modeling, improving the accuracy of anomaly detection. Their attention mechanism formula is as follows:

[0065]

[0066] In the formula, W is the linear transformation matrix, h i and h j These are the features of node i and node j, respectively, h k Let be the feature of node k, representing the hidden representation of node k. α is the learned weight vector of the attention mechanism, N(i) represents the set of neighboring nodes of node i, and || denotes vector concatenation. The formula for aggregating neighboring node information is as follows:

[0067] When aggregating the neighbor information of node i, it is obtained through the characteristics of the neighbor nodes in the software:

[0068]

[0069] In the formula, σ is the activation function (e.g., ReLU), h′ i It is the updated feature of node i.

[0070] Anomaly detection (AE) can effectively capture the internal features of data by learning its low-dimensional representation (encoding) and reconstructing the original data (decoding). In anomaly detection, AE assesses the degree of anomaly of data points based on the reconstruction error. If the reconstruction error exceeds a predetermined threshold, it is judged as an anomaly, as shown in the following formula:

[0071] Encoder: An encoder maps the input x to a latent spatial representation z through a nonlinear transformation.

[0072] z = f enc (x)=σ(W enc x+b enc )

[0073] In the formula, f enc It is the encoding function, W enc It is the encoder's weight matrix, b enc σ is the bias term, and σ is the activation function (e.g., ReLU or sigmoid).

[0074] Decoder: The decoder maps the latent representation z back to the original space (or the same dimension as the original space):

[0075] x′=f dec (z)=σ(W dnc z+b dnc )

[0076] In the formula, f dec It is the decoding function, W dnc It is the weight matrix of the decoder, b dnc is the bias term, and x′ is the input data for reconstruction.

[0077] By combining the two, GNN will be responsible for capturing the relationships between nodes, enabling the model to effectively handle complex topological structures, while AE can learn the low-dimensional representation and reconstruction process of data. The combination of the two not only improves the ability to detect anomalies in sensor data and valve status, but also reduces the dependence on anomaly labeled samples, thereby improving the model's generalization ability.

[0078] In addition, such as Figure 9 As shown, this application also provides an anomaly detection system based on graph neural networks and autoencoders, comprising:

[0079] The data processing module is used to separate the input sensor data and valve control data, construct the sensor differential values, initialize the data, and concatenate the sensor and valve data to form an input matrix.

[0080] The model building and training module is used to encode the input matrix through a graph neural network (GAT), reconstruct sensor and valve data through the GAT structure and decoder, and after standardization, the input data is processed by the GAT encoder. After double decoding and reconstruction, the loss function is calculated, the model parameters are updated, and finally a model that can reconstruct sensor data and valve status is obtained.

[0081] The anomaly detection and judgment module is used to standardize and stitch together the test data, use the trained model for inference, calculate the reconstruction error, and judge anomalies according to the predetermined threshold. The results are then visualized into data and files for easy reading and interpretation by relevant personnel.

[0082] The optimization and multivariate modeling module provides direction for optimizing the model structure, making detection results more accurate. Based on model training and detection, the model structure is further optimized by adjusting the weight coefficients of sensors and valves to achieve more precise anomaly detection. Through the combination of graph neural networks and autoencoders, the model can better identify abnormal patterns and reduce misjudgments caused by ignoring correlations.

[0083] The model evaluation and results output module accurately showcases the model's performance, facilitating researchers' understanding of the model and ensuring its adaptability to most working conditions. After model training, performance is evaluated by assessing reconstruction error and anomaly detection accuracy on the test set. Evaluation metrics include detection accuracy, recall, and F1 score, ensuring the model can stably detect anomalous data under different conditions. All detection data, result files, and statistical information are saved in a designated folder for subsequent analysis and report generation.

[0084] This system proceduralizes the detection of multi-parameter coupled data anomalies in nuclear power process systems, including data processing, model training and storage, test data splicing and standardization, and anomaly detection of test data. It ensures that the parameter stream of the time series is relatively accurate when used, and achieves more efficient and intelligent anomaly detection in nuclear power systems.

[0085] The data processing module provides a standardized matrix for the model input to meet the basic requirements of model building, including input data splitting, data standardization, and concatenation of processed data. Specifically, it first reads the original CSV file, which contains sensor data and valve control status data, and performs preliminary sorting and classification. Sensor data is extracted into sensor_data (n*8), and valve control status data is extracted into valve_data (n*4). Then, the sensor difference value sensor_diff (n*8) is calculated based on the sensor value changes. sensor_data and valve_data are then standardized to ensure a uniform scale of the input data and avoid instability during model training due to differences in numerical ranges. All processed data is concatenated into a unified input matrix input_matrix (n*20) as the input for subsequent model training.

[0086] The model building and training module aims to build a standard model using normal data. By learning the relationships between normal parameters and the changes in parameter values, a complete network structure is constructed to enable anomaly detection when abnormal data is input.

[0087] Specifically, the standardized and concatenated data is fed into the GAT graph neural network. The GAT network calculates the attention weight of each node based on the topological relationships between sensors to effectively capture the correlation between them. The topological relationship structure diagram is as follows: Figure 7 As shown, the topology extraction graph is as follows: Figure 8 As shown, the features of each sensor node are passed to its neighboring nodes to enable information exchange and updates. These updated node features are then input into the decoder for reconstruction using an autoencoder structure.

[0088] Specifically, the standardized and concatenated data is fed into a graph neural network (GAT). The GAT network calculates the attention weight of each node based on the topological relationships between the sensors and updates the node features accordingly, thereby effectively capturing the correlation between sensors. The features of each sensor node are then passed to its neighboring nodes to achieve information interaction and updates.

[0089] In this process, the structure of the autoencoder (AE) works collaboratively. The goal of the autoencoder is to enable the model to reconstruct the original sensor values ​​and valve states based on the input data from the sensors and valves, and then evaluate the model's performance through the reconstruction error. The process is as follows: Figure 3 As shown, the specific steps are as follows:

[0090] (1) GAT Encoding: Data is processed by the encoder of the GAT graph neural network. GAT calculates the attention weight of each node through graph convolution and updates the node's features. The node's features and the feature information of its neighbors are combined to generate a new node representation.

[0091] (2) Autoencoder Reconstruction: The node features processed by GAT are input into the decoder of the autoencoder. The goal of the decoder is to reconstruct the sensor and valve state data. Specifically, the node features updated by GAT are used to reconstruct the sensor and valve states through decoder_sensor and decoder_valve, respectively.

[0092] (3) Loss Function and Optimization: During model training, the reconstruction error is used as the loss function for optimization. SmoothL1Loss is used for sensor state reconstruction, and BCEWithLogitsLoss is used for valve state reconstruction to ensure that their respective structural characteristics are taken into account during data reconstruction. The Adam optimizer is used to update the model parameters, and finally a model that can reconstruct sensor data and valve state is obtained.

[0093] The anomaly detection and judgment module aims to filter out abnormal data and visualize it in the form of charts, helping engineers quickly identify abnormal data and optimize system maintenance and operation. Its process is as follows: Figure 4 As shown. Specifically, in the detection phase, the test data is first standardized and concatenated into an input matrix in the same way as the training data to ensure data format consistency. Next, the trained model is loaded, the test data is input, and inference is performed to calculate the reconstruction error. In the detection process, the reconstruction error is compared with a threshold, which is derived from the error statistics of the training set and is typically set as the mean of the training set error plus three times the standard deviation.

[0094] threshold=np.mean(total_error_train)+3*np.std(total_error_train)

[0095] If the reconstruction error exceeds this threshold, the data point is considered an anomaly. Anomalies detected during the process will be recorded, and their location information will be saved. The results will also be saved to .txt and .csv files for subsequent analysis and processing. The final detection results will be visualized in chart form to help engineers quickly identify anomalies and optimize system maintenance and operation.

[0096] The optimization and multivariate modeling module provides direction for optimizing the model structure, making detection results more accurate. Specifically, based on model training and detection, the model structure is further optimized by adjusting the weight coefficients of sensors and valves to achieve more precise anomaly detection. For example, the sensor weight is currently set to 0.7 and the valve weight to 0.3. This ratio is continuously adjusted through experiments so that the model can better adapt to the anomaly detection needs under different operating conditions. Through joint modeling techniques, the data of sensor and valve states are not only processed independently but also correlated, thereby capturing the complex dependencies between sensors and valves. By combining graph neural networks and autoencoders, the model can better identify anomaly patterns and reduce misjudgments caused by ignoring correlations.

[0097] The model evaluation and results output module aims to accurately demonstrate the model's performance, facilitating researchers' understanding of the model and ensuring its adaptability to most working conditions. Specifically, after model training, the model's performance is evaluated by assessing reconstruction error and anomaly detection accuracy on the test set. Evaluation metrics include detection accuracy, recall, and F1 score, ensuring the model can stably detect anomalous data under different working conditions. The final detection results will undergo statistical analysis, generating detailed reports and presenting them in graphical form to ensure transparency and traceability. All detection data, result files, and statistical information will be stored in a designated folder for subsequent analysis and report generation.

[0098] This application presents a multi-data unsupervised anomaly detection method that integrates graph neural network (GNN) and autoencoder (AE) structures. It can be applied to industrial scenarios with topological structures, such as nuclear power plant process systems. Specifically, it proposes an anomaly detection model based on the fusion of graph neural network and autoencoder to improve the detection accuracy and model generalization ability under complex working conditions.

[0099] The above description is only a specific embodiment of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. An anomaly detection method based on graph neural networks and autoencoders, characterized in that, include: Step 1: Construct a directed graph to represent the sensor network, and fuse the device state encoding with node features to reflect physical connections; Step 2: Extract features, weight and aggregate node information, and automatically learn weights to highlight the contributions of key equipment; Step 3: Reconstruct the sensor state using an autoencoder and perform consistency judgment. Output an abnormal warning when the error exceeds the threshold or the prediction deviation is large. Step 4: Train the model using unsupervised learning and detect anomalies based on reconstruction error and structural inconsistency probability.

2. The anomaly detection method based on graph neural networks and autoencoders according to claim 1, characterized in that, In step 1, the node contains two input dimensions: the current value and the difference value.

3. The anomaly detection method based on graph neural networks and autoencoders according to claim 1, characterized in that, In step 1, the input data is constructed based on graph structure modeling, forming a directed acyclic graph structure of various sensors in the nuclear power plant to reflect the physical connection relationship of the working fluid flow in the process.

4. The anomaly detection method based on graph neural networks and autoencoders according to claim 1, characterized in that, In step 2, a graph attention mechanism is introduced for feature extraction, a graph neural network encoder is constructed, and a multi-layer conditional graph attention module is used to achieve weighted aggregation of information between multiple nodes in the steam flow topology.

5. The anomaly detection method based on graph neural networks and autoencoders according to claim 4, characterized in that, Its attention mechanism formula is: In the formula, W is the linear transformation matrix, h i and h j These are the features of node i and node j, respectively, h k Let be the features of node k, a be the learning weight vector of the attention mechanism, N(i) represent the set of neighboring nodes of node i, and || represent vector concatenation.

6. The anomaly detection method based on graph neural networks and autoencoders according to claim 1, characterized in that, In step 3, the sensor state is reconstructed using an autoencoder structure, and a dedicated valve verification module is introduced to make a consistency judgment on the valve control state. If the sensor error is higher than the normal statistical threshold, or if there is a strong deviation between the valve prediction and the original state, an abnormal warning signal is output.

7. The anomaly detection method based on graph neural networks and autoencoders according to claim 1, characterized in that, In step 4, unsupervised learning is used to adapt to different working conditions. The model is trained with all normal data, and there is no need for manual labeling of abnormality types or locations. The abnormality is judged by the reconstruction error and structural inconsistency probability of the embedding space.

8. The anomaly detection method based on graph neural networks and autoencoders according to claim 1, characterized in that, In anomaly detection, the autoencoder assesses the degree of anomaly of a data point based on the reconstruction error. If the reconstruction error exceeds a predetermined threshold, it is judged as an anomaly, as shown in the following formula: Encoder: An encoder maps the input x to a latent spatial representation z through a nonlinear transformation. z=f ene (x)=σ(W enc x+b enc ) In the formula, f enc It is the encoding function, W enc It is the encoder's weight matrix, b enc σ is the bias term, and σ is the activation function; Decoder: The decoder maps the latent representation z back to the original space. x′=f dec (z)=σ(W dnc Z+b dnc ) In the formula, f dnc It is the decoding function, W dnc It is the weight matrix of the decoder, b dnc is the bias term, and x′ is the input data for reconstruction.

9. An anomaly detection system based on graph neural networks and autoencoders, characterized in that, include: The data processing module is used to separate sensor and valve data, calculate sensor difference values, and concatenate them into an input matrix after initialization. The model building and training module is used to encode the input matrix through GAT, reconstruct sensor and valve data by combining the decoder, calculate loss optimization parameters after standardization, and train the reconstruction model. The anomaly detection and judgment module is used to standardize and stitch test data, use the trained model to infer and calculate the reconstruction error, judge anomalies based on thresholds, and output a visualization result file. The optimization and multivariate modeling module is used to optimize the model structure, adjust the weight coefficients of sensors and valves, and combine GNN and autoencoder to reduce false positives and improve anomaly detection accuracy. The model evaluation and results output module is used to evaluate the reconstruction error and accuracy of the test set and save the results file for analysis.

10. The anomaly detection system based on graph neural networks and autoencoders according to claim 1, characterized in that, The model evaluation and results output module is used to evaluate the performance of the model. Evaluation metrics include detection accuracy, recall, and F1 score.

Images