Pipeline anomaly detection method based on VQSAE and DCGCN

Abstract

The invention discloses a pipeline anomaly detection method based on VQSAE and DCGCN, and the method comprises the steps: 1, collecting the time series data of a sensor on a pipeline, constructing a data set, and carrying out the preprocessing of the data set; 2, constructing an improved VQSAE model, and introducing a new activation function SELU to carry out local time feature extraction on the data; step 3, constructing a DCGCN global space-time relation model according to features extracted by the VQSAE model, and introducing a graph attention mechanism GAM to perform space-time graph convolution operation; and 4, detecting the pipeline anomaly through the DCGCN model, and outputting the pipeline anomaly detection accuracy. According to the method, data are fused through the VQSAE and the DCGCN, complex pipeline topology and multi-modal data can be processed, the false alarm rate and the missing report rate are reduced, more accurate, efficient, high-robustness and extensible pipeline anomaly detection is achieved, managers are helped to recognize potential pipeline risks earlier, maintenance decisions are optimized, and the pipeline anomaly detection efficiency is improved. And economic loss and potential safety hazards caused by faults are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline detection, and in particular to a pipeline anomaly detection method based on VQSAE and DCGCN. Background Art

[0002] Pipeline anomaly detection is a key area in industrial, urban and environmental monitoring, and is widely used in monitoring long-distance transportation systems such as oil, gas, water supply, and heating. Pipeline failures such as leakage, corrosion, or abnormal pressure may cause significant economic losses, environmental pollution, and even endanger safety. Therefore, it is crucial to detect and locate abnormal behavior in pipelines in a timely and effective manner.

[0003] Traditional leakage detection methods include: using CNN-MLC, GCN and IMFs models to detect anomalies in pipelines. Among them, although CNN-MLC can process local features, it is not good at capturing global information in complex network structures. Especially in multi-label classification, it is necessary to process the correlation between different labels, and CNN performs poorly. Although GCN can use the graph structure of the pipeline to capture the topological information of each node in the pipeline network and the dependencies between nodes, its representation ability is insufficient for complex high-dimensional data, especially when there are unstructured patterns in the data (such as local sensor data), and when the scale of the pipeline network is large, the computational complexity and storage requirements of GCN will increase significantly. IMFs have good processing capabilities for non-stationary signals (such as noise signals or nonlinear vibrations in pipelines), but perform poorly when processing the global topological information of pipeline networks. Summary of the invention

[0004] In view of the technical problems caused by the technical defects of the above models, this technical solution provides a pipeline anomaly detection method based on VQSAE and DCGCN. VQSAE can effectively reduce the dimension of sensor signals, reduce data noise and extract key features, while maintaining the sparsity of features; at the same time, VQSAE generates discrete feature representations, which helps to capture discrete patterns in data (such as abnormal events) and has good robustness; DCGCN, based on traditional GCN, strengthens the feature transfer between different layers through a dense connection mechanism, can capture multi-level information in complex network structures, take into account local and global features, and reduce the loss in information transfer, while improving computational efficiency, which is suitable for processing large-scale pipeline networks. VQSAE processes different types of sensor data, compresses and extracts high-dimensional features, effectively removes noise and extracts key features, while DCGCN further utilizes the topological information of the pipeline network to perform structured processing on these features, and combines local and global topological information to improve the accuracy and robustness of anomaly detection; it can effectively make up for the shortcomings of the above models; and effectively solve the above problems.

[0005] The present invention is achieved through the following technical solutions:

[0006] A pipeline anomaly detection method based on VQSAE and DCGCN specifically includes the following steps:

[0007] Step 1: Collect time series data from sensors on the pipeline, build a data set, and preprocess the data set;

[0008] Step 2: Build an improved VQSAE model and introduce a new activation function SELU to extract local time features of the data; the specific operation method includes the following steps:

[0009] Step 2.1: VQSAE is a random autoencoder based on vector quantization, which encodes the input data into a discrete feature space for representation, and obtains an improved VQSAE model. The improved VQSAE model includes the following structures: encoder, vector quantizer and decoder;

[0010] Step 2.2: Calculate the loss function and extract discrete features z q (x);

[0011] Step 3: Based on the features extracted by the VQSAE model, a DCGCN global spatiotemporal relationship model is constructed, and the graph attention mechanism GAM is introduced to perform spatiotemporal graph convolution operations;

[0012] Step 3.1: The discrete features z extracted by VQSAE q (x) is used as the input of the spatiotemporal graph convolutional network; assuming the input data is:

[0013] X∈R N×T×F ,

[0014] Among them, N represents the number of nodes in the graph, T represents the time step, and F represents the feature dimension; after VQSAE processing, the encoding feature Zq∈R N×T×F , used for graph convolution;

[0015] Step 3.2: In the spatiotemporal graph convolution network DCGCN, the key to graph convolution is to perform graph convolution operations on the features of each time step t, using the adjacency matrix A of the graph to capture spatial dependencies; the basic graph convolution operation can be expressed as:

[0016]

[0017] Among them, H (l+1) is the feature matrix of the l+1th layer; is the feature matrix of the lth layer; F l represents the feature dimension of the lth layer, is the normalized adjacency matrix, D is the degree matrix; W(l) is the weight matrix of the graph convolutional layer, σ(·) is the SELU nonlinear activation function;

[0018] In order to model time dependencies, dilated convolution is introduced to expand in the time dimension so that the convolution operation can capture the dependencies of different time steps; the spatiotemporal graph convolution is expressed as:

[0019]

[0020] in, is the weight of the spatial convolution, is the weight of the temporal convolution, which captures the dependencies of different time steps Δt through dilated convolution;

[0021] Step 3.3: Introduce the graph attention mechanism GAM; the goal of the attention mechanism is to dynamically calculate the attention weights between each node and its neighbors to adaptively adjust the calculation process of the graph convolution;

[0022] Step 4: Detect pipeline anomalies through the DCGCN model and output pipeline anomaly detection results.

[0023] Further, the encoder described in step 2.1: Assume that the input is a vector x∈R d , where d is the feature dimension. The encoder maps the input data to the latent space, usually through a convolutional layer or a fully connected layer, and introduces a new activation function SELU for nonlinear transformation; the formula of the SELU function is:

[0024]

[0025] Where: λ=1.0507, α=1.67326, x is the output when it is greater than 0, αe x -α is the output when it is less than 0;

[0026] Then: The output z of the encoder is e (x) is expressed as:

[0027] z e (x) = SELU(W e x+b e )

[0028] Among them, W e and b e are the weights and biases of the encoder, and SELU is used as the activation function to ensure the self-regularization characteristics of the encoder output.

[0029] Further, the vector quantizer described in step 2.1: the encoder output z e(x) is mapped to a discrete codebook index through a vector quantizer. First, find the codebook index corresponding to z. e (x) The nearest vector z q (x):

[0030]

[0031] Among them, C is the codebook set, e i is each vector in the codebook.

[0032] Furthermore, the decoder described in step 2.1: converts the quantized vector z q (x) is input to the decoder and reconstructed into an estimate of the input data Also using the SELU activation function, the decoder output It can be expressed as:

[0033]

[0034] Among them, W d and b d are the decoder weights and biases.

[0035] Furthermore, the calculation loss function described in step 2.2 includes:

[0036] (1) Reconstruction loss: measures the difference between the input x and the reconstructed output The difference between them is usually expressed by mean square error MSE:

[0037]

[0038] (2) Quantization loss: quantization error, which measures the encoder output z e (x) and the quantized vector z q Differences in (x):

[0039]

[0040] Among them, sg(·) represents the operation of stopping the gradient;

[0041] (3) Encoder loss: Encourages the encoder to output z e (x) is close to the codebook vector:

[0042]

[0043] The final loss function is:

[0044] L=L rec +βL vq +γL enc ;

[0045] Among them, β and γ are the weight hyperparameters of the loss term.

[0046] Furthermore, the core steps of the graph attention mechanism described in step 3.3 include:

[0047] (1) Calculate the attention coefficient: For any two nodes i and j in the graph, their attention coefficient e ij According to node characteristics and calculate:

[0048]

[0049] Among them, a T is a learnable attention vector, || represents a vector concatenation operation, W (l) is the weight matrix of the graph convolutional layer;

[0050] (2) Normalized attention coefficient: Use the softmax function to normalize the attention coefficient of neighbor nodes:

[0051]

[0052] Among them, α ij It represents the weight of node i when aggregating the features of neighbor node j, and k represents any node among the neighbor nodes of node i. represents the set of neighbor nodes of node i, e ik Represents the attention coefficient of node i and node k;

[0053] (3) Attention-weighted convolution: The final graph convolution operation can be completed by weighted summation of neighbor node features:

[0054]

[0055] After the introduction of GAM, each node will dynamically adjust the weights of different neighbors according to the attention mechanism when aggregating the features of its neighbor nodes, which enables the model to flexibly capture complex spatiotemporal dependencies.

[0056] Furthermore, the specific operation steps of step 1 include:

[0057] Step 1.1: Collect data from sensors on the pipeline, including vibration acceleration vib, pressure pa, flow rate flow and temperature temp;

[0058] Step 1.2: Use Butterworth filter to remove high-frequency noise from the data. The formula is:

[0059]

[0060] Where: f is the signal frequency, fc is the cutoff frequency, n is the order of the filter, by adjusting f c and n, can control the strength and effect of filtering;

[0061] Step 1.3: Use Min-Max normalization to remove the impact of data dimension. The formula is:

[0062]

[0063] Among them, x min and x max are the minimum and maximum values ​​of the original data, respectively, norm is the normalized data.

[0064] Beneficial Effects

[0065] Compared with the prior art, the pipeline anomaly detection method based on VQSAE and DCGCN proposed in the present invention has the following beneficial effects:

[0066] (1) The present invention can effectively reduce the dimension of sensor signals through VQSAE, reduce data noise and extract key features, while maintaining the sparsity of features; at the same time, VQSAE generates discrete feature representations, which helps to capture discrete patterns in data (such as abnormal events) and has good robustness; DCGCN, based on traditional GCN, strengthens the feature transfer between different layers through a dense connection mechanism, can capture multi-level information in complex network structures, take into account local and global features, and reduce the loss in information transfer, while improving computational efficiency, and is suitable for processing large-scale pipeline networks. VQSAE processes different types of sensor data, compresses and extracts high-dimensional features, effectively removes noise and extracts key features, while DCGCN further utilizes the topological information of the pipeline network to perform structured processing on these features, combining local and global topological information, which can improve the accuracy and robustness of anomaly detection.

[0067] (2) The present invention combines the advantages of the two models to achieve more accurate, efficient, highly robust and scalable anomaly detection, which can handle complex pipeline topologies and multimodal data while reducing false alarm and missed alarm rates, helping managers identify potential risks earlier, optimize maintenance decisions, and reduce economic losses and safety hazards caused by failures.

[0068] (3) The present invention first uses a Butterworth filter to remove high-frequency noise from the data, and then uses MIN-MAX to normalize the data to ensure the quality of the data. The VQSAE model introduces a new activation function SELU to make the model more stable during training, especially in deep networks. Due to the normalization characteristics of SELU, the output value will not disappear, ensuring the normal propagation of the gradient, thereby improving the training efficiency and the generalization ability of the model. At the same time, GAM (graph attention mechanism) is introduced in DCGCN, which further improves the flexibility and accuracy of feature aggregation by dynamically adjusting the importance of neighbor node features, especially in scenarios with heterogeneous sensor networks and complex spatiotemporal dynamic data. VQSAE+DCGCN jointly processes and extracts data features in an innovative way to achieve better results. VQSAE provides compression and discrete representation, DCGCN processes graph structure and local information, and through training, VQSAE and DCGCN simultaneously jointly optimize features. VQSAE provides high-quality node representations, while DCGCN learns global and local structural relationships based on these representations to form richer feature expressions. In the entire architecture, VQSAE and DCGCN do not work independently, but interact with each other through a joint optimization strategy. The features learned by VQSAE are further enhanced through graph convolution, and the graph structure information learned by DCGCN can react to the optimization of the autoencoder part. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is the overall flow chart of the present invention.

[0070] Figure 2 Schematic diagram of the VQSAE model structure in the present invention.

[0071] Figure 3 This is a schematic diagram of the DCGCN model structure in the present invention. DETAILED DESCRIPTION

[0072] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Under the premise of not departing from the design concept of the present invention, various modifications and improvements made by ordinary persons in the art to the technical solutions of the present invention should all fall within the protection scope of the present invention.

[0073] Embodiment 1:

[0074] like Figure 1 As shown, a pipeline anomaly detection method based on VQSAE and DCGCN specifically includes the following steps:

[0075] Step 1: Collect the time series data of the sensors on the pipeline, build a data set and preprocess the data set; the specific operation method is:

[0076] Step 1.1: Collect sensor data, including vibration acceleration (vib), pressure (pa), flow (flow) and temperature (temp).

[0077] Step 1.2: Use Butterworth filter to remove high-frequency noise from the data. The formula is:

[0078]

[0079] Where: f is the signal frequency, f c is the cutoff frequency, n is the order of the filter, by adjusting f c and n, can control the strength and effect of the filter.

[0080] Step 1.3: Use Min-Max normalization to remove the impact of data dimension. The formula is:

[0081]

[0082] Among them, x min and x max are the minimum and maximum values ​​of the original data, respectively, norm is the normalized data.

[0083] Step 2: Build an improved VQSAE model and introduce a new activation function SELU to extract local time features of the data; the specific operation method includes the following steps:

[0084] Step 2.1: VQSAE is a random autoencoder based on vector quantization. It encodes the input data into a discrete feature space for representation, and obtains an improved VQSAE model. The improved VQSAE model includes the following structure:

[0085] (1) Encoder: Assume the input is a vector x∈R d , where d is the feature dimension, the encoder maps the input data to the latent space, usually through a convolutional layer or a fully connected layer, and introduces a new activation function SELU for nonlinear transformation; it is a self-regularized activation function. Compared with common activation functions such as RELU, SELU has a stable activation distribution in deep neural networks, which helps to avoid the problem of gradient disappearance or explosion. The formula of the SELU function is:

[0086]

[0087] Where: λ=1.0507, α=1.67326, x is the output when it is greater than 0, αe x -α is the output when it is less than 0;

[0088] Then: The output z of the encoder is e (x) is expressed as:

[0089] z e (x) = SELU(W e x+b e )

[0090] Among them, W e and b e are the weights and biases of the encoder, and SELU is used as the activation function to ensure the self-regularization characteristics of the encoder output.

[0091] (2) Vector Quantization: The encoder output z e (x) is mapped to a discrete codebook index through a vector quantizer. First, find the codebook index corresponding to z. e (x) The nearest vector z q (x):

[0092]

[0093] Among them, C is the codebook set, e i is each vector in the codebook.

[0094] (3) Decoder: quantizes the vector z q (x) is input to the decoder and reconstructed into an estimate of the input data Also using the SELU activation function, the decoder output It can be expressed as:

[0095]

[0096] Among them, W d and b d are the decoder weights and biases.

[0097] Step 2.2: Calculate the loss function and extract discrete features z q (x); including:

[0098] (1) Reconstruction loss: measures the difference between the input x and the reconstructed output The difference between them is usually expressed by mean square error MSE:

[0099]

[0100] (2) Quantization loss: quantization error, which measures the encoder output z e (x) and the quantized vector z q Differences in (x):

[0101]

[0102] Among them, sg(·) represents the operation of stopping the gradient;

[0103] (3) Encoder loss: Encourages the encoder to output z e (x) is close to the codebook vector:

[0104]

[0105] The final loss function is:

[0106] L=L rec +βL vq +γL enc ;

[0107] Among them, β and γ are the weight hyperparameters of the loss term.

[0108] Step 3: Based on the features extracted by the VQSAE model, a DCGCN global spatiotemporal relationship model is constructed, and the graph attention mechanism GAM is introduced to perform spatiotemporal graph convolution operations; specifically:

[0109] Step 3.1: The discrete features z extracted by VQSAE q (x) is used as the input of the spatiotemporal graph convolutional network; assuming the input data is:

[0110] X∈R N×T×F ,

[0111] Among them, N represents the number of nodes in the graph, T represents the time step, and F represents the feature dimension; after VQSAE processing, the encoding feature Zq∈R N×T×F , used for graph convolution.

[0112] Step 3.2: In the spatiotemporal graph convolution network DCGCN, the key to graph convolution is to perform graph convolution operations on the features of each time step t, using the adjacency matrix A of the graph to capture spatial dependencies; the basic graph convolution operation can be expressed as:

[0113]

[0114] Among them, H (l+1) is the feature matrix of the l+1th layer; is the feature matrix of the lth layer; F l represents the feature dimension of the lth layer, is the normalized adjacency matrix, D is the degree matrix; W(l) is the weight matrix of the graph convolutional layer, and σ(·) is the SELU nonlinear activation function.

[0115] In order to model time dependencies, dilated convolution is introduced to expand in the time dimension so that the convolution operation can capture the dependencies of different time steps; the spatiotemporal graph convolution is expressed as:

[0116]

[0117] in, is the weight of the spatial convolution, is the weight of the temporal convolution, capturing the dependencies of different time steps Δt through dilated convolution.

[0118] Step 3.3: Introduce the graph attention mechanism GAM; the goal of the attention mechanism is to dynamically calculate the attention weights between each node and its neighbors to adaptively adjust the calculation process of graph convolution; the core steps of the graph attention mechanism include:

[0119] (1) Calculate the attention coefficient: For any two nodes i and j in the graph, their attention coefficient e ij According to node characteristics and calculate:

[0120]

[0121] Among them, a T is a learnable attention vector, || represents a vector concatenation operation, W (l) is the weight matrix of the graph convolutional layer.

[0122] (2) Normalized attention coefficient: Use the softmax function to normalize the attention coefficient of neighbor nodes:

[0123]

[0124] Among them, α ij It represents the weight of node i when aggregating the features of neighbor node j, and k represents any node among the neighbor nodes of node i. represents the set of neighbor nodes of node i, e ik Represents the attention coefficient of node i and node k.

[0125] (3) Attention-weighted convolution: The final graph convolution operation can be completed by weighted summation of neighbor node features:

[0126]

[0127] After the introduction of GAM, each node will dynamically adjust the weights of different neighbors according to the attention mechanism when aggregating the features of its neighbor nodes, which enables the model to flexibly capture complex spatiotemporal dependencies.

[0128] Step 4: Detect pipeline anomalies through the DCGCN model and output pipeline anomaly detection results.

[0129] In order to verify the feasibility and superiority of this solution, the inventor conducted a comparative experiment for comparison. The specific experimental steps are as follows:

[0130] Design and conduct comparative experiments, divide the dataset into training set, test set and validation set according to the ratio of 1:1:3 to test the model, and select CNN-MLC, GCN, and IMFs models for comparative experiments. At the same time, use the three evaluation indicators of Precision, Recall and F1-score to measure the actual performance of the model.

[0131] Among them, True Positive (TP) means that the positive sample is successfully labeled as positive. True Negative (TN) means that the negative sample is successfully labeled as negative. False Positive (FP) means that the negative sample is mistakenly labeled as a positive sample. False Negative (FN) means that the positive sample is mistakenly labeled as a negative sample. The specific formulas of the three models of Precision, Recall and F1-score are:

[0132] Precision = TP / (TP+FP);

[0133] Recall = TP / (TP+FN);

[0134] F1-score=(2*Precision*Recall) / (Precision+Recall).

[0135] After the experimental operation, the following experimental data were obtained:

[0136] Model Precision Recall F1-score CNN-MLC 0.821 0.782 0.801 GCN 0.763 0.747 0.755 IMFs 0.796 0.772 0.784 VQSAE+DCGCN 0.954 0.933 0.927

[0137] According to the experimental data in the above table, we can draw the experimental conclusion that the VQSAE+DCGCN model designed in this embodiment has a precision value of 0.954, a recall value of 0.933, and an F1-score value of 0.927. Compared with other models, the performance of various parameters is better. Therefore, this solution has a better effect in pipeline anomaly detection.

[0138] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the present invention.

Claims

1. A pipeline anomaly detection method based on VQSAE and DCGCN, characterized by: The specific steps include: Step 1: Collect time series data from sensors on the pipeline, build a data set, and preprocess the data set; Step 2: Build an improved VQSAE model and introduce a new activation function SELU to extract local time features of the data; the specific operation method includes the following steps: Step 2.1: VQSAE is a random autoencoder based on vector quantization, which encodes the input data into a discrete feature space for representation, and obtains an improved VQSAE model. The improved VQSAE model includes the following structures: encoder, vector quantizer and decoder; Step 2.2: Calculate the loss function and extract discrete features z q (x); Step 3: Based on the features extracted by the VQSAE model, a DCGCN global spatiotemporal relationship model is constructed, and the graph attention mechanism GAM is introduced to perform spatiotemporal graph convolution operations; Step 3.1: The discrete features z extracted by VQSAE q (x) is used as the input of the spatiotemporal graph convolutional network; assuming the input data is: X∈R N×T×F , Among them, N represents the number of nodes in the graph, T represents the time step, and F represents the feature dimension; after VQSAE processing, the encoding feature Zq∈R N×T×F , used for graph convolution; Step 3.2: In the spatiotemporal graph convolution network DCGCN, the key to graph convolution is to perform graph convolution operations on the features of each time step t, using the adjacency matrix A of the graph to capture spatial dependencies; the basic graph convolution operation can be expressed as: Among them, H (l+1) is the feature matrix of the l+1th layer; is the feature matrix of the lth layer; F l represents the feature dimension of the lth layer, is the normalized adjacency matrix, D is the degree matrix; W (l) is the weight matrix of the graph convolutional layer, σ(·) is the SELU nonlinear activation function; In order to model the temporal dependency, the dilated convolution is introduced to expand in the time dimension so that the convolution operation can capture the dependencies of different time steps; the spatiotemporal graph convolution is expressed as: in, is the weight of the spatial convolution, is the weight of the temporal convolution, which captures the dependencies of different time steps Δt through dilated convolution; Step 3.3: Introduce the graph attention mechanism GAM; the goal of the attention mechanism is to dynamically calculate the attention weights between each node and its neighbors to adaptively adjust the calculation process of the graph convolution; Step 4: Detect pipeline anomalies through the DCGCN model and output pipeline anomaly detection results.

2. The pipeline anomaly detection method based on VQSAE and DCGCN according to claim 1, characterized in that: Encoder as described in step 2.1: Assume the input is a vector x∈R d , where d is the feature dimension. The encoder maps the input data to the latent space, usually through a convolutional layer or a fully connected layer, and introduces a new activation function SELU for nonlinear transformation; the formula of the SELU function is: Where: λ=1.0507, α=1.67326, x is the output when it is greater than 0, αe x -α is the output when it is less than 0; Then: The output z of the encoder is e (x) is expressed as: z e (x)=SELU(W e x+b e ) Among them, W e and b e are the weights and biases of the encoder, and SELU is used as the activation function to ensure the self-regularization characteristics of the encoder output.

3. The pipeline anomaly detection method based on VQSAE and DCGCN according to claim 2 is characterized in that: The vector quantizer described in step 2.1: The output z of the encoder e (x) is mapped to a discrete codebook index through a vector quantizer. First, find the codebook index corresponding to z. e (x) The nearest vector z q (x): Among them, C is the codebook set, e i is each vector in the codebook.

4. The pipeline anomaly detection method based on VQSAE and DCGCN according to claim 3 is characterized in that: The decoder described in step 2.1: converts the quantized vector z q (x) is input to the decoder and reconstructed into an estimate of the input data Also using the SELU activation function, the decoder output It can be expressed as: Among them, W d and b d are the decoder weights and biases.

5. The pipeline anomaly detection method based on VQSAE and DCGCN according to claim 4 is characterized in that: Calculate the loss function as described in step 2.2, including: (1) Reconstruction loss: measures the difference between the input x and the reconstructed output The difference between them is usually expressed by mean square error MSE: (2) Quantization loss: quantization error, which measures the encoder output z e (x) and the quantized vector z q Differences in (x): Among them, sg(·) represents the operation of stopping the gradient; (3) Encoder loss: Encourages the encoder to output z e (x) is close to the codebook vector: The final loss function is: L=L rec +βL vq +γL enc ; Among them, β and γ are the weight hyperparameters of the loss term.

6. The pipeline anomaly detection method based on VQSAE and DCGCN according to claim 1 is characterized in that: The core steps of the graph attention mechanism described in step 3.3 include: (1) Calculate the attention coefficient: For any two nodes i and j in the graph, their attention coefficient e ij According to node characteristics and calculate: Among them, a T is a learnable attention vector, ∥ represents a vector concatenation operation, W (l) is the weight matrix of the graph convolutional layer; (2) Normalized attention coefficient: Use the softmax function to normalize the attention coefficient of neighbor nodes: Among them, α ij It represents the weight of node i when aggregating the features of neighbor node j, and k represents any node among the neighbor nodes of node i. represents the set of neighbor nodes of node i, e ik Represents the attention coefficient of node i and node k; (3) Attention-weighted convolution: The final graph convolution operation can be completed by weighted summation of neighbor node features: After the introduction of GAM, each node will dynamically adjust the weights of different neighbors according to the attention mechanism when aggregating the features of its neighbor nodes, which enables the model to flexibly capture complex spatiotemporal dependencies.

7. The pipeline anomaly detection method based on VQSAE and DCGCN according to claim 1 is characterized in that: The specific operation steps of step 1 include: Step 1.1: Collect data from sensors on the pipeline, including vibration acceleration vib, pressure pa, flow rate flow and temperature temp; Step 1.2: Use Butterworth filter to remove high-frequency noise from the data. The formula is: Where: f is the signal frequency, f c is the cutoff frequency, n is the order of the filter, by adjusting f c and n, can control the strength and effect of filtering; Step 1.3: Use Min-Max normalization to remove the impact of data dimension. The formula is: Among them, x min and x max are the minimum and maximum values ​​of the original data, respectively, norm is the normalized data.

Images