A method and system for risk perception and early warning of the operation status of oil and gas pipeline networks

Abstract

The present invention discloses a method and system for risk perception and early warning of the operation state of oil and gas pipeline networks. The method includes: fusing multi-dimensional features of the pipeline network operation state according to the real-time operation data of the oil and gas pipeline network to generate a spatio-temporal feature matrix; inputting the spatio-temporal feature matrix into a deep Gaussian process model to probabilistically model the abnormal features in the pipeline network operation state and output a risk feature vector with confidence; constructing a three-dimensional risk concentration distribution field based on the risk feature vector to predict the propagation path and diffusion trend of the risk concentration distribution field; for the prediction result, constructing an adaptive early warning threshold generation model based on a generative adversarial network to generate multi-level early warning thresholds that change with the operation state of the pipeline network, and realizing risk perception and early warning of the operation state of the oil and gas pipeline network according to the comparison result between the real-time risk features and the early warning thresholds. Using the embodiments of the present invention, dynamic risk assessment and adaptive early warning can be realized, enhancing the safety and emergency response ability of the pipeline network.

Description

Technical Field

[0001] The present invention belongs to the technical field of oil and gas pipeline networks, and in particular to a method and system for perceiving and warning the risk of operating status of an oil and gas pipeline network. Background Art

[0002] As global energy demand continues to grow, oil and gas pipeline networks, as important energy transportation channels, bear huge transportation pressure and flow. Therefore, ensuring the safe and efficient operation of oil and gas pipeline networks has become a major issue that needs to be urgently addressed in the energy industry. Traditional pipeline network monitoring methods mainly rely on regular inspections and manual monitoring, which makes it difficult to achieve real-time perception and accurate early warning of complex operating conditions. This method has many defects such as slow response and high misjudgment rate when responding to sudden dangers, which may lead to significant economic losses and environmental damage. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for risk perception and early warning of the operating status of an oil and gas pipeline network, so as to address the deficiencies in the prior art, realize dynamic risk assessment and adaptive early warning, and enhance the safety and emergency response capabilities of the pipeline network.

[0004] An embodiment of the present application provides a method for early warning of oil and gas pipeline network operation status risk perception, the method comprising:

[0005] Based on the real-time operation data of the oil and gas pipeline network, the multi-dimensional characteristics of the pipeline network operation status are integrated to generate a spatiotemporal feature matrix containing pressure, temperature, flow and vibration information;

[0006] The spatiotemporal feature matrix is ​​input into the deep Gaussian process model to perform probabilistic modeling on the abnormal features in the operation status of the pipeline network and output a risk feature vector with confidence.

[0007] According to the risk characteristic vector, a three-dimensional risk concentration distribution field is constructed, and the propagation path and diffusion trend of the risk concentration distribution field are predicted based on the graph neural network;

[0008] According to the prediction results of the risk concentration distribution field, an adaptive warning threshold generation model based on the generative adversarial network is constructed. Through the dynamic game between the normal operating condition generator and the real-time discriminator, a multi-level warning threshold that changes with the operating status of the pipeline network is generated. Based on the comparison results of the real-time risk characteristics and the warning threshold, the risk-aware warning of the operating status of the oil and gas pipeline network is realized.

[0009] Optionally, the multi-dimensional features of the pipeline network operation status are integrated according to the real-time operation data of the oil and gas pipeline network to generate a spatiotemporal feature matrix containing pressure, temperature, flow and vibration information, including:

[0010] Based on the real-time operation data collected by multi-source sensors of oil and gas pipeline networks, timestamp alignment and missing value filling are performed. A multi-source data alignment algorithm based on dynamic time warping is used to eliminate the differences in sensor sampling frequencies, and a time-synchronized multi-source data sequence is obtained. Among them, the real-time operation data includes pressure, temperature, flow rate, and vibration data;

[0011] The multi-source data sequence is input into a spatio-temporal convolutional neural network. Among them, in the spatial dimension, a graph convolutional network is used to capture the topological structure features of the pipeline network, and in the time dimension, a one-dimensional convolutional kernel is used to extract time series features, obtaining a preliminary spatio-temporal feature representation;

[0012] For the extracted spatio-temporal feature representation, the multi-head self-attention mechanism is used to calculate the correlation weights between different sensor data, and the features are weighted and fused according to the weights to highlight the contribution of key sensor data to risk perception, obtaining a weighted spatio-temporal feature representation;

[0013] The weighted spatio-temporal features are input into a sparse autoencoder, and through non-linear mapping for dimensionality reduction, redundant information is removed, key features are retained, and a low-dimensional spatio-temporal feature matrix containing pressure, temperature, flow rate, and vibration information is generated.

[0014] Optionally, the spatio-temporal feature matrix is input into a deep Gaussian process model to probabilistically model the abnormal features in the pipeline network operation state, and a risk feature vector with confidence is output, including:

[0015] The spatio-temporal feature matrix is input into a multi-layer perceptron, and through multi-layer non-linear transformations, the original features are mapped into a high-dimensional latent space to obtain a deep feature representation. Among them, each layer uses an activation function with a residual connection to avoid the problem of gradient disappearance and enhance the feature expression ability;

[0016] For the deep feature representation, a probability model based on Gaussian process regression is constructed, and a linear combination of a radial basis kernel function and a periodic kernel function is used to capture the non-linear relationships and periodic change laws in the pipeline network operation state, obtaining a preliminary probabilistically characterized representation;

[0017] According to the probabilistically characterized representation, the predicted mean and variance of each feature point are calculated, and through the variational inference algorithm, the uncertainty of the model is quantified to obtain a probability feature distribution with a confidence interval;

[0018] For the probability feature distribution, an anomaly detection algorithm based on Mahalanobis distance is used to calculate the distance between each feature point and the normal operating condition distribution, and according to a preset confidence threshold, the abnormal feature points are screened out to generate a risk feature vector with confidence.

[0019] Optionally, constructing a three-dimensional risk concentration distribution field based on the risk feature vector and predicting the propagation path and diffusion trend of the risk concentration distribution field based on a graph neural network, including:

[0020] According to the risk feature vector, combining with the topological structure information of the oil and gas pipeline network, map each risk feature point to the three-dimensional space coordinates of the pipeline network, and use the Kriging interpolation algorithm to perform spatial interpolation on the discrete risk feature points to generate a preliminary three-dimensional risk concentration distribution field;

[0021] According to the preliminary three-dimensional risk concentration distribution field, combining with the time dimension information, use the spatio-temporal Kriging interpolation algorithm, combined with historical risk propagation data, to dynamically correct the risk concentration distribution field to obtain a three-dimensional risk concentration distribution field that changes with time;

[0022] Input the three-dimensional risk concentration distribution field that changes with time into the graph neural network, construct a graph model according to the pipeline network topology structure, where the nodes of the graph model represent the key positions of the pipeline network, and the edges of the graph model represent the pipe segment connection relationships, and use the graph attention mechanism to capture the risk propagation dependence relationship between nodes to obtain an initial prediction result of risk propagation;

[0023] According to the initial prediction result of risk propagation, use a sequence prediction model based on a time convolutional network, combined with historical risk propagation path data, to predict the risk propagation path and diffusion trend in future time steps, and generate a propagation prediction result of the dynamic risk concentration distribution field.

[0024] Optionally, for the prediction result of the risk concentration distribution field, construct an adaptive early warning threshold generation model based on a generative adversarial network, generate multi-level early warning thresholds that change with the operating state of the pipeline network through the dynamic game between the normal condition generator and the real-time discriminator, and realize the risk perception early warning of the operating state of the oil and gas pipeline network according to the comparison result between the real-time risk features and the early warning thresholds, including:

[0025] According to the historical normal operation data, train a normal condition generator based on a generative adversarial network to generate simulation data that conforms to the normal operation characteristics of the pipeline network; among them, the generator uses a variational autoencoder with conditional constraints to ensure the diversity and authenticity of the generated data;

[0026] Construct a real-time discriminator based on a deep residual network, input the prediction result of the risk concentration distribution field and the normal condition data generated by the generator, learn the boundary features between normal and abnormal states through dynamic game, and output the discriminant result and its confidence;

[0027] According to the discrimination results, an adaptive clustering algorithm is used to perform multi-level partitioning on the discrimination results. Combining with the dynamic changes of the risk concentration distribution field, multi-level warning thresholds that change with the operating state of the pipeline network are generated. The multi-level warning thresholds include low-risk, medium-risk, and high-risk thresholds;

[0028] The real-time risk characteristics are compared with the multi-level warning thresholds, and a decision-making model based on fuzzy logic is adopted. According to the confidence level of the risk characteristics and the degree of deviation from the threshold, the warning level is dynamically adjusted to achieve risk perception warning of the operating state of the oil and gas pipeline network.

[0029] Another embodiment of the present application provides a risk perception warning system for the operating state of an oil and gas pipeline network. The system includes:

[0030] A fusion module for fusing multi-dimensional characteristics of the pipeline network operating state according to the real-time operating data of the oil and gas pipeline network, and generating a spatio-temporal feature matrix containing pressure, temperature, flow rate, and vibration information;

[0031] A modeling module for inputting the spatio-temporal feature matrix into a deep Gaussian process model, probabilistically modeling the abnormal features in the pipeline network operating state, and outputting a risk feature vector with confidence;

[0032] A construction module for constructing a three-dimensional risk concentration distribution field according to the risk feature vector, and predicting the propagation path and diffusion trend of the risk concentration distribution field based on a graph neural network;

[0033] A warning module for constructing an adaptive warning threshold generation model based on a generative adversarial network for the prediction results of the risk concentration distribution field. Through the dynamic game between the normal working condition generator and the real-time discriminator, multi-level warning thresholds that change with the operating state of the pipeline network are generated, and according to the comparison result between the real-time risk characteristics and the warning thresholds, risk perception warning of the operating state of the oil and gas pipeline network is realized.

[0034] Another embodiment of the present application provides a storage medium in which a computer program is stored. Wherein, the computer program is set to execute the method described in any one of the above when running.

[0035] Another embodiment of the present application provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is set to run the computer program to execute the method described in any one of the above.

[0036] Compared with the prior art, a risk perception and early warning method for the operation state of an oil and gas pipeline network provided by the present invention fuses multi-dimensional features of the pipeline network operation state according to the real-time operation data of the oil and gas pipeline network to generate a spatio-temporal feature matrix; inputs the spatio-temporal feature matrix into a deep Gaussian process model to probabilistically model the abnormal features in the pipeline network operation state and outputs a risk feature vector with confidence; constructs a three-dimensional risk concentration distribution field according to the risk feature vector to predict the propagation path and diffusion trend of the risk concentration distribution field; for the prediction result, constructs an adaptive early warning threshold generation model based on a generative adversarial network to generate multi-level early warning thresholds that change with the pipeline network operation state, and realizes risk perception and early warning of the oil and gas pipeline network operation state according to the comparison result between the real-time risk features and the early warning thresholds, so as to be able to realize dynamic risk assessment and adaptive early warning, enhance the safety of the pipeline network and the emergency response ability. Description of the Drawings

[0037] Figure 1 It is a hardware structure block diagram of a computer terminal for a risk perception and early warning method for the operation state of an oil and gas pipeline network provided by an embodiment of the present invention;

[0038] Figure 2 It is a schematic flow chart of a risk perception and early warning method for the operation state of an oil and gas pipeline network provided by an embodiment of the present invention;

[0039] Figure 3 It is a schematic structural diagram of a risk perception and early warning system for the operation state of an oil and gas pipeline network provided by an embodiment of the present invention. Detailed Embodiments

[0040] The embodiments described below with reference to the drawings are exemplary and are only used to explain the present invention and should not be construed as limiting the present invention.

[0041] An embodiment of the present invention first provides a risk perception and early warning method for the operation state of an oil and gas pipeline network. This method can be applied to electronic devices, such as computer terminals, specifically, ordinary computers, etc.

[0042] The following takes running on a computer terminal as an example to describe it in detail. Figure 1 It is a hardware structure block diagram of a computer terminal for a risk perception and early warning method for the operation state of an oil and gas pipeline network provided by an embodiment of the present invention. As Figure 1 shown, this computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the memory can include a non-volatile storage medium and an internal memory.

[0043] The non-volatile storage medium can store an operating system and a computer program. This computer program includes program instructions. When the program instructions are executed, the processor can execute any risk perception and early warning method for the operation state of an oil and gas pipeline network.

[0044] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0045] The internal memory provides an environment for the operation of a computer program in a non-volatile storage medium. When the computer program is executed by the processor, the processor can be made to execute any one of the risk perception and early warning methods for the operation status of the oil and gas pipeline network.

[0046] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 The structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0047] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0048] See Figure 2 , Embodiments of the present invention provide a risk perception and early warning method for the operation status of an oil and gas pipeline network, which may include the following steps:

[0049] S201, according to the real-time operation data of the oil and gas pipeline network, fuse the multi-dimensional features of the pipeline network operation status to generate a spatio-temporal feature matrix including pressure, temperature, flow rate, and vibration information;

[0050] This step aims to generate a spatio-temporal feature matrix by fusing multi-dimensional features of real-time operation data from oil and gas pipe networks, so as to reflect the comprehensive operation status of the pipe networks. First, the system collects key data related to the operation of oil and gas pipe networks from multi-source sensors, including pressure, temperature, flow rate, and vibration. These data usually exist in the form of time series, and the sampling frequencies of each sensor may be different, resulting in problems of time alignment of the data. To solve this problem, the system adopts the dynamic time warping (DTW) algorithm to ensure the synchronization of data from different sensors in the time dimension, thus obtaining an integrated multi-source data sequence.

[0051] Then, after the data alignment process, the system integrates these real-time operation data in two dimensions: space and time, to form a spatio-temporal feature matrix. In this process, the spatio-temporal convolutional neural network (ST-CNN) is applied to extract spatial topological information and time series features. Specifically, the graph convolutional network is used in the spatial dimension to capture the topological structure features of the pipe network, and the one-dimensional convolutional kernel is used in the time dimension to extract time series features. Finally, the fused spatio-temporal feature matrix contains various operation status information of the pipe network, laying a foundation for subsequent risk identification and prediction.

[0052] Through the multi-dimensional feature fusion of the real-time operation data of oil and gas pipe networks, the generated spatio-temporal feature matrix can effectively reflect the overall picture of the operation status of the pipe networks. This process not only improves the utilization efficiency of data, but also provides more accurate input data for subsequent risk identification. For example, by comprehensively considering various factors such as pressure, temperature, flow rate, and vibration, the system can more comprehensively understand the health status of the pipe network, and thus issue early warnings in a timely manner when abnormalities occur. The establishment of such comprehensive features helps to improve the operation safety of oil and gas pipe networks, provides a basis for risk management, and reduces potential safety hazards.

[0053] Specifically, according to the real-time operation data collected by the multi-source sensors of the oil and gas pipe network, timestamp alignment and missing value filling processing can be carried out, and a multi-source data alignment algorithm based on dynamic time warping is adopted to eliminate the differences in sensor sampling frequencies, so as to obtain a time-synchronized multi-source data sequence, where the real-time operation data includes pressure, temperature, flow rate, and vibration data;

[0054] In this step, the system first collects real-time operation data from multi-source sensors of the oil and gas pipe network, and the data types cover key parameters such as pressure, temperature, flow rate, and vibration. However, the data from different sensors often have problems of inconsistent timestamps, resulting in difficulties in subsequent analysis and modeling. To effectively solve this problem, the system adopts the dynamic time warping (DTW) algorithm to align data with different time frequencies, ensuring that all sensor data can be analyzed on the same time scale.

[0055] In the specific implementation process, the system will first preprocess the data of each sensor, including removing noise and filling missing values. The filling of missing values can be achieved through interpolation methods such as linear interpolation or spline interpolation to ensure the continuity and integrity of the data. Then, the DTW algorithm is used to compare the time series of each sensor, identify and adjust the shortest time axis to obtain the best alignment. Finally, the system will generate a time-synchronized multi-source data sequence, providing a standardized data basis for subsequent feature extraction.

[0056] By performing time alignment and missing value processing on multi-source sensor data, the system can obtain high-quality data sequences that can be used for analysis. This step not only improves the integrity of the data but also lays a solid foundation for subsequent feature extraction. By effectively eliminating the differences in sensor sampling frequencies, it can ensure the accuracy of the data used in subsequent analysis, thereby improving the risk identification and early warning capabilities of the operation status of oil and gas pipelines. This improvement in data quality helps to achieve real-time monitoring and fault prediction of the pipeline network, enhancing the safety of pipeline network operation.

[0057] In this step, the system first collects real-time operation data from various multi-source sensors installed at different positions in the oil and gas pipeline network to monitor key parameters such as pressure, temperature, flow rate, and vibration. Since the operating frequencies of different sensors may vary, resulting in inconsistent data timestamps, it is necessary to perform time alignment processing on the data. The system will use the Dynamic Time Warping (DTW) algorithm, which is a method capable of handling the similarity of time series. It calculates the optimal matching path between different time series for time alignment.

[0058] In the specific implementation process, first, the data collected by each sensor is preprocessed, including removing noise and filling missing values. The system will use the linear interpolation method to fill in the missing data, that is, fill in the missing part by interpolating the existing data points before and after to ensure the continuity of the data. Next, the DTW algorithm is used for time alignment. The system will construct a distance matrix to record the distances between all sensor data points and determine the time matching path through this distance matrix to eliminate the influence of different sampling frequencies.

[0059] Finally, the aligned multi-source data sequences will be integrated to form a time-synchronized multi-source data sequence. In this process, the system will ensure that the data of all sensors are balanced at the same timestamp, enabling subsequent processing and analysis to be carried out on a consistent data basis. The completion of this step provides a solid data foundation for subsequent feature extraction, ensuring the integrity and availability of the data.

[0060] Input the multi-source data sequence into a spatio-temporal convolutional neural network, where the graph convolutional network is used in the spatial dimension to capture the topological structure features of the pipe network, and a one-dimensional convolutional kernel is used in the temporal dimension to extract the time series features, obtaining a preliminary spatio-temporal feature representation;

[0061] In this step, the multi-source data sequence after time alignment and missing value processing will be input into a spatio-temporal convolutional neural network (ST-CNN) to extract richer feature information. This network structure is specifically designed to process data in both spatial and temporal dimensions and is particularly suitable for describing systems with complex topological structures such as pipe networks. The spatial dimension of the system uses a graph convolutional network (GCN) to capture the topological features of the pipe network, such as the connection relationships of pipes, the relative positions of each node, and their attributes. In addition, in the temporal dimension, the system uses a one-dimensional convolutional kernel to extract time series features and captures time-varying information through a sliding window method.

[0062] Specifically, in implementation, the system divides the multi-source data sequence into multiple small windows, enabling parallel processing and extraction of relevant features for each time period. Through the GCN, the system can identify the features of each node (such as pumping stations, valves, etc.) and their connected pipes in the pipe network, generating a spatial feature matrix. After obtaining the spatial features, the one-dimensional convolutional kernel continues to perform a convolutional operation on the time series to capture the key change trends and periodic signals in the time series. Finally, the spatio-temporal convolutional neural network will generate a preliminary spatio-temporal feature representation, comprehensively reflecting the operating state of the oil and gas pipeline network.

[0063] Processing the multi-source data using a spatio-temporal convolutional neural network can effectively combine the characteristics of space and time and extract deeper feature information. The fusion of spatial and temporal features provides strong data support for subsequent anomaly detection and risk assessment. By identifying the topological structure and dynamic changes of the pipe network, the system can better understand the operating state of the pipe network and then achieve the prediction of potential risks. This comprehensive feature representation greatly improves the accuracy and efficiency of data analysis and lays a solid foundation for the intelligent management of oil and gas pipeline networks.

[0064] In this step, the system will input the multi-source data sequence that has undergone time alignment and preprocessing into a spatio-temporal convolutional neural network (ST-CNN). The design of this network aims to capture both the spatial and temporal features of the data and is particularly suitable for processing oil and gas pipeline networks with complex topological structures. The system first structures the input data to ensure that each data point can correspond to the nodes of the pipe network, facilitating subsequent extraction of spatial features.

[0065] To extract the spatial features of the pipeline network, the system uses a graph convolutional network (GCN) to process the data. By processing the topological structure information of the pipeline network, such as the connection relationship of the pipelines and the location of the sensors, GCN can effectively identify the relationship between the nodes in the pipeline network and generate a spatial feature representation. At the same time, GCN can use the characteristics of the graph structure to mine the connection patterns and local characteristics that are particularly important for risk perception. The effect of this process is directly related to the accuracy of the subsequent risk assessment, so the focus is on ensuring that the information of each node is fully transmitted and processed.

[0066] The feature extraction of the time dimension is performed using a one-dimensional convolution kernel. The system performs sliding window processing on each time series and adjusts the size of the convolution kernel to capture the change information in different time periods. For example, in order to identify abnormal fluctuations in flow in a specific time period, the system uses a small convolution kernel to extract local features, while for the analysis of longer time trends, a larger convolution kernel is used. Ultimately, through feature extraction in both space and time dimensions, the system can generate a preliminary spatiotemporal feature representation that comprehensively reflects the operating status of the pipeline network.

[0067] For the extracted spatiotemporal feature representation, the multi-head self-attention mechanism is used to calculate the correlation weights between different sensor data, and the features are weighted fused according to the weights to highlight the contribution of key sensor data to risk perception, thus obtaining the weighted spatiotemporal feature representation;

[0068] In this step, the system applies a multi-head self-attention mechanism to the previously extracted spatiotemporal feature representations to calculate the correlation weights between different sensor data. The self-attention mechanism allows the system to dynamically focus on the most important information in a specific input data, which is particularly important for risk perception in oil and gas pipeline networks. Specifically, the system calculates the correlation between each group of features based on the extracted spatiotemporal feature representations to determine which sensor data has a greater contribution in risk identification.

[0069] The specific implementation process includes that the system first inputs the spatiotemporal feature representation into the multi-head self-attention module, where each head independently calculates the weighted average of the feature vector, reflecting the importance of specific features to the overall risk assessment. In this way, the system can effectively identify key sensor data. For example, under a certain working condition, flow data and pressure changes may have a higher correlation. Ultimately, the weighted feature representation will highlight the contribution of key sensors to risk perception, which will help subsequent anomaly detection and risk assessment.

[0070] This process significantly enhances the system's sensitivity to key features. By focusing on the key sensor data that affects risk perception, the system can more accurately reflect the operating state of the pipeline network, effectively improving the accuracy of risk prediction. For example, when the pressure is abnormal, the changes in flow rate or temperature may be more significant. Using the self-attention mechanism can ensure that these key data are properly emphasized in risk assessment. The final weighted feature representation provides a solid information basis for subsequent risk feature extraction and prediction, improving the effectiveness of pipeline network risk management.

[0071] In this step, the system processes the spatio-temporal feature representation extracted in the previous step, and uses the multi-head self-attention mechanism to calculate the correlation weights between different sensor data. Specifically, the system will identify the features that have a greater impact on risk perception in a specific state through the self-attention mechanism, and determine which data have a more significant impact on the overall risk assessment by comparing the correlations between different sensor data.

[0072] In the specific implementation, the system first performs a linear transformation on the spatio-temporal feature representation to generate query, key, and value vectors for multiple heads. Each head independently calculates the attention weights between its features, and then applies these weights to the corresponding value vectors to achieve weighted averaging. Through the multi-head mechanism, the system can extract information from different subspaces, thereby comprehensively considering the relationships between multiple features and strengthening the influence of important features. For example, when a significant change in flow rate is detected, the system can automatically adjust the weights so that the data from the flow rate sensor occupies a larger proportion in subsequent analysis.

[0073] Finally, the system concatenates the output results of all heads to form a weighted spatio-temporal feature representation. This weighting process ensures that the sensor data crucial for risk perception is highlighted, reducing the interference of redundant information and providing more accurate input data for subsequent risk assessment. In this way, the system can effectively improve its sensitivity to the operating state of the oil and gas pipeline network, further enhancing the effectiveness of the entire risk management process.

[0074] Input the weighted spatio-temporal features into a sparse autoencoder, perform dimensionality reduction through non-linear mapping, remove redundant information, retain key features, and generate a low-dimensional spatio-temporal feature matrix containing pressure, temperature, flow rate, and vibration information.

[0075] In this step, the system inputs the weighted spatio-temporal features into a sparse autoencoder to achieve dimensionality reduction of the features. The sparse autoencoder is an effective non-linear dimensionality reduction tool that can remove redundant data while retaining key information. In the analysis of the operating status of oil and gas pipeline networks, due to the complexity of multi-dimensional features, directly processing high-dimensional data may lead to problems such as low computational efficiency and overfitting. Therefore, through the non-linear mapping of the sparse autoencoder, the system can obtain a low-dimensional feature representation, which is beneficial for subsequent data analysis and risk assessment.

[0076] In the specific implementation process, the system will first define the network structure of the autoencoder, including the input layer, hidden layer, and output layer. The input layer receives the weighted spatio-temporal features, and after being transformed by the non-linear activation function (such as ReLU) in the hidden layer, the system will reconstruct the input data in the output layer. By introducing a sparsity constraint, the system can force some nodes in the hidden layer to be zero, thus achieving the effects of feature selection and dimensionality reduction. Finally, the sparse autoencoder will generate a low-dimensional spatio-temporal feature matrix containing pressure, temperature, flow rate, and vibration information, which serves as the basis for subsequent risk feature modeling.

[0077] Through the dimensionality reduction process of the sparse autoencoder, the system can efficiently remove redundant information and retain the features crucial for risk identification. This feature compression significantly improves the efficiency of data processing, reduces computational complexity, and simultaneously reduces the risk of overfitting. In addition, the low-dimensional feature matrix after removing redundancy makes subsequent model training and prediction more efficient, ensuring that the system still has good response capabilities and prediction accuracy when facing complex oil and gas pipeline network data. This effective feature processing method provides strong technical support for the risk perception and management of oil and gas pipeline networks.

[0078] In this process, the system inputs the weighted spatio-temporal features into a sparse autoencoder to achieve dimensionality reduction of the features and removal of redundant information. The sparse autoencoder is an efficient neural network structure that can map high-dimensional data to a low-dimensional space through learning while retaining the most crucial information in the data. In the specific implementation process, the system first constructs the network structure of the sparse autoencoder, including the input layer, hidden layer, and output layer. The input layer directly receives the weighted spatio-temporal features, and the hidden layer is responsible for feature compression and mapping.

[0079] The system uses a non-linear activation function to perform non-linear transformation of the features in the hidden layer. For example, ReLU or LSTM activation functions can help the model capture complex feature associations. During the training process, the system optimizes the network weights through the backpropagation algorithm, with the goal of making the data reconstructed in the output layer as close as possible to the data in the input layer. To enhance sparsity, the system introduces L1 regularization during training to force some activations of neurons to be zero, thereby removing redundant features and only retaining key feature information.

[0080] Finally, after being processed by the sparse autoencoder, the system will generate a low-dimensional spatio-temporal feature matrix, which contains key pressure, temperature, flow rate, and vibration information. This transformation provides a more concise and efficient data input for subsequent risk modeling and anomaly detection, and is more suitable for the processing and analysis of large-scale data.

[0081] Exemplary:

[0082] 1. Multi-source data collection and preprocessing

[0083] 1.1. Sensor types

[0084] Pressure sensors (such as Honeywell ST3000 series, sampling frequency 1 Hz);

[0085] Temperature sensors (such as PT100 thermocouples, sampling frequency 0.5 Hz);

[0086] Flow meters (such as Emerson high-precision turbine flow meters, sampling frequency 2 Hz);

[0087] Vibration sensors (such as accelerometer ICP models, sampling frequency 10 Hz).

[0088] 1.2. Data alignment and missing value processing

[0089] Input: Pressure sequence P = [p1, p2,...] (1 Hz), temperature sequence T = [t1, t3,...] (0.5 Hz).

[0090] Operation: Calculate the optimal path to align the time axis through DTW, for example, interpolate the temperature sequence to 1 Hz.

[0091] Missing value filling: Use linear interpolation (for example, when t2 is missing, t2 = (t1 + t3) / 2).

[0092] Output: Synchronized multi-source data matrix, each row represents a timestamp, and each column represents a sensor parameter, as shown in Table 1:

[0093] Table 1

[0094] Timestamp Pressure (MPa) Temperature (°C) Flow rate (m³ / s) Vibration (g) t1 5.2 30.1 10.5 0.01 t2 5.3 30.5 10.7 0.02

[0095] 2. Spatio-temporal feature extraction

[0096] 2.1. Spatio-temporal convolutional neural network (ST-CNN)

[0097] Graph model construction: The nodes are pumping stations (such as nodes A, B), valves (such as V1, V2), and the edges represent pipeline connections.

[0098] Adjacency matrix: If node A is connected to V1, then A[0][1] = 1.

[0099] Feature aggregation: Use GCN layers (such as the PyTorch Geometric library) to aggregate the features of adjacent nodes.

[0100] Time dimension (1D convolution):

[0101] Convolution kernel size = 3, and a sliding window is used to extract the traffic time series features (such as local fluctuation trends).

[0102] 2.2 Multi-Head Self-Attention Mechanism

[0103] Input: Spatio-temporal feature matrix (dimension [time step × number of nodes × number of features]).

[0104] Operation: Generate Q, K, and V vectors through linear transformation (number of heads = 4); calculate the attention weights (such as the correlation weight between flow and pressure = 0.7); output the feature matrix after weighted fusion.

[0105] 2.3 Sparse Autoencoder (SAE)

[0106] Structure: Input layer (64 dimensions) → Hidden layer (32 dimensions, ReLU activation, L1 regularization) → Output layer (64 dimensions).

[0107] Output: A low-dimensional matrix after dimensionality reduction (such as 32 dimensions).

[0108] S202, Input the spatio-temporal feature matrix into the deep Gaussian process model to probabilistically model the abnormal features in the operation state of the pipe network, and output a risk feature vector with confidence;

[0109] In this step, the spatio-temporal feature matrix generated through the previous processing is input into the deep Gaussian process model to probabilistically model the abnormal features in the operation state of the oil and gas pipeline network. The advantage of the deep Gaussian process model lies in its ability to flexibly capture complex non-linear relationships in the data and handle uncertainties. By performing multi-layer non-linear transformations on the input spatio-temporal features, the model can map the original features to a high-dimensional latent space, thereby extracting deeper feature information and enhancing the feature expression ability. In this process, each layer of the network uses an activation function with residual connections to avoid the common gradient vanishing problem in the training of traditional deep networks, thus ensuring the effective training and stability of the model.

[0110] After the model training is completed, the system will construct a probability model of Gaussian process regression based on the extracted deep feature representation. This model can effectively capture the nonlinear relationship and periodic change law in the operation status of the pipeline network by combining the radial basis kernel function and the periodic kernel function. This structure enables the model to not only obtain the predicted mean of the feature, but also calculate the corresponding predicted variance, and finally generate a probability feature distribution with a confidence interval for each feature point. This information will help with subsequent risk assessment and decision making.

[0111] By inputting the spatiotemporal feature matrix into the deep Gaussian process model, the system can accurately identify and quantitatively evaluate potential abnormal features in the operation status of the pipeline network. The implementation of this step not only enhances the ability to perceive risks, but also provides high-confidence decision support in dynamic situations. By outputting risk feature vectors with confidence, the system can quickly judge the safety status of the oil and gas pipeline network in actual operation, provide a scientific basis for the formulation of emergency response strategies, thereby effectively reducing the risk of accidents and ensuring the safety and stability of the oil and gas pipeline network.

[0112] Specifically, the spatiotemporal feature matrix can be input into a multi-layer perceptron, and the original features can be mapped to a high-dimensional latent space through multi-layer nonlinear transformation to obtain a deep feature representation, wherein each layer uses an activation function with a residual connection to avoid the gradient vanishing problem and enhance the feature expression capability;

[0113] In this step, the system inputs the processed spatiotemporal feature matrix into a multilayer perceptron (MLP) to achieve deep learning and abstraction of the original features. The design of the multilayer perceptron includes multiple fully connected layers, and the neurons in each layer are connected to all the neurons in the previous layer. This structure enables the MLP to capture complex data patterns. Through layer-by-layer nonlinear transformations, the system can gradually map the original features into a high-dimensional latent space, thereby extracting deeper and more abstract feature information.

[0114] In order to solve the gradient vanishing problem in traditional deep networks, the system uses an activation function with residual connections in each layer. Residual connections allow the network to skip the direct output of a certain layer and combine the output of the previous layer with the output of the next layer, thereby effectively improving the training performance of the model. This structure can not only accelerate convergence, but also enhance the ability to express features, so that the model can still maintain good learning effects when facing high-dimensional features. After multiple layers of processing, the system will obtain a deep feature representation that contains rich information about the input spatiotemporal features.

[0115] The implementation of this process enables the system to perform efficient feature extraction and representation learning in a high-dimensional feature space, greatly enhancing the model's expressive power and the ability to recognize complex patterns. The resulting deep feature representation will provide a more comprehensive basis for subsequent probabilistic modeling, enabling more accurate identification of risk features, thereby assisting in the safety monitoring and risk management of oil and gas pipelines.

[0116] In the initial stage of this process, the system first designs the architecture of a multi-layer perceptron (MLP), which typically consists of an input layer, multiple hidden layers, and an output layer. The input layer is responsible for receiving the processed spatio-temporal feature matrix, which contains important information such as pressure, temperature, flow rate, and vibration. Each hidden layer uses a fully connected structure to take the output of the current layer as the input of the next layer. For example, assuming the feature dimension of the input layer is 32, after passing through the first hidden layer, the feature dimension may expand to 64, and this change is achieved through the learned weight matrix.

[0117] In each hidden layer, the system applies a non-linear activation function, such as ReLU (Rectified Linear Unit), to increase the non-linear expressive power of the model. This process also introduces the design of residual connections. Specifically, some features are directly passed from the previous layer to subsequent layers without undergoing the transformation of the current layer. This method effectively solves the problem of vanishing gradients that may occur in deep networks, enabling the network to be trained faster and more effectively. For example, in this design, if the output of a certain layer is small, the residual connection can retain the information of the previous layer, preventing the information from disappearing during propagation, thereby increasing the learning ability of the network.

[0118] Finally, after multiple non-linear transformations, the system will obtain a deep feature representation. This representation not only extracts the important information in the original features but also captures more complex feature correlations. These deep features will be used for anomaly detection in the subsequent Gaussian process regression model because they can provide more discriminative inputs, improving the overall performance and accuracy of the model.

[0119] For the said deep feature representation, a probability model based on Gaussian process regression is constructed, using a linear combination of a radial basis kernel function and a periodic kernel function to capture the non-linear relationships and periodic change patterns in the pipeline network operation state, obtaining a preliminary probabilistic feature representation;

[0120] In this step, the system constructs a Gaussian Process Regression (GPR) model based on the deep feature representation obtained in the previous step. Gaussian Process Regression is a flexible non-parametric Bayesian modeling method that can effectively handle complex functional relationships and is suitable for state prediction of oil and gas pipeline networks. When constructing the GPR model, the system uses a linear combination of the radial basis kernel function and the periodic kernel function. This combination strategy can capture both the non-linear relationships and the periodic variation patterns in the pipeline network operation state.

[0121] The radial basis kernel function can well represent the similarity between input features, and its characteristics enable the model to reflect the similarity of data points that are close to each other in the feature space. The periodic kernel function is specifically used to model the periodic patterns in the data. By introducing a periodic component into the Gaussian process, the system can better adapt to the periodic fluctuations in the pipeline network operation state. After being processed by these kernel functions, the system will obtain a preliminary probabilistic feature representation, which contains rich descriptions of the risk states.

[0122] By constructing a Gaussian Process Regression model, the system can synthesize the operation state of the oil and gas pipeline network into a probabilistic feature representation. This representation not only reflects the current state of the pipeline network but also can predict the future state and provide the confidence level of each feature point. This information not only enriches the dimension of risk assessment but also provides a reliable basis for subsequent anomaly detection, enabling the system to respond in a timely manner in a dynamic environment and reducing the risk of accidents.

[0123] In this step, the system constructs a Gaussian Process Regression (GPR) model using the deep feature representation. First, the system selects the kernel function, and here a combination of the radial basis kernel function and the periodic kernel function is used. The radial basis kernel function (RBF) can effectively capture the similarity between feature points, while the periodic kernel function is targeted at the periodic characteristics of the data, which is particularly important in an environment like the oil and gas pipeline network with seasonal variations. For example, the flow rate often shows different variation patterns in summer and winter, and combining the periodic kernel function can help the model accurately capture these characteristics.

[0124] Specifically in implementation, the system first defines the prior distribution of the Gaussian process and constructs the covariance matrix based on the training data. By linearly combining the radial basis kernel and the periodic kernel, the system can obtain a covariance function that comprehensively considers the influence of distance and periodicity. In this process, the system optimizes the hyperparameters of the kernel function, such as selecting appropriate length scales and periodicity parameters, to ensure the best performance of the model in practical applications. After the model training is completed, the system can generate a probability distribution representing the possible states of the input features under specific conditions.

[0125] After this kind of processing, the obtained preliminary probabilized feature representation can reflect the complexity in the operation state of the pipe network. For example, the flow characteristics of a certain node may show abnormal fluctuations within a specific time window, and the Gaussian process model can identify this abnormal state by measuring the gap between it and the normal feature distribution. The final output is the mean and variance corresponding to each feature point, forming a set of probabilistic features that can guide subsequent risk assessment.

[0126] According to the probabilized feature representation, calculate the predicted mean and variance of each feature point, and through the variational inference algorithm, quantify the uncertainty of the model to obtain a probabilistic feature distribution with a confidence interval.

[0127] In this process, the system uses the variational inference algorithm to further analyze the probabilized feature representation obtained from the Gaussian process regression model to calculate the predicted mean and variance of each feature point. Variational inference is an effective inference method mainly used to handle the uncertainty in the model. It approximates the true distribution by constructing an approximate distribution. In this step, the system will use an optimization algorithm to minimize the difference between the approximate posterior distribution and the true posterior distribution, thereby realizing the quantification of the model uncertainty and risk.

[0128] Specifically, the system will calculate the predicted mean and predicted variance of each feature point. Among them, the predicted mean represents the best estimate of the feature point state, while the predicted variance indicates the uncertainty of this estimate. This information is obtained through weighted calculation of the observed data. Furthermore, the system will construct a confidence interval based on the obtained variance to generate a probabilistic feature distribution for each feature point, reflecting the uncertainty characteristics under different risk states.

[0129] The implementation of this process enables the system to quantitatively evaluate the operation state of the oil and gas pipeline network and provides the uncertainty information of the prediction model. The probabilistic feature distribution with a confidence interval not only helps to identify potential risk points in the pipeline network but also provides a reliable basis for subsequent decision-making. This ability enables the operators of the oil and gas pipeline network to make more robust decisions when facing a complex environment, reduce potential safety hazards, and ensure the safety of operation.

[0130] In this step, the system will use the probabilized feature representation obtained from the Gaussian process regression model to calculate the predicted mean and variance of each feature point. In specific implementation, the system will apply the variational inference algorithm to each feature point to realize the quantification of the model uncertainty. Variational inference is an approximate inference technique. The system will set an appropriate prior distribution and optimize it using the observed data to obtain an approximation of the posterior distribution. The system maximizes the variational lower bound to optimize the parameters of the approximate posterior distribution to make it as close as possible to the true posterior distribution.

[0131] For example, the system can set a Gaussian distribution prior for each feature point, where the mean represents the possible state and the variance reflects the uncertainty of that state. During inference, the system calculates each feature point independently and updates the mean and variance based on the observed data by introducing sample points. Finally, the system will generate a probability feature distribution with a confidence interval, and the confidence interval provides a basis for quantifying the risk level.

[0132] This process improves the interpretability of the model because the predicted mean with a confidence interval not only indicates the possible state of each feature point but also enables decision-makers to understand the reliability of the data. By combining the mean and variance, the system can provide the necessary information for subsequent risk assessment, enabling managers to make more scientific decisions when judging risks.

[0133] For the probability feature distribution, an anomaly detection algorithm based on Mahalanobis distance is adopted to calculate the distance between each feature point and the normal operating condition distribution. According to a preset confidence threshold, abnormal feature points are screened out to generate a risk feature vector with confidence.

[0134] In this step, the system will analyze the obtained probability feature distribution through an anomaly detection algorithm based on Mahalanobis distance to identify abnormal feature points. Mahalanobis distance is an effective distance metric that can consider the covariance matrix of the data, thus better reflecting the relationship between different features. In this process, the system first needs to establish a distribution model of normal operating conditions, usually obtained through historical data statistics, and this model will be used as a reference standard.

[0135] Then, for each feature point, the system calculates its Mahalanobis distance from the normal operating condition distribution to evaluate the abnormality degree of the current feature point. If the Mahalanobis distance exceeds the preset confidence threshold, the system marks this feature point as abnormal. This process can efficiently screen out potential risk points and integrate these risk features into a risk feature vector with confidence, facilitating subsequent processing and decision-making.

[0136] Through the anomaly detection algorithm based on Mahalanobis distance, the system can effectively identify the abnormal state of the oil and gas pipeline network and timely feedback potential risks. The implementation of this step greatly enhances the safety monitoring ability of the oil and gas pipeline network, enabling the operator to take countermeasures immediately and reducing the risk of accidents. The risk feature vector with confidence provides strong data support for subsequent decision-making, making the system more flexible and efficient in the complex oil and gas pipeline network environment.

[0137] In this step, the system will adopt an anomaly detection algorithm based on Mahalanobis distance for the previously obtained probability feature distribution to identify potential anomaly feature points. First, the system will establish a data distribution model for normal operating conditions, which is usually achieved by statistically analyzing historical normal operation data. This distribution model will provide a reference standard for subsequent anomaly detection. The system will calculate the mean and covariance of each feature point in the normal state to determine the aggregation region of normal feature points.

[0138] Next, the system will calculate the Mahalanobis distance of each feature point to the normal operating condition distribution. Different from Euclidean distance, Mahalanobis distance can consider the correlation and covariance between features, making the distance metric more appropriate. During the calculation process, the system will use the Mahalanobis distance formula to calculate the Mahalanobis distance. Through such calculation, the system can effectively judge the deviation degree of each feature point compared with the normal state.

[0139] Finally, the system will screen out anomaly feature points according to a preset confidence threshold. If the calculated Mahalanobis distance exceeds this threshold, then this feature point is considered abnormal. After this process, the risk feature vector with confidence generated by the system will contain the feature points marked as abnormal and their relevant information. This not only provides data support for subsequent risk management but also helps the operation team take timely measures to prevent potential safety accidents.

[0140] Exemplary:

[0141] 1. Multilayer Perceptron (MLP) and Residual Connection

[0142] Technical Feature: Input the spatio-temporal feature matrix into a multilayer perceptron for non-linear transformation, and use an activation function with residual connection to avoid gradient vanishing.

[0143] Input Data: The dimension of the spatio-temporal feature matrix is 32 (for example, including 8 time-step features of pressure, temperature, flow rate, and vibration each).

[0144] Network Structure: Input Layer (32 dimensions) → Hidden Layer 1 (64 dimensions, ReLU activation, residual connection) → Hidden Layer 2 (128 dimensions, ReLU activation, residual connection) → Output Layer (64-dimensional depth features).

[0145] Residual Connection Design: The output of each layer is H(x) = F(x) + x, where F(x) is the result after transformation of the current layer.

[0146] Weight Initialization: He normal distribution (suitable for ReLU).

[0147] Batch Normalization: Add batch normalization after each layer to accelerate convergence.

[0148] Actual scenario: The data of the pressure sensor of an oil pipeline fluctuates within a time window. The MLP captures long-term dependencies through residual connections to avoid signal attenuation caused by deepening the number of layers.

[0149] 2. Gaussian Process Regression (GPR) combined with kernel functions

[0150] Technical feature: A linear combination of the Radial Basis Function (RBF) and the periodic kernel is used to model non-linearity and periodicity.

[0151] Kernel function definition: RBF kernel: k_RBF(x,x') = exp(-||x - x'||² / (2l²)), where the length scale l = 0.5; Periodic kernel: k_PER(x,x') = exp(-2sin²(π|x - x'| / p) / σ²), where the period p = 24 (daily period) and the variance σ = 1; Combined kernel: k_combined = 0.6*k_RBF + 0.4*k_PER.

[0152] Training data: Flow data of a natural gas pipeline (including daily peak periodic fluctuations).

[0153] Hyperparameter optimization: Optimize l, p, σ through maximum marginal likelihood estimation.

[0154] Output: The predicted average flow rate of a certain node in the next 3 hours is 50 m³ / s, the variance is 2.5, and the confidence interval is [45, 55].

[0155] 3. Variational inference algorithm for quantifying uncertainty

[0156] Technical feature: Calculate the mean and variance of feature points through variational inference to generate a confidence interval.

[0157] Setting of variational distribution: The approximate posterior q(f) ~ N(μ, Σ), where μ is the mean vector and Σ is the diagonal covariance matrix.

[0158] Optimization objective: Maximize the evidence lower bound (ELBO), using stochastic gradient descent (learning rate 1e-3).

[0159] Result example: The predicted mean of the pressure at the feature point is 10 MPa, the variance is 0.4, and the corresponding 95% confidence interval is [9.2, 10.8]. The temperature feature deviates from the confidence interval, triggering anomaly detection.

[0160] Application scenario: In a liquefied natural gas pipeline, variational inference quantifies the measurement uncertainty of temperature sensors to assist in determining whether it belongs to normal fluctuations.

[0161] 4. Anomaly detection based on Mahalanobis distance

[0162] Technical feature: Calculate the Mahalanobis distance between feature points and the normal distribution to screen for anomalies.

[0163] 4.1. Normal operating condition modeling

[0164] The mean value μ of historical data = [10 MPa, 50 °C, 200 m³ / h].

[0165] Covariance matrix Σ (calculated from 1000 sets of normal data):

[0166] Real-time data point: x = [12 MPa, 60 °C, 180 m³ / h].

[0167] 4.2. Mahalanobis distance calculation

[0168] Threshold setting: The chi-square distribution threshold corresponding to a 95% confidence level is 7.81 (3 degrees of freedom).

[0169] Result: 15.6 > 7.81, marked as a high-risk anomaly.

[0170] S203. According to the risk feature vector, construct a three-dimensional risk concentration distribution field, and predict the propagation path and diffusion trend of the risk concentration distribution field based on the graph neural network;

[0171] In this step, the system first maps each risk feature point to the three-dimensional spatial coordinates of the pipeline network according to the previously generated risk feature vector and in combination with the topological structure information of the oil and gas pipeline network. This process involves corresponding the nodes of the pipeline network (such as valves, pump stations, etc.) to the three-dimensional coordinate axes to visually represent the risk levels of each feature point in space. Subsequently, using the Kriging interpolation algorithm, spatial interpolation is performed on these discrete risk feature points to form a preliminary three-dimensional risk concentration distribution field. This distribution field reflects the risk levels and distribution conditions of the oil and gas pipeline network at different spatial positions. In addition, to dynamically capture the changes in risk, the system also combines the information in the time dimension and corrects the risk concentration through the spatio-temporal Kriging interpolation algorithm to ensure that the generated distribution field can reflect the time evolution of the risk.

[0172] By constructing a three-dimensional risk concentration distribution field, the system can visually display the risk status of the oil and gas pipeline network, enabling operators to quickly identify potential high-risk areas during monitoring. The real-time updated risk concentration distribution field provides data support for subsequent risk management, enabling decision-makers to monitor and respond more precisely when dealing with possible risk events. This process not only improves the accuracy of pollutant diffusion prediction but also provides a strong basis for dynamically adjusting the pipeline network operation strategy.

[0173] Specifically, according to the risk feature vector and combined with the topological structure information of the oil and gas pipeline network, each risk feature point can be mapped to the three-dimensional space coordinates of the pipeline network, and the Kriging interpolation algorithm is used to perform spatial interpolation on the discrete risk feature points to generate a preliminary three-dimensional risk concentration distribution field.

[0174] In this process, the system first needs to clarify the composition and topological structure of the oil and gas pipeline network. The topological structure usually includes the connection relationship of pipelines and the coordinate information of main nodes. The system will associate each risk feature point with these nodes to ensure that each feature point can find its corresponding position in three-dimensional space. With the help of the Kriging interpolation algorithm, these discrete risk feature points on the pipeline are integrated into a continuous risk concentration distribution field. Kriging interpolation is a statistical-based spatial interpolation technique that can use the spatial relationship of known feature points to estimate the values of unknown points. For example, if the pressure of a certain section of pipeline is abnormal, Kriging interpolation can use the information of surrounding normal points to predict the risk concentration of this section of pipeline, thus generating a continuous risk distribution field.

[0175] The implementation of this step will make the oil and gas pipeline network clear in space, enabling visual monitoring of risks in different regions. This intuitive visual presentation helps operators respond quickly, detect and handle potential leaks or other safety hazards in a timely manner. In addition, the generated three-dimensional risk concentration distribution field provides basic data for subsequent risk prediction and management, which will help enhance the safety and stability of the pipeline network.

[0176] In this step, the system first needs to obtain the topological structure information of the oil and gas pipeline network, including the three-dimensional space coordinates of key nodes such as pipelines, valves, and pumping stations. This information usually comes from the design drawings or digital models of the pipeline network. The system associates each risk feature point (for example, abnormal pressure or flow in a certain section of pipeline) with the corresponding three-dimensional space coordinates. To ensure that these feature points can be effectively displayed in three-dimensional space, the system will use identifiers to correspond the feature points to the actual pipeline network positions one by one.

[0177] Next, the system will use the Kriging interpolation algorithm to perform spatial interpolation on these discrete risk feature points. Kriging interpolation is an effective statistical interpolation method that can rely on the distribution of known points to predict the values of unknown points. Specifically, the system will construct a covariance matrix based on the risk values and spatial relationships of the feature points to characterize the mutual relationships between feature points. In this way, the system can construct a continuous risk concentration distribution field in the three-dimensional space of the pipeline network. For example, if the pressure value of a certain section of pipeline is significantly higher than other parts, the Kriging interpolation method can use the risk value of this point to infer the risk concentration of its surrounding area, thus forming a smooth risk concentration map.

[0178] Finally, after Kriging interpolation, the system will obtain a preliminary three-dimensional risk concentration distribution field. This distribution field will provide basic data for subsequent dynamic analysis and decision-making, allowing pipeline network operators to visually identify high-risk areas and quickly take actions to ensure the safety and stability of the pipeline network.

[0179] Based on the preliminary three-dimensional risk concentration distribution field and combined with time dimension information, the spatio-temporal Kriging interpolation algorithm is used, combined with historical risk propagation data, to dynamically correct the risk concentration distribution field and obtain a three-dimensional risk concentration distribution field that changes over time.

[0180] In this step, the system will combine historical risk propagation data and introduce the dynamic time dimension into the generation process of the risk concentration distribution field. First, the system will extract the risk propagation trend from historical data, such as the risk evolution in specific conditions in the past few months. These data will provide valuable reference for the system. Next, the system will apply the spatio-temporal Kriging interpolation algorithm, which can introduce time factors on the basis of spatial interpolation to dynamically correct the risk concentration distribution field. Through the analysis of time series, the system can capture how the risk concentration changes over time. For example, if a certain section of pipeline had risks due to abnormal temperature in a past time period, then in future risk assessments, the system will adjust the current risk distribution according to the change law of historical data.

[0181] By introducing the time dimension, the system can generate a risk concentration distribution field with strong timeliness, which ensures that risks changing over time can be captured and updated in a timely manner. This dynamic feature not only improves the accuracy of risk monitoring but also provides a more reliable data basis for subsequent risk assessment and decision-making. Operators can actively take measures by analyzing the change trend of the time dimension to reduce potential risks and ensure the safe and stable operation of the oil and gas pipeline network.

[0182] In this step, the system combines the preliminary three-dimensional risk concentration distribution field with time dimension information for dynamic correction. First, the system will extract historical risk propagation data, including risk event records in the past few months or years. These data can include the change of risk characteristics at each node, such as leakage events and flow fluctuations at a specific time point for a certain section of pipeline. These historical data will be used to construct a time series model to analyze the risk propagation trend in the pipeline network.

[0183] Next, the system will use the spatio-temporal Kriging interpolation algorithm to achieve this dynamic correction. The spatio-temporal Kriging interpolation algorithm is different from traditional spatial interpolation methods. It takes into account the influence of time and can better reflect the change process of risk concentration. The system will input historical risk propagation data, establish a time series model, analyze the change pattern of risk over time, and thus calculate the risk concentration in a future time period under the current pipeline network state. For example, if historical data shows that a certain section of pipeline is prone to leakage during the rainy season, the system will take this seasonal trend into account and adjust the current risk concentration prediction accordingly.

[0184] Finally, after dynamic correction, the system will obtain a three-dimensional risk concentration distribution field that changes over time. This dynamic risk concentration distribution field can provide more accurate information support for the management and maintenance of oil and gas pipeline networks, enabling operators to take timely preventive measures based on the latest risk status and reducing the probability of accidents.

[0185] Input the three-dimensional risk concentration distribution field that changes over time into the graph neural network, construct a graph model according to the pipeline network topology structure, where the nodes of the graph model represent the key positions of the pipeline network, the edges of the graph model represent the pipe segment connection relationships, and use the graph attention mechanism to capture the risk propagation dependence relationships between nodes to obtain the initial prediction result of risk propagation;

[0186] In this step, the system first inputs the three-dimensional risk concentration distribution field that changes over time into the graph neural network (GNN) to establish a graph model that reflects the topology structure of the oil and gas pipeline network. Specifically, the nodes in the graph model represent the key positions of the pipeline network (such as pumping stations, valves, etc.), and the edges represent the connection relationships of pipe segments. Through this representation, the system can better understand the internal relationships and dependencies within the pipeline network. Then, the system will use the graph attention mechanism to capture the risk propagation dependence relationships between different nodes. The graph attention mechanism can dynamically adjust its influence weight with adjacent nodes by learning the importance of each node. For example, assuming that the risk level of a certain node is relatively high, the system will pay more attention to other nodes connected to this node to ensure that important information is not missed in the risk propagation analysis.

[0187] By combining the graph neural network with the topology structure of the pipeline network, the system can discover the mutual influences between each node and their corresponding risk propagation paths. This ability enables decision-makers to more clearly identify the direction and potential impact of risk propagation, enhancing the scientific nature and accuracy of real-time monitoring and decision-making. In addition, the initial prediction result of risk propagation will provide a basis for subsequent risk management measures, enabling the management team to respond in a timely manner and prevent accidents from occurring.

[0188] In this step, the system inputs the three-dimensional risk concentration distribution field that changes over time into a graph neural network (GNN) to construct a graph model. First, the system needs to clarify the topological structure of the oil and gas pipeline network and define the nodes and edges of the graph model. Nodes generally represent key positions in the pipeline network, such as pumping stations, valves, and monitoring points, while edges represent the connection relationships between these nodes. The system combines the risk eigenvalue with the nodes to establish a graph model representing the risk state.

[0189] Next, the system uses the graph attention mechanism to analyze and capture the risk propagation dependence relationships between different nodes. The graph attention mechanism is a dynamic weighting model that can assign different importance weights to each node, so as to pay more attention to those nodes that have a greater impact on the final result in the risk propagation analysis. For example, if the risk value of a certain valve is high and this valve is connected to multiple pipe segments, the system will assign a higher weight to the nodes connected to this valve to ensure that its impact on the risk propagation analysis is fully considered.

[0190] Finally, after being processed by the graph neural network, the system will obtain the initial prediction results of risk propagation. These results can not only reveal the distribution characteristics of the current risk in the pipeline network, but also reflect the possible propagation paths and impact areas, providing a reference for subsequent risk countermeasures.

[0191] According to the initial prediction results of risk propagation, a sequence prediction model based on a temporal convolutional network is adopted. Combining with the historical risk propagation path data, the risk propagation path and diffusion trend in future time steps are predicted to generate the propagation prediction results of the dynamic risk concentration distribution field.

[0192] In this step, the system uses a temporal convolutional network (TCN) to further analyze and predict the preliminary prediction results of risk propagation. The temporal convolutional network is a model specialized for processing time series data, which captures the temporal features in the data through convolutional operations. According to the preliminary prediction results, the system inputs the historical risk propagation path data and combines the advantages of the temporal convolutional network to generate the risk propagation prediction for future time steps. In this process, the model will be trained based on historical data to learn the propagation pattern and diffusion trend of risks. For example, by analyzing the risk changes in the past period of time, the system can judge how the risk will propagate in the pipeline network under specific conditions.

[0193] The implementation of this process enables the system to predict the future risk state, significantly improving the ability of managers to respond to emergencies. By generating the propagation prediction results of the dynamic risk concentration distribution field, the system can identify high-risk areas in advance, giving decision-makers sufficient time to take countermeasures. In addition, such predictions not only reduce the probability of accidents, but also improve the overall safety of the oil and gas pipeline network, ensuring the stable energy supply.

[0194] In this step, the system will use a sequence prediction model based on Temporal Convolutional Network (TCN), leveraging the initial prediction results of risk propagation obtained previously and historical risk propagation path data. First, the system needs to organize the historical risk propagation data to provide rich training information for the TCN model. This historical data may include risk concentration values at different time points and their corresponding propagation paths, such as the risk distribution of a certain pipeline section during a certain time period.

[0195] Next, the system will input this data into the Temporal Convolutional Network. TCN uses convolutional layers to process time series data and can effectively capture features in the time dimension. During the input of data, the system takes into account the temporal sequence of risk propagation and trains through multiple time windows to ensure that the model can learn the potential propagation patterns in the historical data. For example, if the system identifies that a certain risk spreads rapidly under specific conditions, TCN will remember this pattern and apply it when predicting future risks.

[0196] Finally, the system outputs the risk propagation path and diffusion trend for future time steps predicted by TCN. In this way, the system can not only provide predictions on the future development of risks but also form a propagation prediction result of a dynamic risk concentration distribution field. These prediction results will provide necessary data support for the operation and management of oil and gas pipe networks, enabling decision-makers to actively respond to potential risks, optimize operation strategies, and ensure the safe and efficient operation of the pipe networks.

[0197] Exemplary:

[0198] 1. Data Acquisition and Generation of Spatiotemporal Feature Matrix

[0199] 1.1 Multi-source Sensor Data

[0200] Pressure sensors: P1(50MPa), P2(48MPa), P3(52MPa);

[0201] Temperature sensors: T1(25℃), T2(28℃);

[0202] Flow meters: F1(100m³ / s), F2(105m³ / s);

[0203] Vibration sensors: V1(0.5mm / s²), V2(0.6mm / s²).

[0204] 1.2 Dynamic Time Warping (DTW) Alignment

[0205] Assume that the sampling frequency of the pressure sensor is 1Hz and that of the flow meter is 0.5Hz. The DTW algorithm stretches and aligns the time axis of the flow data to the time axis of the pressure data to ensure time synchronization.

[0206] 1.3, Spatiotemporal Convolutional Neural Network (ST-CNN)

[0207] Spatial dimension: The graph convolutional network (GCN) processes the pipeline network topology (such as pipeline connection relationships: node A → node B → node C).

[0208] Temporal dimension: A one-dimensional convolutional kernel (window size = 5) extracts the periodic fluctuation characteristics of the flow rate data.

[0209] Output: A spatiotemporal feature matrix (size: 10×10×4, containing four channels of pressure, temperature, flow rate, and vibration).

[0210] 2. Deep Gaussian Process Model and Risk Feature Vector

[0211] 2.1, Multilayer Perceptron (MLP)

[0212] Input the spatiotemporal feature matrix, map it to a high-dimensional space through a 3-layer residual network (each layer dimension: 256→512→256), and output a deep feature representation (dimension: 128).

[0213] 2.2, Gaussian Process Regression (GPR)

[0214] Kernel function combination: RBF kernel (length scale = 1.0) + periodic kernel (period = 24 hours).

[0215] Prediction result: The average pressure of a certain pipe segment is 51 MPa, and the variance is 0.5 (confidence interval: 50.2~51.8 MPa).

[0216] 2.3, Mahalanobis Distance Anomaly Detection

[0217] Calculate the distance D = 2.1 between the current pressure feature and the normal operating condition (mean 50 MPa, covariance matrix Σ).

[0218] If the threshold = 2.0, then it is determined as abnormal, and a risk feature vector [risk type = high pressure, confidence level = 90%] is generated.

[0219] 3. Construction of a Three-Dimensional Risk Concentration Distribution Field

[0220] 3.1, Kriging Interpolation

[0221] Discrete risk point coordinates: Node A(1,2,0), Node B(2,3,0), and the risk values are 0.8 and 0.5 respectively.

[0222] After interpolation, a continuous field is generated, such as the risk concentration at coordinate (1.5,2.5,0) = 0.65.

[0223] 3.2, Spatiotemporal Kriging Correction

[0224] Combined with historical data (such as the risk at node A during the rainy season last year increased by 20%), the current risk concentration is dynamically adjusted to 0.78.

[0225] 4. Graph Neural Network and Risk Propagation Prediction

[0226] 4.1. Graph Model Construction

[0227] Nodes: {Pumping Station 1, Valve 2, Monitoring Point 3}

[0228] Edges: {(Pumping Station 1 → Valve 2, length = 5 km), (Valve 2 → Monitoring Point 3, length = 3 km)}

[0229] 4.2. Graph Attention Mechanism (GAT)

[0230] The risk weight of Pumping Station 1 = 0.7, Valve 2 = 0.3, predicted risk propagation path: Pumping Station 1 → Valve 2 → Monitoring Point 3.

[0231] 4.3. Temporal Convolutional Network (TCN)

[0232] Input historical risk path data (such as the risk diffusion speed in the past 24 hours = 0.5 km / h), predict that the risk will cover Monitoring Point 3 in the next 6 hours.

[0233] S204. For the prediction results of the risk concentration distribution field, construct an adaptive early warning threshold generation model based on the generative adversarial network. Through the dynamic game between the normal condition generator and the real-time discriminator, generate multi-level early warning thresholds that change with the operation state of the pipeline network, and based on the comparison results of the real-time risk characteristics and the early warning thresholds, realize the risk perception and early warning of the operation state of the oil and gas pipeline network.

[0234] In this step, the system realizes the dynamic monitoring and risk perception of the operation state of the oil and gas pipeline network by constructing an adaptive early warning threshold generation model based on the generative adversarial network (GAN). This model mainly consists of two parts: a normal condition generator and a real-time discriminator. The task of the generator is to generate simulated data that conforms to the normal state of the pipeline network according to historical normal operation data. These data not only need to reflect the typical characteristics of the pipeline network but also have a certain degree of diversity and authenticity to better simulate the real operation environment. The real-time discriminator is responsible for comparing the generated normal condition data with the prediction results of the risk concentration distribution field, learning to identify the boundary characteristics between normal and abnormal states, so as to improve the accuracy of the model. This process continuously optimizes the two parts through a dynamic game, making the early warning threshold more targeted and effective.

[0235] The early warning thresholds generated based on this dynamic game mechanism can timely reflect the real-time operation status of the pipeline network, enabling the risk management team to quickly identify potential risk areas and take corresponding preventive measures. Specifically, as the operation status of the oil and gas pipeline network changes, the system will generate corresponding low-risk, medium-risk, and high-risk thresholds, which will serve as important reference standards for the monitoring system to help decision-makers make scientific judgments and decisions at critical moments. In this way, the safety and stability of the oil and gas pipeline network are further improved, effectively reducing the probability of accidents and ensuring the smooth energy transmission.

[0236] Specifically, based on historical normal operation data, a normal condition generator based on a generative adversarial network can be trained to generate simulation data that conforms to the normal operation characteristics of the pipeline network. Among them, the generator uses a variational autoencoder with conditional constraints to ensure the diversity and authenticity of the generated data.

[0237] In this step, the system will use historical normal operation data to train the generator to generate simulation data that conforms to the normal state of the oil and gas pipeline network. The design of the generator is based on the variational autoencoder (VAE) architecture. By introducing conditional constraints, the system can ensure that the generated data not only has diversity but also truly reflects the operation characteristics of the pipeline network. The key to this method is to encode historical data to obtain the distribution of the latent space, thereby generating new samples. Specifically, the system will extract various features of the samples, including pressure, flow rate, temperature, and vibration, etc. Through structured data processing means, the system can form a representative dataset for the generator to learn.

[0238] By generating simulation data that conforms to historical data, the system can still maintain the monitoring of the normal state of the oil and gas pipeline network in the absence of real-time data. This powerful generation ability enables managers to fully consider various changes under normal operation conditions when evaluating the safety of the pipeline network, providing a scientific basis for subsequent risk assessment. In addition, the generated diverse data helps to identify potential abnormal behaviors, providing more comprehensive information support for the risk monitoring system.

[0239] In this step, the system first collects historical normal operation data, including information such as pressure, flow rate, temperature, and vibration. These data come from various monitoring sensors of the oil and gas pipeline network. For example, the pipeline pressure value recorded by the pressure sensor, the flow rate data monitored by the flow meter, and the real-time data provided by the temperature and vibration sensors. In the data collection stage, the system needs to preprocess the original data, such as removing outliers and filling missing values, to ensure the quality and accuracy of the data. These processed data will become the training samples of the generator.

[0240] Next, a normal operating condition generator based on a Generative Adversarial Network (GAN) is constructed. The generator adopts the architecture of a Conditional Variational Autoencoder (CVAE) with conditional constraints, which can ensure that the generated data is both diverse and conforms to the real operating characteristics. The generator uses the learned features together with the conditional input to generate new samples. For example, if the input condition is "normal operating pressure", the generator will generate diverse pressure values and corresponding flow rates, temperatures, etc. under this condition based on historical data. During the model training process, the generator will be continuously optimized to reduce the gap between the generated data and the real data.

[0241] Finally, by comparing with the real data, the system will continuously adjust the parameters of the generator to make its output more realistic simulated data. This process will be achieved by using adversarial loss and reconstruction loss to ensure that the generated data is not only statistically similar to the real data but also physically meaningful and usable. The trained generator can generate a large number of simulated data representing normal operating conditions, which will provide a reliable basis for risk prediction and early warning strategies.

[0242] A real-time discriminator based on a deep residual network is constructed. The prediction results of the risk concentration distribution field and the normal operating condition data generated by the generator are input. Through dynamic game learning, the boundary features between normal and abnormal states are learned, and the discrimination results and their confidence levels are output.

[0243] In this step, the system constructs a real-time discriminator, adopting the architecture of a deep residual network to improve the model's recognition ability. The main function of the discriminator is to compare the prediction results of the input risk concentration distribution field with the normal operating condition data generated by the generator, and learn to distinguish what is a normal state and what is an abnormal state. During the dynamic game process, the discriminator will continuously optimize its performance and improve the discrimination accuracy of the input data. Specifically, the system will train the discriminator to recognize the boundary features between the two by comparing normal data with abnormal risk features, ensuring that the model can make effective risk judgments in practical applications.

[0244] The establishment of the real-time discriminator will greatly enhance the risk monitoring ability of the oil and gas pipeline network. By accurately identifying normal and abnormal states, the system can early warn of potential risks and provide key decision-making support for managers. This function not only improves the scientific nature of risk management but also effectively reduces safety accidents caused by misjudgment, ensuring the smooth operation and safety and stability of the pipeline network.

[0245] In this step, the system needs to construct a real-time discriminator based on the Deep Residual Network (ResNet). The main task of the discriminator is to distinguish the normal operating condition data generated by the generator from the real-time risk concentration distribution field prediction results. First, the system will design the network structure to consist of multiple residual blocks to ensure the effective transmission of information in the deep network and avoid the problem of gradient disappearance. The discriminator will take as input the normal operating condition samples generated by the generator and the real-time risk data, extract key features through multiple non-linear transformations, and establish the boundary between normal and abnormal states.

[0246] During the training process, the discriminator will conduct dynamic games, that is, continuously self-update when competing with the generator. If the data generated by the generator is realistic, the discriminator will have difficulty distinguishing it from the real data; conversely, if the data generated by the generator is not realistic enough, the discriminator will be able to easily identify it. Through this adversarial training, the discriminator can more accurately learn the feature differences between normal and abnormal states. For example, when there are abnormal fluctuations in real-time sensor data, the discriminator can promptly identify this state and mark it as abnormal.

[0247] Finally, after multiple rounds of training, the discriminator will output the discrimination results and confidence levels for each input data. The higher the confidence level value, the more certain the model is that the sample is in a normal or abnormal state. In this way, the system can monitor the operating state of the pipe network in real time and promptly detect potential risks.

[0248] According to the discrimination results, an adaptive clustering algorithm is used to perform multi-level partitioning on the discrimination results. Combining with the dynamic changes of the risk concentration distribution field, multi-level warning thresholds that change with the operating state of the pipe network are generated. The multi-level warning thresholds include low-risk, medium-risk, and high-risk thresholds;

[0249] In this step, the system will rely on the results output by the discriminator and apply an adaptive clustering algorithm to perform multi-level partitioning of the risk states. The advantage of the adaptive clustering algorithm is that it can automatically determine the number and categories of clusters according to the distribution of the current data, making the classification results more reasonable and effective. Combining with the dynamic changes of the risk concentration distribution field, the system will generate multi-level warning thresholds, including low-risk, medium-risk, and high-risk thresholds. These thresholds not only reflect the risk characteristics of the current operating state of the pipe network but also provide a quantitative reference standard for real-time monitoring.

[0250] Through the setting of multi-level warning thresholds based on clustering, the system can achieve a detailed division of the risks in the operating state of the oil and gas pipeline network. This not only improves the response speed of the warning system but also enables managers to arrange monitoring and emergency measures in a targeted manner. The multi-level warning mechanism ensures that decision-makers can actively respond under different risk states, thereby improving the safety management level of the oil and gas pipeline network. At the same time, this method reduces the subjectivity of human judgment and provides support for scientific decision-making.

[0251] In this step, the system will apply an adaptive clustering algorithm to perform multi-level partitioning on the discrimination results according to the results output by the discriminator. The key feature of the adaptive clustering algorithm is its ability to automatically determine the number and categories of clusters based on the distribution of data. The system will first extract abnormal and normal sample data from the discrimination results and sort them according to their confidence scores. Next, the system will perform feature extraction on these data, including their time features, spatial features, etc., for more effective clustering analysis.

[0252] By introducing the adaptive clustering algorithm, the system can divide the data into three main categories: low risk, medium risk, and high risk based on the discrimination results. For example, samples with confidence scores higher than a certain threshold in the discrimination results can be classified as low risk, those in the middle range are medium risk, and those lower than the threshold are classified as high risk. This division will help managers take corresponding response measures in a timely manner under different states, thereby optimizing the risk management strategy.

[0253] Finally, the system will integrate the generated multi-level warning thresholds into a dynamically updated system so that when the operating state of the oil and gas pipeline network changes, the warning thresholds at all levels can be adjusted in real time. Combining with the dynamic changes in the risk concentration distribution field, the system can ensure the effectiveness and accuracy of risk monitoring, thus better coping with potential risks.

[0254] Compare the real-time risk features with the multi-level warning thresholds, adopt a decision-making model based on fuzzy logic, and dynamically adjust the warning level according to the confidence of the risk features and the degree of deviation from the threshold to achieve risk perception warning for the operating state of the oil and gas pipeline network.

[0255] In this step, the system will compare the real-time risk features with the generated multi-level warning thresholds and dynamically adjust the warning level in combination with a decision-making model based on fuzzy logic. The fuzzy logic decision-making model can handle uncertainty and ambiguity. Based on the analysis of the confidence of the risk features and their degree of deviation from the warning thresholds, the system will generate a more accurate warning level. For example, if the real-time risk features are significantly higher than the high-risk threshold, the system will issue a high-risk warning; if the risk features are within the low-risk range, the system may remain in a normal state.

[0256] The dynamic adjustment decision based on fuzzy logic can significantly improve the risk management ability of the oil and gas pipeline network. This mechanism not only ensures the efficiency of real-time monitoring but also helps managers respond in a timely manner in a complex and changing operating environment, reducing the probability of risk accidents. Through accurate risk perception and warning, the system improves the safe operation level of the pipeline network, creating conditions for ensuring the stability of energy supply.

[0257] In this step, the system compares the risk characteristics collected in real time with the generated multi-level warning thresholds, and adopts a decision-making model based on fuzzy logic to dynamically adjust the warning level. First, the system integrates and compares the real-time risk characteristic data with multiple warning thresholds to evaluate the risk level of the current pipeline network. For example, when the pressure value of a certain section of pipeline is detected to exceed the medium-risk threshold in real time, the system will mark this section of pipeline as a high-risk area.

[0258] Next, the fuzzy logic model calculates by considering the confidence level of the risk characteristics and the degree of deviation from the warning threshold. By establishing fuzzy rules, such as "if the risk characteristics are higher than the medium-risk threshold and the confidence level is high, then the warning level is set to high risk", the system can flexibly respond to different situations. This flexibility is a significant advantage of the fuzzy logic model because it can handle uncertainty and ambiguity, making the decision-making more scientific and reasonable.

[0259] Finally, based on the calculation results, the system will dynamically adjust the warning level and output the corresponding warning information. These information will be presented to the pipeline network monitoring personnel through a visual interface to help them make decisions and take measures in a timely manner. For example, dispatching more monitoring personnel, reducing pressure, and performing equipment maintenance. This risk perception warning mechanism will significantly improve the safety of oil and gas pipeline networks, reduce the risk of accidents, and ensure the stable operation of the pipeline network.

[0260] 1. Historical data collection and preprocessing

[0261] 1.1. Example data sources

[0262] Pressure data: Pressure sensor (model PTX5000) for a certain section of the natural gas pipeline in Xi'an, sampling frequency 1Hz, normal range 4 - 6MPa.

[0263] Flow data: Turbine flowmeter (model TUF - 2000), data interval 5 seconds, normal flow range 20 - 30m³ / s.

[0264] Temperature data: Infrared temperature sensor (model IR - TC300), sampling frequency 0.5Hz, normal temperature range 10 - 50°C.

[0265] Vibration data: Accelerometer (model ADXL345), sampling frequency 100Hz, normal vibration amplitude <0.5g.

[0266] 1.2. Preprocessing steps

[0267] Missing value filling: For the 5 - minute missing data in the temperature data, cubic spline interpolation is used for filling.

[0268] Outlier removal: Use the 3σ principle to remove the outliers (such as a sudden 8MPa peak) in the pressure data that exceed ±3 standard deviations.

[0269] 2. Training of the Normal Condition Generator (CVAE)

[0270] 2.1 Model Architecture

[0271] Encoder: A 3 - layer fully connected network (256 nodes in the input layer, 128 nodes in the hidden layer), outputting the mean μ and variance σ of the latent variable.

[0272] Decoder: A 3 - layer transposed convolution network, with the input condition "flow rate = 25 m³ / s ± 5%", generating corresponding pressure, temperature, and vibration data.

[0273] Loss function: ELBO (Evidence Lower Bound) loss + KL divergence to constrain the latent space distribution.

[0274] 2.2 Example of Generated Data

[0275] With the input condition "pressure = 5 MPa", 10 groups of simulated data are generated:

[0276] Flow rate: 24.8 m³ / s, 25.1 m³ / s,...;

[0277] Temperature: 32 °C, 31.5 °C,...;

[0278] Vibration: 0.3 g, 0.35 g,...

[0279] 3. Construction of the Real - Time Discriminator (ResNet)

[0280] 3.1 Network Structure

[0281] Residual block: 4 residual modules, each module containing 2 convolutional layers (kernel size 3×3) + skip connection.

[0282] Input: Risk concentration distribution field (64×64 grid) + simulated data output by the generator.

[0283] Output: Discrimination result (0 / 1) and confidence level (0 - 1, e.g., 0.92 indicates "normal").

[0284] 3.2 Example of Dynamic Game

[0285] The generator generates a set of data with "pressure = 7 MPa (abnormal)", and the discriminator determines it as "abnormal" with a confidence level of 0.85.

[0286] After the generator adjusts the parameters, it generates data with "pressure = 5.2 MPa", and the confidence level of the discriminator drops to 0.55 (close to misjudgment).

[0287] 4. Adaptive Clustering (DBSCAN Algorithm)

[0288] 4.1 Parameter Settings

[0289] The neighborhood radius ε = 0.5 and the minimum number of samples min_samples = 10.

[0290] Input: Confidence levels of 500 groups of data output by the discriminator (such as [0.1, 0.3, 0.8,...]).

[0291] 4.2 Clustering Results

[0292] Low risk: Confidence level > 0.7 (such as 0.8, 0.9), corresponding pressure < 5.5 MPa.

[0293] Medium risk: 0.4 ≤ Confidence level ≤ 0.7 (such as 0.6), corresponding pressure 5.5 - 6.5 MPa.

[0294] High risk: Confidence level < 0.4 (such as 0.2), corresponding pressure > 6.5 MPa or vibration > 1 g.

[0295] 5. Fuzzy Logic Decision Model

[0296] 5.1 Rule Base Example

[0297] IF the pressure deviation from the threshold > 10% AND the confidence level > 0.8 THEN high - risk warning.

[0298] IF the flow rate fluctuates ±5% AND the temperature is normal THEN medium - risk warning.

[0299] 5.2 Dynamic Adjustment Case

[0300] Real - time data: Pressure = 6.8 MPa (exceeding the threshold by 13%), confidence level 0.75.

[0301] Fuzzy output: Risk value = 0.82 (high risk), triggering a red alarm and starting the pressure reducing valve.

[0302] 6. Visualization and Warning Output

[0303] 6.1 System Interface Example

[0304] Three - dimensional risk field: In the three - dimensional model of the pipeline, high - risk sections are shown in red (such as the Xi'an section), and low - risk sections are green.

[0305] 6.2 Warning Log

[0306] [2023 - 10 - 01 14:30] Warning ID: #2023 - 001;

[0307] Location: KP32 + 500 of the Xi'an section;

[0308] Risk level: High risk (abnormal pressure);

[0309] Suggested measure: Immediately repair Valve No. 3.

[0310] It can be seen that, based on the real-time operation data of the oil and gas pipeline network, the multi-dimensional features of the pipeline network operation state are fused to generate a spatio-temporal feature matrix; the spatio-temporal feature matrix is input into the deep Gaussian process model to probabilistically model the abnormal features in the pipeline network operation state, and a risk feature vector with confidence is output; according to the risk feature vector, a three-dimensional risk concentration distribution field is constructed, and the propagation path and diffusion trend of the risk concentration distribution field are predicted; for the prediction results, an adaptive early warning threshold generation model based on the generative adversarial network is constructed to generate multi-level early warning thresholds that change with the pipeline network operation state, and based on the comparison results of the real-time risk features and the early warning thresholds, the risk perception early warning of the oil and gas pipeline network operation state is realized, so as to achieve dynamic risk assessment and adaptive early warning, and enhance the safety and emergency response ability of the pipeline network.

[0311] Another embodiment of the present invention provides a risk perception early warning system for the operation state of an oil and gas pipeline network. Refer to Figure 3 , the system may include:

[0312] A fusion module 301, configured to fuse the multi-dimensional features of the pipeline network operation state according to the real-time operation data of the oil and gas pipeline network, and generate a spatio-temporal feature matrix including pressure, temperature, flow rate, and vibration information;

[0313] A modeling module 302, configured to input the spatio-temporal feature matrix into the deep Gaussian process model to probabilistically model the abnormal features in the pipeline network operation state, and output a risk feature vector with confidence;

[0314] A construction module 303, configured to construct a three-dimensional risk concentration distribution field according to the risk feature vector, and predict the propagation path and diffusion trend of the risk concentration distribution field based on the graph neural network;

[0315] An early warning module 304, configured to construct an adaptive early warning threshold generation model based on the generative adversarial network for the prediction results of the risk concentration distribution field, generate multi-level early warning thresholds that change with the pipeline network operation state through the dynamic game between the normal working condition generator and the real-time discriminator, and realize the risk perception early warning of the oil and gas pipeline network operation state based on the comparison results of the real-time risk features and the early warning thresholds.

[0316] It can be seen that according to the real-time operation data of the oil and gas pipeline network, the multi-dimensional features of the pipeline network operation state are fused to generate a spatio-temporal feature matrix; the spatio-temporal feature matrix is input into the deep Gaussian process model to probabilistically model the abnormal features in the pipeline network operation state, and a risk feature vector with confidence is output; according to the risk feature vector, a three-dimensional risk concentration distribution field is constructed, and the propagation path and diffusion trend of the risk concentration distribution field are predicted; for the prediction result, an adaptive early warning threshold generation model based on a generative adversarial network is constructed to generate multi-level early warning thresholds that change with the pipeline network operation state, and based on the comparison result between the real-time risk features and the early warning thresholds, the risk perception early warning of the oil and gas pipeline network operation state is realized, so that dynamic risk assessment and adaptive early warning can be achieved, and the safety and emergency response ability of the pipeline network are enhanced.

[0317] An embodiment of the present invention further provides a storage medium, in which a computer program is stored, and the computer program is configured to execute the steps in any one of the above method embodiments when running.

[0318] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for executing the following steps:

[0319] S201, according to the real-time operation data of the oil and gas pipeline network, fuse the multi-dimensional features of the pipeline network operation state to generate a spatio-temporal feature matrix including pressure, temperature, flow rate, and vibration information;

[0320] S202, input the spatio-temporal feature matrix into the deep Gaussian process model to probabilistically model the abnormal features in the pipeline network operation state, and output a risk feature vector with confidence;

[0321] S203, according to the risk feature vector, construct a three-dimensional risk concentration distribution field, and predict the propagation path and diffusion trend of the risk concentration distribution field based on a graph neural network;

[0322] S204, for the prediction result of the risk concentration distribution field, construct an adaptive early warning threshold generation model based on a generative adversarial network, generate multi-level early warning thresholds that change with the pipeline network operation state through the dynamic game between the normal working condition generator and the real-time discriminator, and based on the comparison result between the real-time risk features and the early warning thresholds, realize the risk perception early warning of the oil and gas pipeline network operation state.

[0323] It can be seen that according to the real-time operation data of the oil and gas pipeline network, the multi-dimensional characteristics of the pipeline network operation state are fused to generate a spatio-temporal feature matrix; the spatio-temporal feature matrix is input into the deep Gaussian process model to probabilistically model the abnormal characteristics in the pipeline network operation state, and a risk feature vector with confidence is output; according to the risk feature vector, a three-dimensional risk concentration distribution field is constructed, and the propagation path and diffusion trend of the risk concentration distribution field are predicted; for the prediction result, an adaptive early warning threshold generation model based on the generative adversarial network is constructed to generate multi-level early warning thresholds that change with the pipeline network operation state, and according to the comparison result between the real-time risk characteristics and the early warning thresholds, the risk perception early warning of the oil and gas pipeline network operation state is realized, so as to achieve dynamic risk assessment and adaptive early warning, and enhance the safety and emergency response ability of the pipeline network.

[0324] An embodiment of the present invention further provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0325] Specifically, the above electronic device may further include a transmission device and an input / output device. Among them, the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0326] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program:

[0327] S201, according to the real-time operation data of the oil and gas pipeline network, fuse the multi-dimensional characteristics of the pipeline network operation state to generate a spatio-temporal feature matrix including pressure, temperature, flow rate, and vibration information;

[0328] S202, input the spatio-temporal feature matrix into the deep Gaussian process model to probabilistically model the abnormal characteristics in the pipeline network operation state, and output a risk feature vector with confidence;

[0329] S203, according to the risk feature vector, construct a three-dimensional risk concentration distribution field, and predict the propagation path and diffusion trend of the risk concentration distribution field based on a graph neural network;

[0330] S204, for the prediction result of the risk concentration distribution field, construct an adaptive early warning threshold generation model based on the generative adversarial network, generate multi-level early warning thresholds that change with the pipeline network operation state through the dynamic game between the normal working condition generator and the real-time discriminator, and according to the comparison result between the real-time risk characteristics and the early warning thresholds, realize the risk perception early warning of the oil and gas pipeline network operation state.

[0331] It can be seen that according to the real-time operation data of the oil and gas pipeline network, the multi-dimensional characteristics of the pipeline network operation state are fused to generate a spatio-temporal feature matrix; the spatio-temporal feature matrix is input into the deep Gaussian process model to probabilistically model the abnormal characteristics in the pipeline network operation state, and a risk feature vector with confidence is output; according to the risk feature vector, a three-dimensional risk concentration distribution field is constructed to predict the propagation path and diffusion trend of the risk concentration distribution field; for the prediction result, an adaptive early warning threshold generation model based on the generative adversarial network is constructed to generate multi-level early warning thresholds that change with the pipeline network operation state, and based on the comparison result between the real-time risk characteristics and the early warning thresholds, the risk perception early warning of the oil and gas pipeline network operation state is realized, so as to achieve dynamic risk assessment and adaptive early warning, enhance the safety and emergency response ability of the pipeline network.

[0332] The structure, features and effects of the present invention have been described in detail based on the embodiments shown in the drawings. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the drawings. Any changes made according to the concept of the present invention, or equivalent embodiments modified to equivalent changes, still within the spirit covered by the specification and drawings, shall be within the protection scope of the present invention.

Claims

1. A method for early warning of oil and gas pipeline network operation status risk perception, characterized in that: include: According to the real-time operation data of the oil and gas pipeline network, the multi-dimensional features of the pipeline network operation status are integrated to generate a spatiotemporal feature matrix containing pressure, temperature, flow and vibration information; wherein, according to the real-time operation data collected by the multi-source sensors of the oil and gas pipeline network, timestamp alignment and missing value filling processing are performed, and a multi-source data alignment algorithm based on dynamic time warping is used to eliminate the difference in sensor sampling frequency and obtain a time-synchronized multi-source data sequence, wherein the real-time operation data includes pressure, temperature, flow and vibration data; The multi-source data sequence is input into the spatiotemporal convolutional neural network, wherein the spatial dimension uses a graph convolutional network to capture the topological structure characteristics of the pipeline network, and the time dimension uses a one-dimensional convolution kernel to extract the time series characteristics, so as to obtain a preliminary spatiotemporal feature representation; for the extracted spatiotemporal feature representation, a multi-head self-attention mechanism is used to calculate the correlation weights between different sensor data, and the features are weightedly fused according to the weights to highlight the contribution of key sensor data to risk perception, so as to obtain a weighted spatiotemporal feature representation; the weighted spatiotemporal features are input into a sparse autoencoder, and the dimension is reduced by nonlinear mapping, redundant information is removed, key features are retained, and a low-dimensional spatiotemporal feature matrix containing pressure, temperature, flow and vibration information is generated; The spatiotemporal feature matrix is ​​input into the deep Gaussian process model to perform probabilistic modeling on the abnormal features in the operation status of the pipeline network and output a risk feature vector with confidence. According to the risk characteristic vector, a three-dimensional risk concentration distribution field is constructed, and the propagation path and diffusion trend of the risk concentration distribution field are predicted based on the graph neural network; wherein, according to the risk characteristic vector, combined with the topological structure information of the oil and gas pipeline network, each risk characteristic point is mapped to the three-dimensional spatial coordinates of the pipeline network, and the Kriging interpolation algorithm is used to perform spatial interpolation on discrete risk characteristic points to generate a preliminary three-dimensional risk concentration distribution field; Based on the preliminary three-dimensional risk concentration distribution field, combined with the time dimension information, the spatiotemporal Kriging interpolation algorithm is used, combined with the historical risk propagation data, to dynamically correct the risk concentration distribution field, and obtain the three-dimensional risk concentration distribution field that changes with time; the three-dimensional risk concentration distribution field that changes with time is input into the graph neural network, and a graph model is constructed according to the topological structure of the pipeline network. The nodes of the graph model represent the key positions of the pipeline network, and the edges of the graph model represent the connection relationship of the pipeline sections. The graph attention mechanism is used to capture the risk propagation dependency between nodes, and the initial prediction result of risk propagation is obtained; According to the initial prediction results of risk propagation, a sequence prediction model based on a time convolutional network is used, combined with historical risk propagation path data, to predict the risk propagation path and diffusion trend of future time steps, and generate propagation prediction results of a dynamic risk concentration distribution field; According to the prediction results of the risk concentration distribution field, an adaptive warning threshold generation model based on the generative adversarial network is constructed. Through the dynamic game between the normal operating condition generator and the real-time discriminator, a multi-level warning threshold that changes with the operating status of the pipeline network is generated. Based on the comparison results of the real-time risk characteristics and the warning threshold, the risk-aware warning of the operating status of the oil and gas pipeline network is realized.

2. The method according to claim 1, characterized in that The spatiotemporal feature matrix is ​​input into the deep Gaussian process model, the abnormal features in the operation status of the pipeline network are probabilistically modeled, and the risk feature vector with confidence is output, including: Input the spatiotemporal feature matrix into a multi-layer perceptron, and map the original features to a high-dimensional latent space through multi-layer nonlinear transformation to obtain a deep feature representation, wherein each layer uses an activation function with a residual connection to avoid the gradient vanishing problem and enhance the feature expression capability; For the deep feature representation, a probability model based on Gaussian process regression is constructed, and a linear combination of radial basis kernel function and periodic kernel function is used to capture the nonlinear relationship and periodic change law in the operation state of the pipeline network, and obtain a preliminary probabilistic feature representation; According to the probabilistic feature representation, the predicted mean and variance of each feature point are calculated, and the uncertainty of the model is quantified through a variational inference algorithm to obtain a probabilistic feature distribution with a confidence interval; For the probability feature distribution, an anomaly detection algorithm based on Mahalanobis distance is used to calculate the distance between each feature point and the normal operating condition distribution, and the abnormal feature points are screened out according to a preset confidence threshold to generate a risk feature vector with confidence.

3. The method according to claim 2, characterized in that Based on the prediction results of the risk concentration distribution field, an adaptive warning threshold generation model based on a generative adversarial network is constructed. Through the dynamic game between the normal operating condition generator and the real-time discriminator, a multi-level warning threshold that changes with the operation status of the pipeline network is generated. Based on the comparison results of the real-time risk characteristics and the warning threshold, the risk perception warning of the operation status of the oil and gas pipeline network is realized, including: Based on historical normal operation data, a normal operating condition generator based on a generative adversarial network is trained to generate simulated data that conforms to the normal operation characteristics of the pipeline network. The generator uses a variational autoencoder with conditional constraints to ensure the diversity and authenticity of the generated data. Construct a real-time discriminator based on a deep residual network, input the prediction results of the risk concentration distribution field and the normal operating condition data generated by the generator, learn the boundary characteristics of normal and abnormal states through dynamic game, and output the discrimination results and their confidence; According to the discrimination results, an adaptive clustering algorithm is used to divide the discrimination results into multiple levels, and combined with the dynamic changes of the risk concentration distribution field, a multi-level warning threshold that changes with the operation status of the pipeline network is generated. The multi-level warning threshold includes low risk, medium risk and high risk thresholds; The real-time risk characteristics are compared with the multi-level warning thresholds, and a fuzzy logic-based decision model is used to dynamically adjust the warning level according to the confidence level of the risk characteristics and the degree of threshold deviation, thereby realizing risk-aware warning of the operating status of the oil and gas pipeline network.

4. An oil and gas pipeline network operation status risk perception and early warning system, characterized in that: The system comprises: A fusion module is used to fuse the multi-dimensional features of the operation status of the oil and gas pipeline network according to the real-time operation data of the pipeline network, and generate a spatiotemporal feature matrix containing pressure, temperature, flow and vibration information; wherein, according to the real-time operation data collected by the multi-source sensors of the oil and gas pipeline network, timestamp alignment and missing value filling processing are performed, and a multi-source data alignment algorithm based on dynamic time warping is adopted to eliminate the difference in sensor sampling frequency and obtain a time-synchronized multi-source data sequence, wherein the real-time operation data includes pressure, temperature, flow and vibration data; The multi-source data sequence is input into the spatiotemporal convolutional neural network, wherein the spatial dimension uses a graph convolutional network to capture the topological structure characteristics of the pipeline network, and the time dimension uses a one-dimensional convolution kernel to extract the time series characteristics, so as to obtain a preliminary spatiotemporal feature representation; for the extracted spatiotemporal feature representation, a multi-head self-attention mechanism is used to calculate the correlation weights between different sensor data, and the features are weightedly fused according to the weights to highlight the contribution of key sensor data to risk perception, so as to obtain a weighted spatiotemporal feature representation; the weighted spatiotemporal features are input into a sparse autoencoder, and the dimension is reduced by nonlinear mapping, redundant information is removed, key features are retained, and a low-dimensional spatiotemporal feature matrix containing pressure, temperature, flow and vibration information is generated; The modeling module is used to input the spatiotemporal feature matrix into the deep Gaussian process model, perform probabilistic modeling on the abnormal features in the operation status of the pipeline network, and output a risk feature vector with confidence; A construction module is used to construct a three-dimensional risk concentration distribution field according to the risk characteristic vector, and predict the propagation path and diffusion trend of the risk concentration distribution field based on a graph neural network; wherein, according to the risk characteristic vector, combined with the topological structure information of the oil and gas pipeline network, each risk characteristic point is mapped to the three-dimensional spatial coordinates of the pipeline network, and the Kriging interpolation algorithm is used to perform spatial interpolation on discrete risk characteristic points to generate a preliminary three-dimensional risk concentration distribution field; Based on the preliminary three-dimensional risk concentration distribution field, combined with the time dimension information, the spatiotemporal Kriging interpolation algorithm is used, combined with the historical risk propagation data, to dynamically correct the risk concentration distribution field, and obtain the three-dimensional risk concentration distribution field that changes with time; the three-dimensional risk concentration distribution field that changes with time is input into the graph neural network, and a graph model is constructed according to the topological structure of the pipeline network. The nodes of the graph model represent the key positions of the pipeline network, and the edges of the graph model represent the connection relationship of the pipeline sections. The graph attention mechanism is used to capture the risk propagation dependency between nodes, and the initial prediction result of risk propagation is obtained; According to the initial prediction results of risk propagation, a sequence prediction model based on a time convolutional network is used, combined with historical risk propagation path data, to predict the risk propagation path and diffusion trend of future time steps, and generate propagation prediction results of a dynamic risk concentration distribution field; The early warning module is used to build an adaptive early warning threshold generation model based on the generative adversarial network according to the prediction results of the risk concentration distribution field. Through the dynamic game between the normal operating condition generator and the real-time discriminator, a multi-level early warning threshold that changes with the operation status of the pipeline network is generated. Based on the comparison results of the real-time risk characteristics and the early warning threshold, the risk perception early warning of the operation status of the oil and gas pipeline network is realized.

5. The system according to claim 4, characterized in that The modeling module is specifically used for: Input the spatiotemporal feature matrix into a multi-layer perceptron, and map the original features to a high-dimensional latent space through multi-layer nonlinear transformation to obtain a deep feature representation, wherein each layer uses an activation function with a residual connection to avoid the gradient vanishing problem and enhance the feature expression capability; For the deep feature representation, a probability model based on Gaussian process regression is constructed, and a linear combination of radial basis kernel function and periodic kernel function is used to capture the nonlinear relationship and periodic change law in the operation state of the pipeline network, and obtain a preliminary probabilistic feature representation; According to the probabilistic feature representation, the predicted mean and variance of each feature point are calculated, and the uncertainty of the model is quantified through a variational inference algorithm to obtain a probabilistic feature distribution with a confidence interval; For the probability feature distribution, an anomaly detection algorithm based on Mahalanobis distance is used to calculate the distance between each feature point and the normal operating condition distribution, and the abnormal feature points are screened out according to a preset confidence threshold to generate a risk feature vector with confidence.

6. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 3 when executed.

7. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 3.

Images