Industrial control system fault diagnosis system based on GAT-2DCNN double-branch feature fusion

Abstract

The invention discloses an industrial control system fault diagnosis system based on GAT-2DCNN double-branch feature fusion. The system comprises a data processing module, a graph attention network GAT, a two-dimensional convolutional neural network 2DCNN and a feature fusion module WFC. According to the invention, time sequence signals are converted into frequency domain signals through the data processing module, and data of a graph structure is established, so that the problem of lack of time correlation in the prior art is effectively overcome; and a feature fusion module WFC is adopted to realize adaptive deep fusion between the topological correlation features extracted by the graph attention network and the spatial-temporal features extracted by the two-dimensional convolutional network, so that the system can significantly improve the accuracy and robustness of fault diagnosis in a complex industrial scene.

Description

Technical Field

[0001] This invention relates to the field of industrial control system fault diagnosis technology, specifically to an industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion. Background Technology

[0002] Modern chemical processes achieve low-cost and high-safety continuous production through distributed control systems (DCS), but numerous accidents have still been reported in recent years. The root cause lies in the fact that operational condition monitoring still heavily relies on operator judgment, making it difficult to guarantee real-time performance and reliability. Furthermore, delayed responses to anomalies can easily escalate safety incidents. Fault Detection and Diagnosis (FDD) technology, through real-time condition monitoring, can effectively reduce unplanned equipment downtime and accident losses. FDD technology can extend equipment status to multi-classification tasks, not only detecting the presence of faults but also identifying their types. Compared to model-driven and knowledge-driven methods that rely on physicochemical mechanisms, data-driven methods significantly reduce reliance on prior knowledge by constructing classification models through analyzing historical data features. Traditional intelligent diagnostic methods mainly involve two steps: feature extraction using signal processing methods and fault classification using machine learning methods. However, traditional intelligent diagnostic methods often rely on expert knowledge and struggle to accurately describe complex industrial processes.

[0003] Currently, deep learning-based intelligent monitoring systems have been applied in areas such as industrial equipment fault diagnosis and process optimization. However, they have significant limitations when processing multi-sensor signals. Standard convolution operations only perform a weighted summation of sensor measurements and the convolution kernel, neglecting the interaction relationships between sensor nodes. This results in standard convolution operations being insufficient for mining the topological correlation features inherent in industrial systems. This invention provides an industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion to solve the above problems. Summary of the Invention

[0004] This invention provides a fault diagnosis system for industrial control systems based on GAT-2DCNN dual-branch feature fusion, which improves the accuracy of fault diagnosis through multi-level feature fusion.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: like Figure 1 As shown, an industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion includes a data processing module, a graph attention network (GAT), a two-dimensional convolutional neural network (2DCNN), and a feature fusion module (WFC). The data processing module is used to receive data collected by the sensor and output the processed data to the graph attention network (GAT) and the two-dimensional convolutional neural network (2DCNN). The graph attention network (GAT) and the two-dimensional convolutional neural network (2DCNN) are configured in parallel, and each has two layers. The Graph Attention Network (GAT) is used to process structured data between sensors, extracting graph features through multi-layer GAT and edge pooling operations. The 2D Convolutional Neural Network (2DCNN) is used to analyze the time-frequency representation of sensor signals, utilizing multi-layer 2DCNN and max pooling to capture local features of the sequence. The graph attention network (GAT) and the two-dimensional convolutional neural network (2DCNN) are used to receive graph structure data and temporal data output by the data processing module, respectively, and extract intermediate features. The feature fusion module WFC is located between the graph attention network GAT and the two-dimensional convolutional neural network 2DCNN, after the first layer of the graph attention network GAT and the first layer of the two-dimensional convolutional neural network 2DCNN. It receives and fuses the intermediate features output by the first layer of the graph attention network GAT and the first layer of the two-dimensional convolutional neural network 2DCNN, and feeds back the fused features to the second layer of the graph attention network GAT and the second layer of the two-dimensional convolutional neural network 2DCNN, respectively. The second layer of the graph attention network (GAT) and the second layer of the two-dimensional convolutional neural network (2DCNN) output the fused features to the fully connected layer, thereby achieving accurate classification of different fault types.

[0006] Furthermore, the graph attention network (GAT) introduces an attention mechanism, allowing the model to dynamically assign different importance weights to each neighbor when aggregating neighbor information, thus capturing the complex relationships between sensors more flexibly. The GAT first performs a linear transformation on the node features, then calculates the attention coefficients of neighbor nodes to the center node through a shared attention mechanism. These attention coefficients are normalized using a softmax function and then used as weights to perform a weighted summation of the transformed features of the neighbor nodes, thereby updating the feature representation of the center node. The attention coefficient of the graph attention network GAT is calculated using formula (1). (1), In the formula, h i and h j Let be the features of sensor i and sensor j, respectively; W be the weight matrix of the linear transformation; α be the shared learnable weight vector; and || denotes the concatenation operation. Representing sensor i itself and other sensors associated with it, LeakyReLU is the activation function; Equation (1) calculates and normalizes the attention weight of node j to node i.

[0007] The graph attention network GAT employs a multi-head attention mechanism to enhance the model's representation ability across different feature subspaces. Each attention head independently learns different patterns of node relationships, and finally integrates multi-perspective information through feature concatenation, specifically calculated using formula (2). (2), In the formula, K is the number of attention heads. and W (K) These are the normalized attention coefficients and weight matrix of the Kth attention head, respectively, and σ is a non-linear activation function; Formula (2) concatenates the outputs of multiple independent attention heads to form the final new feature representation of node i.

[0008] Furthermore, sensor signals exhibit specific local patterns in the frequency domain. These patterns are closely related to the device's operating state. The convolutional kernels of a 2D Convolutional Neural Network (2DCNN) can effectively capture these local patterns, thereby identifying key information in the data. The core operation of the 2DCNN is 2D convolution. 2D convolution involves a learnable convolutional kernel sliding across the input feature map and calculating a weighted sum of local regions to extract features. The convolution operation is specifically calculated using formula (3). (3), In the formula, Y i,j,k X represents the activation value at position (i,j) of the k-th channel of the output feature. i+u,j+v,c K represents the pixel value of the c-th input channel at position (i+u,j+v) after padding. u,v,c,k B represents the weight of the k-th convolutional kernel at position (u,v) corresponding to the c-th input channel. K H is the bias of the k-th output channel. K and W K C represents the height and width of the convolution kernel, respectively. in This represents the number of channels in the input feature map.

[0009] like Figure 4 As shown, the feature fusion module WFC further integrates 2D convolutional features xc1 with graph convolutional features xg1 through parallel and complementary adaptive weighted feature fusion and channel-expansion-based feature fusion. The parallel and complementary adaptive weighted feature fusion and channel-expansion-based feature fusion work synergistically, enabling the model to enhance the contribution of feature channels more relevant to the target task and enrich feature representation by expanding channel dimensions, thereby improving the overall discriminative ability of the model.

[0010] Furthermore, the feature adaptive weighted fusion is used to fuse 2D convolutional feature xc1 to graph convolutional feature xg1; by learning the correlation between 2D convolutional feature xc1 and graph convolutional feature xg1, the contribution of different channels in feature xc1 is dynamically adjusted, and finally feature xc1 and feature xg1 are combined.

[0011] The fusion process is as follows: The last dimension of the input feature xc1 is adjusted through a 1D convolutional layer to match the dimension of feature xg1, so that the shape of the adjusted feature xc1 is consistent with the last dimension of feature xg1, facilitating subsequent similarity calculation and weighting operations. After the adjusted features xc1 are flattened in their respective dimensions, their correlation is measured by calculating the cosine similarity between them. The similarity calculation result is normalized by the softmax function to generate a set of attention weights W. The generated attention weights are applied element-wise to the adjusted features xc1. The weighted features xc1 are summed along the channel dimension to obtain a summarized feature representation. The summarized features are added element-wise to features xg1 to complete feature fusion. Through weighted summation, the model can dynamically allocate attention according to the correlation between features, enhancing the contribution of xc1 features that are more relevant to the target feature xg1, thereby improving the expressive power of the fused features.

[0012] The cosine similarity function assesses the proximity of two vectors in a direction, thus quantifying the similarity between each channel of xc1 and the overall features of xg1. The cosine similarity is specifically calculated using formula (4). (4), In the formula, A and B are two vectors, A·B is the dot product of vectors A and B, ‖A‖ and ‖B‖ are the norms of vectors A and B, and the cosine similarity value ranges from -1 to 1: 1 means that the two vectors have the same direction, 0 means that they have no similarity, and -1 means that the two vectors have completely opposite directions. Attention weight , where k is the number of the kth channel of feature xc1; the attention weight reflects the degree of contribution of each channel in feature xc1 to feature xg1.

[0013] Furthermore, the specific fusion process of the feature fusion based on channel expansion is as follows: First, the input feature xg1 is reshaped into a batch processing form. Then, a 1D convolutional layer is introduced to transform the feature dimension of feature xg1, making it consistent with the last dimension of feature xc1. Next, the transformed feature xg1 is further dimensionally adjusted to make it compatible with feature xc1 in the channel dimension, allowing it to be added as a new channel. Finally, the adjusted feature xg1 is concatenated with feature xc1 along the channel dimension to complete feature fusion. This channel-expanded feature fusion method enriches the feature representation by expanding the channel dimension without changing the original feature space, enabling the model to utilize complementary information from xc1 and xg1 simultaneously, thereby enhancing the model's discriminative ability.

[0014] like Figure 2 As shown, the data processing module further includes a sliding window processing technology module, a Fourier transform module, and a graph generation module. First, the sliding window technology is used to generate time series slices of the data. Then, a Fourier transform is applied to each slice to extract the frequency domain features of the sensor data. Finally, the edges of the node data are constructed by sequential connection, and a Gaussian kernel function is used to assign corresponding weights to each edge.

[0015] The graph generation module constructs frequency domain information into graph structure data.

[0016] like Figure 3 As shown, the sliding window processing module is further used to receive the raw data collected by the sensor, normalize the collected raw data to obtain time series data, and then extract samples from the time series data to form time series slice data as input to the graph attention network GAT and the two-dimensional convolutional neural network 2DCNN.

[0017] In industrial control processes, various types of sensors are typically installed. The data collected by each sensor is generally presented in the form of a time series, which reflects the operating status of the system. To reduce the scale differences between the characteristics of different sensors, it is necessary to normalize the data. The sliding window processing module normalizes the original data using formula (5). (5), In the formula, x C The current value, σ is the mean of the current sequence, σ is the variance of the current sequence, and C is the number of data points in the current sequence; To construct the time dependency of sensor readings, a sliding window technique is used to extract samples from the time series data. Specifically, a fixed-size window is moved along the time axis to convert the time series data into time slices, which are then labeled according to the fault type corresponding to each slice. After normalization, the raw data forms a multi-sensor time series dataset C and a single-sensor time series dataset c. m , (6), (7), In the formula, c m This represents the time series data collected by each sensor, where M represents the number of sensors, and x represents the time series data collected by each sensor. m,t The data collected by the m-th sensor at time t represents the total number of data collected by that sensor. The time-series slice data output by the sliding window processing module forms a set D = ω(∂, λ, C), where set D is defined as... (8), In the formula, d n This represents the sample output by the sliding window; Each sample contains joint observations from multiple sensors within a time window, as shown in formula (9). (9), In the formula, M represents the number of sensors, and x m,t Let denot be the data collected by the m-th sensor at time t, λ represent the movement step of the sliding window, ∂ represent the width of the sliding window, and N represent the number of samples output by the sliding window.

[0018] This can be obtained by processing with a sliding window. The nth sample d is used as the input to the graph neural network, and the nth sample d is constructed by the sliding window. n It includes not only data from M different sensors, but also the variations of these data over time n∂ to n∂+λ.

[0019] Furthermore, while time-series data can be used as input to a two-dimensional convolutional neural network (2DCNN), it cannot be directly used as input to a graph attention network (GAT). Therefore, it must be converted into graph-structured data. The Fourier transform module transforms the time-series slice data into frequency domain information through time-frequency conversion, effectively extracting the periodic components of the signal.

[0020] The specific operation involves first decomposing the original sequence into two smaller discrete Fourier transform subproblems using the Fast Fourier Transform (FFT), and then recursively decomposing each subproblem until the sequence length is reduced to the required length, at which point the degree of discrete Fourier transform can be directly calculated. Finally, the sequences are merged using a butterfly operation to obtain the discrete Fourier transform result of the entire sequence.

[0021] The basic principle of the Discrete Fourier Transform is shown in Equation (10). (10) In the formula, X K It is the Kth frequency point in the frequency domain. is the rotation factor, and i is the imaginary unit.

[0022] Furthermore, the graph generation module abstracts the sensor frequency domain features transformed by the Fourier transform module into a graph structure, which serves as the input to the graph attention network (GAT). Specifically, it treats the frequency domain information of each sensor as a node, connects the nodes sequentially, uses a coordinate format to represent the adjacency matrix, and calculates the edge weights between nodes using a Gaussian kernel function to construct the frequency domain information into graph structure data.

[0023] The beneficial effects of this invention are as follows: A dynamic weighting and feature splicing fusion module is set up. It realizes the fusion of features between different modalities through feature adaptive weighting fusion and feature fusion based on channel expansion. The two fusion strategies work together to enhance the contribution of feature channels that are more relevant to the target task, and enrich the feature representation by expanding the channel dimension, thereby improving the overall discriminative ability of the model. In the data processing stage, the sliding window technique is used to capture data features at different time periods; then, Fourier transform is used to convert the time-domain signal into the frequency domain, thereby comprehensively extracting the periodicity and frequency characteristics of the data; finally, a graph structure is constructed for the data. The Graph Attention Network (GAT) and the 2D Convolutional Neural Network (2DCNN) form a dual-branch, multi-stage feature fusion architecture. This architecture fully combines the advantages of graph-structured data and sequence and image data to collaboratively complete classification tasks and effectively improve classification accuracy. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a schematic diagram of the data processing module architecture of the present invention; Figure 3 This is a schematic diagram of the sliding window processing procedure of the present invention; Figure 4 This is a schematic diagram illustrating the principle of the dynamic weighting and feature splicing fusion module of the present invention; Figure 5 This is a schematic diagram of the TE platform simulation process of the present invention; Figure 6 This is a list of operating conditions for the TE dataset of this invention; Figure 7 This is a list of operating conditions for the water tank dataset of this invention; Figure 8 This is a schematic diagram of the results of Experiment 1 of the TE dataset of this invention; Figure 9 This is a schematic diagram of the results of Experiment 2 of the TE dataset of this invention; Figure 10 This is a schematic diagram of the experimental results of the water tank dataset of the present invention; Figure 11 This is a schematic diagram of the branch ablation experimental results of the present invention; Figure 12 This is a schematic diagram of the ablation experiment results of the feature fusion module of the present invention. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0027] In a specific embodiment of the present invention, the effectiveness and accuracy of the system are verified by experiments. When conducting experiments, the TE dataset and the water tank dataset are used as datasets. After setting relevant experimental parameters, the experiments are carried out, and the effectiveness and accuracy of the system are verified according to the experimental structure.

[0028] like Figure 5 , 6As shown, the data generated by the TE simulation platform is characterized by time-varying, strong coupling, and nonlinearity, and is widely used to test control and fault diagnosis models for complex industrial processes. The TE dataset involves 41 measurement variables and 11 operational variables, and pre-defines 21 fault types and 1 normal operating state. Each training set contains 500 samples, while each test set contains 960 samples. The first 20 samples in each training set and the first 160 samples in each test set are collected under the normal operating state. The 21 fault types are categorized as step change, random change, slow drift, and valve sticking. In this embodiment, all operating states were selected to construct the system. During construction, normal samples from the 21 fault data were first removed. Then, 1280 samples were collected under each state, with 480 samples in the training set and 800 samples in the test set. All samples were first labeled, then randomly shuffled and normalized, and the training and test sets were divided according to a 67% and 33% ratio, respectively.

[0029] like Figure 7 As shown, the water tank dataset is a simulation model built based on the oil storage process, simulating several possible faults during this process. The dataset uses 17 features, including five liquid level values ​​for tanks HT1-HT5, five pressure values ​​for pipes PT1-PT5, flow rate QT1 for the inlet pipe and QT2 for the outlet pipe, and temperature values ​​TT1-TT5 for the five tanks. A total of eight fault types with simulated parameter deviations and one normal state are preset; there are 178,256 normal samples and 66,607 abnormal samples. The system samples every second. After data preprocessing, all detected abnormal samples are retained. One-fifth of the total number of abnormal samples are randomly selected from all normal samples to balance the class distribution and avoid bias in model training. The selected abnormal samples are merged with normal samples, and the model's generalization ability is enhanced by random shuffling. The dataset is then divided into training and test sets at a ratio of 67% and 33%, respectively.

[0030] The experimental parameters in this embodiment include data preprocessing parameters and system architecture parameters.

[0031] When setting the data preprocessing parameters, the window width for the TE dataset was set to 12 and the step size to 1; the window width for the water tank dataset was set to 12 and the step size to 3.

[0032] When setting the system architecture parameters, two sets of two-dimensional convolutional layers, Conv1 and Conv2, were constructed in the CNN branch. Image features were extracted layer by layer through cascaded 3×3 convolutional kernels, stride of 1, padding of 1, batch normalization, and ReLU activation function. An adaptive max pooling layer was used after each convolutional block. Conv1 downsampled the feature map to a fixed size (52, 8), while Conv2 further downsampled it to (26, 4) and (8, 2), enabling the model to handle inputs of different sizes and gradually focus on key features. Due to the fusion of feature information from the graph neural network, the first convolutional layer of Conv2 received 33 input channels.

[0033] The GCN branch employs a two-layer graph attention network, GATConv, to progressively map graph features from the input dimension and maintain them at 1024 dimensions. It enhances feature representation through batch normalization and the ReLU activation function. An edge pooling layer is introduced for graph downsampling to effectively aggregate node information and extract hierarchical graph features. Finally, a fully connected classifier receives the concatenated features, passes them through a hidden layer and regularization, and outputs class probabilities using the Softmax activation function, thus achieving multi-class classification.

[0034] The Adam optimizer was selected for system optimization, with an initial learning rate of 0.001. To improve model performance and stability, the learning rate was halved in a stepwise manner at the 100th and 200th training epochs. The entire training process was planned to run for 300 epochs, with the fault diagnosis rate (FDR) and false positive rate (FPR) used as the core evaluation metrics. TP, True Positive: The number of correctly diagnosed fault samples. FN, False Negative: The number of false negatives, i.e., the number of missed diagnoses of faulty samples. FP, False Positive: The number of false positives, i.e., the number of normal samples that are falsely reported. TN, True Negative: The number of correctly identified normal samples.

[0035] Experiments were conducted using the TE dataset, with comparisons made to the following networks: Long Short-Term Memory (LSTM), a hybrid architecture of convolutional neural networks and LSTM (CNN-LSTM), a deep convolutional neural network enhanced by GRU (GRU-EDCNN), an improved Transformer model (TT) combining self-attention and target attention mechanisms, and a network (TGGL) that combines physical knowledge with graph neural networks.

[0036] like Figure 8 As shown, the results demonstrate that our system consistently exhibits excellent classification accuracy across a wide range of fault categories. Furthermore, our system achieves significant performance improvements not only in generally easy-to-classify categories but also in categories that are typically difficult to identify. For example, in categories 0, 3, 9, and 10, other methods such as LSTM, CNN-LSTM, and GRU-EDCNN show significantly lower accuracy, while our system maintains high diagnostic accuracy.

[0037] like Figure 9 As shown, the effectiveness of the proposed system is further verified by using the mechanistic model A-CM based on anomaly prior knowledge and convolutional neural networks, convolutional neural networks (CNN), and graph convolutional networks (GCN) as experimental baselines. Experiments on the TE dataset demonstrate that the proposed system maintains a high fault diagnosis rate across various fault types and can achieve accurate classification for multiple fault types.

[0038] like Figure 10 As shown in the experiment using the water tank dataset, the results indicate that the average fault diagnosis rate of this system increased by 2.34% and 1.22% compared to A-CM and CNN, respectively, and by 34.22% compared to GCN. This suggests that the proportion of misidentifications among all samples identified as abnormal by the model is lower; the false positive rate decreased by 2.9%-0.13% compared to the three methods, meaning that the probability of misidentifying a sample as another sample is lower.

[0039] like Figure 11 As shown, an ablation experiment was conducted to evaluate the independent contributions of each branch of GAT and 2DCNN and their fusion effect. Specifically, simplified systems were constructed, one retaining the 2DCNN branch and removing the GAT branch, and the other retaining the GAT branch and removing the 2DCNN branch. The simplified systems were trained uniformly for 200 epochs, and their performance was evaluated by comparing the accuracy of the simplified systems under different sliding window step sizes. The results show that the proposed system achieves an accuracy of 98.50% on the TE dataset and 88.66% on the water tank dataset. The integration of GAT and 2DCNN significantly enhances the predictive ability of a single model on different datasets and with different λ values, demonstrating the effectiveness of the dual-branch hybrid architecture, overcoming the feature representation limitations of a single model, and improving diagnostic accuracy under complex conditions.

[0040] like Figure 12As shown, an ablation experiment was conducted to evaluate the impact of the Feature Fusion Module (WFC) on the system, comparing the performance of the system with and without the WFC on the TE dataset and the water tank dataset. The results show that the WFC significantly improves system performance. On both the TE and water tank datasets, regardless of the value of the sliding window step size λ, the system with the WFC exhibits higher accuracy and greater stability. Furthermore, at larger λ values, the system demonstrates stronger robustness and a more pronounced performance advantage. This demonstrates that the WFC effectively fuses features from graph neural networks and 2D CNNs, thereby enhancing the model's feature representation and fault diagnosis capabilities.

[0041] This system converts time-series signals into frequency-domain signals and establishes graph-structured data through a data processing module, effectively overcoming the problem of missing time correlation in existing technologies. Furthermore, the WFC feature fusion module achieves adaptive deep fusion between topological correlation features extracted by the graph attention network and spatiotemporal features extracted by the two-dimensional convolutional network, enabling the system to significantly improve the accuracy and robustness of fault diagnosis in complex industrial scenarios.

[0042] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A fault diagnosis system for industrial control systems based on GAT-2DCNN dual-branch feature fusion, characterized by: It includes a data processing module, a graph attention network (GAT), a two-dimensional convolutional neural network (2DCNN), and a feature fusion module (WFC). The data processing module is used to receive data collected by the sensor; The graph attention network (GAT) and the two-dimensional convolutional neural network (2DCNN) are configured in parallel and each has multiple layers. The graph attention network (GAT) and the two-dimensional convolutional neural network (2DCNN) are used to receive graph structure data and temporal data output by the data processing module, respectively, and extract intermediate features. The feature fusion module WFC is located between the graph attention network GAT and the two-dimensional convolutional neural network 2DCNN. It is used to fuse the intermediate features extracted by the graph attention network GAT and the two-dimensional convolutional neural network 2DCNN, and feed the fused features back to the subsequent stages of the graph attention network GAT and the two-dimensional convolutional neural network 2DCNN respectively. The graph attention network (GAT) and the two-dimensional convolutional neural network (2DCNN) output the fused features to the fully connected layer.

2. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 1, characterized in that: The graph attention network (GAT) first performs a linear transformation on the node features, and then calculates the attention coefficients of neighboring nodes to the center node through a shared attention mechanism. After the attention coefficients are normalized by the softmax function, they are used as weights to perform a weighted summation on the transformed features of neighboring nodes, thereby updating the feature representation of the center node. The attention coefficient of the graph attention network GAT is calculated using formula (1). （1）， In the formula, h i and h j Let be the features of sensor i and sensor j, respectively; W be the weight matrix of the linear transformation; α be the shared learnable weight vector; and || denotes the concatenation operation. LeakyReLU is the activation function, representing sensor i itself and other sensors associated with it. The graph attention network GAT employs a multi-head attention mechanism to enhance the model's representation ability across different feature subspaces. Each attention head independently learns different patterns of node relationships, and finally integrates multi-perspective information through feature concatenation, specifically calculated using formula (2). （2）， In the formula, K is the number of attention heads. and W (K) These are the normalized attention coefficients and weight matrix of the Kth attention head, respectively, and σ is the non-linear activation function.

3. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 1, characterized in that: The two-dimensional convolutional neural network (2DCNN) extracts features by sliding a learnable convolutional kernel across the input feature map and calculating a weighted sum of local regions. The convolution operation is specifically calculated using formula (3). （3）， In the formula, Y i,j,k X represents the activation value at position (i,j) of the k-th channel of the output feature. i+u,j+v,c K represents the pixel value of the c-th input channel at position (i+u,j+v) after padding. u,v,c,k B represents the weight of the k-th convolutional kernel at position (u,v) corresponding to the c-th input channel. K H is the bias of the k-th output channel. K and W K C represents the height and width of the convolution kernel, respectively. in This represents the number of channels in the input feature map.

4. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 1, characterized in that: The feature fusion module WFC integrates 2D convolutional features xc1 with graph convolutional features xg1 through parallel and complementary adaptive weighted feature fusion and channel-expansion-based feature fusion.

5. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 4, characterized in that: The adaptive weighted fusion of features is used to fuse 2D convolutional features xc1 onto graph convolutional features xg1; The fusion process is as follows: The last dimension of the input feature xc1 is adjusted through a 1D convolutional layer to match the dimension of feature xg1. The shape of the adjusted feature xc1 is consistent with the last dimension of feature xg1. After the adjusted features xc1 are flattened in their respective dimensions, their correlation is measured by calculating the cosine similarity between them. The similarity calculation result is normalized by the softmax function to generate a set of attention weights W. The attention weights are applied element-wise to the adjusted feature xc1. The weighted feature xc1 is summed along the channel dimension to obtain a summarized feature representation. The summarized feature is added to feature xg1 element-wise to complete the feature fusion. The cosine similarity is specifically calculated using formula (4). （4）， In the formula, A and B are two vectors, A·B is the dot product of vectors A and B, ‖A‖ and ‖B‖ are the norms of vectors A and B, and the cosine similarity value ranges from -1 to 1: 1 means that the two vectors have the same direction, 0 means that they have no similarity, and -1 means that the two vectors have completely opposite directions. Attention weight , where k is the kth channel number of feature xc1.

6. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 4, characterized in that: The specific fusion process of the feature fusion based on channel expansion is as follows: First, the input feature xg1 is reshaped and converted into a batch processing form. Then, a 1D convolutional layer is introduced to transform the feature dimension of feature xg1 so that it is consistent with the last dimension of feature xc1. Next, the transformed feature xg1 is dimensionally adjusted so that feature xg1 is compatible with feature xc1 in the channel dimension. Finally, the adjusted feature xg1 is concatenated with feature xc1 along the channel dimension to complete feature fusion.

7. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 1, characterized in that: The data processing module includes a sliding window processing technology module, a Fourier transform module, and a graph generation module; The sliding window processing module is used to receive raw data collected by the sensor, normalize the collected raw data to obtain time series data, and then extract samples from the time series data to form time series slice data as input to the graph attention network GAT and the two-dimensional convolutional neural network 2DCNN. The Fourier transform module converts the time series slice data into frequency domain information; The graph generation module constructs frequency domain information into graph structure data.

8. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 7, characterized in that: The sliding window processing module normalizes the original data using formula (5). （5）， In the formula, x C The current value, σ is the mean of the current sequence, σ is the variance of the current sequence, and C is the number of data points in the current sequence; After normalization, the raw data is divided into a multi-sensor time series data set C and a single-sensor time series data set c. m , （6）， （7）， In the formula, c m This represents the time series data collected by each sensor, where M represents the number of sensors, and x represents the time series data collected by each sensor. m,t The data collected by the m-th sensor at time t represents the total number of data collected by that sensor. The time-series slice data output by the sliding window processing module forms a set D, and the set D is defined as follows: （8）， （9）， In the formula, M represents the number of sensors, and x m,t Let denot be the data collected by the m-th sensor at time t, λ represent the movement step of the sliding window, ∂ represent the width of the sliding window, and N represent the number of samples output by the sliding window.

9. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 7, characterized in that: The Fourier transform module first decomposes the original sequence into two smaller discrete Fourier transform subproblems using fast Fourier transform, and then recursively decomposes each subproblem until the sequence length is reduced to the required length. Finally, the sequences are merged using butterfly operations.

10. The industrial control system fault diagnosis system based on GAT-2DCNN dual-branch feature fusion as described in claim 1, characterized in that: The graph generation module treats the frequency domain information of each sensor as a node, connects the nodes sequentially, represents the adjacency matrix using a coordinate format, and calculates the edge weights between nodes using a Gaussian kernel function to construct graph structure data from the frequency domain information.

Images