Classified fault detection method based on graph neural network

Abstract

The invention discloses a classification fault detection method based on a graph neural network, and the method comprises the steps: inputting time series data, projecting the input data to a set dimension through a linear layer, employing a two-layer one-dimensional convolutional neural network to extract time features, dynamically generating an adjacent matrix through a radius according to the Euclidean distance, and based on the adjacent matrix, in each layer of graph convolution, carrying out the classification fault detection of each layer of graph convolution. The method comprises the following steps of: performing spatial aggregation on adjacent node features, learning the weight of important nodes by using an attention mechanism, repeating three-layer GATv2 convolution, adding residual connection and normalization processing into output of each layer, finally obtaining node feature representation of spatial aggregation, and finally converting extracted spatial and temporal features into classification results through a decoder to detect faults; the method solves the problem that the prior art cannot be competent for multi-source data and diversified features in a modern industrial system, combines a memory one-dimensional convolutional neural network and a GNN deep learning method, can extract spatial and temporal features at the same time, and is significantly superior to a traditional statistical method in fault detection accuracy.

Description

Technical Field

[0001] The present invention relates to the field of industrial big data computing, and in particular to a classification fault detection method based on graph neural network. Background Art

[0002] In the field of industrial equipment fault detection, classification algorithms have become the preferred method for processing abundant fault data. Traditional anomaly detection methods, such as rule-based models based on statistical analysis or threshold settings, while practical when data is scarce, often struggle to adapt to the complex and diverse fault modes of modern industrial systems. With advances in data collection technology, the amount of recorded industrial equipment operating data has become increasingly abundant, and the size of fault datasets has increased significantly, providing a foundation for the widespread application of deep learning-based classification algorithms.

[0003] Classification algorithms leverage sufficient fault samples and achieve accurate distinction between normal and abnormal conditions through a data-driven approach. In this scenario, the classification model can directly utilize labeled data for supervised training, gradually learning the hidden characteristic patterns in complex equipment with multiple measurement points, and achieving high fault identification accuracy during prediction.

[0004] Traditional statistical fault detection algorithms, relying on preset thresholds and fixed rules, are typically only able to identify simple anomaly patterns and struggle to adapt to the complex spatiotemporal data associations and diverse fault characteristics found in modern industrial systems. These methods are often limited in accuracy by assumptions of linearity and independence, making it difficult to capture complex inter-device relationships and implicit temporal features. Consequently, they are prone to high rates of false positives and missed detections. Therefore, a classification-based fault detection method based on graph neural networks is needed to address these issues. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a classification fault detection method based on graph neural network, which aims to solve the problem that the existing technology relies on preset thresholds and fixed rules, can only be used for the identification of simple abnormal patterns, and is not competent for the multi-source data and diversified features in modern industrial systems. By combining the memory one-dimensional convolutional neural network and the deep learning method of GNN, it can simultaneously extract spatiotemporal features, automatically learn the correlation between devices and the dynamic change characteristics of fault occurrence, and thus significantly outperform traditional statistical methods in the accuracy of fault detection.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: The classification fault detection method based on graph neural network includes the following steps: S1, input time series data ,in is the batch size, is the time window length, is the number of nodes, each node has Features S2, projects the input data into dimensions through a linear layer : ; in, , represents the time series data tensor after projection; represents the projection weight matrix, represents the projection bias vector; S3 uses a two-layer one-dimensional convolutional neural network to extract temporal features: First layer output: ; Second layer output: ; in, , after normalization and ReLU activation function processing; represents a nonlinear activation function, Indicates neural network normalization; represents one-dimensional convolution; The final features The temporal structure information of the input is retained and mapped to the convolutional feature dimension; S4, build radius neighborhood graph: According to the Euclidean distance Dynamically generate an adjacency matrix: ; Where i, j represent node indexes, and represents the node feature vector; S5, based on the adjacency matrix, spatially aggregates the features of adjacent nodes in each layer of graph convolution, and uses the attention mechanism to learn the weights of important nodes: ; Repeat three layers of GATv2 convolution, add residual connection and normalization to the output of each layer, and finally obtain the spatially aggregated node feature representation ; S6, through the decoder, converts the extracted spatiotemporal features into classification results to detect faults; the fully connected layer decoder classifies the features to obtain the fault classification probability of each node; the softmax function is used to obtain the classification probability: ; ; in, is the probability of each node being faulty or normal; represents the weight matrix, Represents the bias vector.

[0007] Preferably, the linear layer in step S2 inputs feature dimensions Mapped to the same dimension as the CNN output channel .

[0008] Preferably, each layer of the one-dimensional convolutional neural network in step S3 includes a convolution operation, layer normalization and a ReLU activation function, and a residual connection is introduced in the second layer.

[0009] Preferably, the radius in step S4 The adjacency relationship is dynamically adjusted according to the distance in node feature space.

[0010] Preferably, the GATv2 convolutional layer in step S5 adopts an attention mechanism to weightedly aggregate neighborhood node features.

[0011] Preferably, the classifier output is a binary probability: normal 0 or fault 1.

[0012] Preferably, a one-dimensional convolutional neural network is combined with a GNN to realize spatiotemporal feature extraction: One-dimensional convolutional neural networks capture dynamic changes in the time dimension, and GNN models the spatial correlation between nodes.

[0013] Preferably, a classification fault detection device based on a graph neural network comprises: Data acquisition and preprocessing module: used to obtain time series data from multiple measurement points of industrial equipment; The initial projection module is connected to the data acquisition and preprocessing module to perform a linear projection operation and map the input feature dimension to ; A temporal feature extraction module, connected to the initial projection module, comprising two layers of one-dimensional convolutional neural networks, for performing first-layer convolution and second-layer convolution; The spatial relationship modeling module is connected to the temporal feature extraction module and includes: Dynamic graph building unit: based on radius Generate adjacency matrix , Spatial feature aggregation unit: Process the second layer of convolution and , output spatial features ; The fault classification module is connected to the spatial relationship modeling module and includes a fully connected classifier for calculating And output softmax probability ; The device is used to execute the classification fault detection method based on graph neural network described in any one of claims 1-7.

[0014] Preferably, an electronic device includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the classification fault detection method based on graph neural network as described above is implemented.

[0015] Preferably, a computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the graph neural network-based classification fault detection method.

[0016] The beneficial effects of the present invention are as follows: 1. Fusion of spatiotemporal features improves prediction accuracy: Traditional time series prediction models can only capture single features in time or space and are unable to effectively process complex industrial data that incorporates both temporal and spatial correlations. This algorithm combines a graph neural network (GNN) with a long short-term memory (LSTM) network to extract spatial features from control data and temporal features from status data, respectively, thereby achieving a deep fusion of spatiotemporal features. This innovative design enables the model to more accurately model the relationships and temporal dependencies between different measurement points, effectively improving prediction accuracy. It is particularly suitable for multi-point prediction in complex industrial scenarios.

[0017] 2. Parallel processing of spatial features to avoid error accumulation: Existing recursive multi-step prediction methods are often prone to error accumulation, which is particularly significant in long-term prediction tasks. This invention processes the spatial features of the control variables in parallel and incorporates future control variable information during the step-by-step prediction process, thereby reducing the impact of error accumulation on prediction accuracy and enhancing model stability. This structural design significantly improves the model's predictive performance and effectively addresses the error amplification problem common in recursive methods.

[0018] 3. Improved real-time performance and computational efficiency: By introducing a combined architecture of GNN and LSTM, the present invention enables efficient processing of spatiotemporal data while simplifying the structure. Compared to traditional pure LSTM or pure GNN models, the present invention achieves step-by-step processing of different features through a modular design, significantly reducing the model's computational burden. In industrial real-time prediction applications, this efficient structure not only meets the needs of high-frequency data processing, but is also suitable for deployment on edge devices with limited computing resources, providing a more practical solution for intelligent decision-making in industrial sites.

[0019] 4. Model Adaptability and Scalability: The modular design of this invention enables flexible adaptation to diverse measurement point configurations and control variable combinations in various industrial scenarios. Because the GNN and LSTM modules process spatial and temporal features, respectively, the model can be easily expanded to more measurement points or new data sources without redesigning the entire network structure. This adaptability significantly enhances the algorithm's application value in real-world industrial scenarios, reduces the cost of model adjustment and retraining, and enhances its potential for commercialization. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the overall process of the system of the present invention; Figure 2 It is a connection block diagram of the device part of the present invention; Figure 3 It is a schematic diagram of a comparison curve diagram between the embodiment of the present invention and the prior art. DETAILED DESCRIPTION

[0021] Example 1: like Figure 1 As shown in FIG, the classification fault detection method based on graph neural network includes the following steps: S1, input time series data ,in is the batch size, is the time window length, is the number of nodes, each node has Features S2, projects the input data into dimensions through a linear layer : ; in, , represents the time series data tensor after projection; represents the projection weight matrix, represents the projection bias vector; S3 uses a two-layer one-dimensional convolutional neural network to extract temporal features: First layer output: ; Second layer output: ; in, , after normalization and ReLU activation function processing; represents a nonlinear activation function, Indicates neural network normalization; represents one-dimensional convolution; The final features The temporal structure information of the input is retained and mapped to the convolutional feature dimension; S4, build radius neighborhood graph: According to the Euclidean distance Dynamically generate an adjacency matrix: ; Where i, j represent node indexes, and represents the node feature vector; S5, based on the adjacency matrix, spatially aggregates the features of adjacent nodes in each layer of graph convolution, and uses the attention mechanism to learn the weights of important nodes: ; Repeat three layers of GATv2 convolution, add residual connection and normalization to the output of each layer, and finally obtain the spatially aggregated node feature representation ; S6, through the decoder, converts the extracted spatiotemporal features into classification results to detect faults; the fully connected layer decoder classifies the features to obtain the fault classification probability of each node; the softmax function is used to obtain the classification probability: ; ; in, is the probability of each node being faulty or normal; represents the weight matrix, Represents the bias vector.

[0022] Preferably, the linear layer in step S2 inputs feature dimensions Mapped to the same dimension as the CNN output channel .

[0023] Preferably, each layer of the one-dimensional convolutional neural network in step S3 includes a convolution operation, layer normalization and a ReLU activation function, and a residual connection is introduced in the second layer.

[0024] Preferably, the radius in step S4 The adjacency relationship is dynamically adjusted according to the distance in node feature space.

[0025] Preferably, the GATv2 convolutional layer in step S5 adopts an attention mechanism to weightedly aggregate neighborhood node features.

[0026] Preferably, the classifier output is a binary probability: normal 0 or fault 1.

[0027] Preferably, a one-dimensional convolutional neural network is combined with a GNN to realize spatiotemporal feature extraction: One-dimensional convolutional neural networks capture dynamic changes in the time dimension, and GNN models the spatial correlation between nodes.

[0028] Preferably, a classification fault detection device based on a graph neural network comprises: Data acquisition and preprocessing module: used to obtain time series data from multiple measurement points of industrial equipment; The initial projection module is connected to the data acquisition and preprocessing module to perform a linear projection operation and map the input feature dimension to ; A temporal feature extraction module, connected to the initial projection module, comprising two layers of one-dimensional convolutional neural networks, for performing first-layer convolution and second-layer convolution; The spatial relationship modeling module is connected to the temporal feature extraction module and includes: Dynamic graph building unit: based on radius Generate adjacency matrix , Spatial feature aggregation unit: Process the second layer of convolution and , output spatial features ; The fault classification module is connected to the spatial relationship modeling module and includes a fully connected classifier for calculating And output softmax probability ; The device is used to execute the classification fault detection method based on graph neural network described in any one of claims 1-7.

[0029] like Figure 2 As shown, an electronic device includes a processor and a memory, wherein the memory stores a computer program. When the computer program is executed by the processor, the classification fault detection method based on graph neural network as described above is implemented.

[0030] Preferably, a computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the graph neural network-based classification fault detection method.

[0031] Example 2: This embodiment provides an application of a classification fault detection method based on a graph neural network in a specific device. The following is an example of fault detection in a hydropower plant using this method. The specific process is as follows: 1. Application scenarios: Equipment: Francis turbine generator set with a rated power of 100MW; Monitoring nodes: 16 key measuring points; 4 upper frame vibration sensors; 4 lower frame vibration sensors; 4 bearing temperature sensors; 4 cooling water pressure sensors; Input features: ;Data collection is 1 year of operating data, sampling frequency is 100Hz; 2. The comparison method is shown in Table 1 below: Table 1: Comparison with existing technologies:

[0032] 3. Performance comparison data is shown in Table 2 below. The test set consists of 2000 samples: Table 2: Performance comparison with existing methods:

[0033] 4. Typical fault detection comparison is shown in Table 3 below: Table 3: Failure scenario: early bearing wear;

[0034] From Tables 1-3 above and Figure 3 It can be seen that the traditional method independently analyzes vibration and temperature data, ignoring the vibration-temperature coupling effect that is inevitably accompanied by bearing wear; the GATv2 attention mechanism of the present invention automatically weights key associations, and the effect is better than the traditional method.

Claims

1. A classification fault detection method based on graph neural network, characterized in that: The following steps are involved: S1, input time series data ,in is the batch size, is the time window length, is the number of nodes, each node has Features S2, projects the input data into dimensions through a linear layer : ； in, , represents the time series data tensor after projection; represents the projection weight matrix, represents the projection bias vector; S3 uses a two-layer one-dimensional convolutional neural network to extract temporal features: First layer output: ; Second layer output: ; in, , after normalization and ReLU activation function processing; represents a nonlinear activation function, Indicates neural network normalization; Represents one-dimensional convolution; The final features The temporal structure information of the input is retained and mapped to the convolutional feature dimension; S4, build radius neighborhood graph: According to the Euclidean distance Dynamically generate an adjacency matrix: ； Where i, j represent node indexes, and represents the node feature vector; S5, based on the adjacency matrix, spatially aggregates the features of adjacent nodes in each layer of graph convolution, and uses the attention mechanism to learn the weights of important nodes: ； Repeat three layers of GATv2 convolution, add residual connection and normalization to the output of each layer, and finally obtain the spatially aggregated node feature representation ; S6, through the decoder, converts the extracted spatiotemporal features into classification results to detect faults; the fully connected layer decoder classifies the features to obtain the fault classification probability of each node; the softmax function is used to obtain the classification probability: ； ； in, is the probability of each node being faulty or normal; represents the weight matrix, Represents the bias vector.

2. The classification fault detection method based on graph neural network according to claim 1 is characterized in that: The linear layer in step S2 will input the feature dimension Mapped to the same dimension as the CNN output channel .

3. The classification fault detection method based on graph neural network according to claim 2 is characterized in that: In step S3, each layer of the one-dimensional convolutional neural network includes convolution operation, layer normalization and ReLU activation function, and the second layer introduces residual connection.

4. The classification fault detection method based on graph neural network according to claim 3 is characterized in that: The radius in step S4 The adjacency relationship is dynamically adjusted according to the distance in node feature space.

5. The classification fault detection method based on graph neural network according to claim 4 is characterized in that: The GATv2 convolutional layer in step S5 uses an attention mechanism to weightedly aggregate neighborhood node features.

6. The classification fault detection method based on graph neural network according to claim 5 is characterized in that: In step S6, the classifier outputs a binary probability: 0 for normal or 1 for fault.

7. The classification fault detection method based on graph neural network according to claim 6 is characterized in that: Combine one-dimensional convolutional neural network and GNN to realize spatiotemporal feature extraction: One-dimensional convolutional neural networks capture dynamic changes in the time dimension, and GNN models the spatial correlation between nodes.

8. A classification fault detection device based on graph neural network, characterized in that: include: Data acquisition and preprocessing module: used to obtain time series data from multiple measurement points of industrial equipment; The initial projection module is connected to the data acquisition and preprocessing module to perform a linear projection operation and map the input feature dimension to ; A temporal feature extraction module, connected to the initial projection module, comprising two layers of one-dimensional convolutional neural networks, for performing first-layer convolution and second-layer convolution; The spatial relationship modeling module is connected to the temporal feature extraction module and includes: Dynamic graph building unit: based on radius Generate adjacency matrix , Spatial feature aggregation unit: Process the second layer of convolution and , output spatial features ; The fault classification module is connected to the spatial relationship modeling module and includes a fully connected classifier for calculating And output softmax probability ; The device is used to execute the classification fault detection method based on graph neural network described in any one of claims 1-7.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program. When the computer program is executed by the processor, the classification fault detection method based on the graph neural network as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the steps of the classification fault detection method based on graph neural network described in any one of claims 1 to 7 are implemented.

Images