A CO2 pipeline network status sensing method based on multi-source heterogeneous data fusion

Abstract

This invention belongs to the field of data processing technology and relates to a method for sensing the state of carbon dioxide pipeline networks based on multi-source heterogeneous data fusion. The method first acquires multi-source heterogeneous time-series data on the operation of the carbon dioxide pipeline network, calculates the causal relationship strength matrix between each data source based on a preset time window, and uses a spectral clustering algorithm to divide spatiotemporal causal clusters. Then, the Grey Wolf optimization algorithm is used to adaptively optimize the variational mode decomposition parameters, performing multi-scale decomposition of the data within each cluster, thereby constructing a weighted graph, and using a graph attention network to aggregate and generate cluster-level state features. Finally, the cluster-level state features are fused with the pipeline network's physical location encoding, and a Transformer encoder is used to capture long-distance global spatiotemporal dependencies and calculate the current state of the pipeline network. This invention achieves a layer-by-layer in-depth analysis from local causal relationships to global spatiotemporal dependencies, effectively improving the accuracy and real-time performance of carbon dioxide pipeline network state sensing.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method for sensing the status of carbon dioxide pipeline networks based on the fusion of multi-source heterogeneous data. Background Technology

[0002] During operation, carbon dioxide pipeline systems are susceptible to disturbances from various internal and external factors, and the accuracy and real-time performance of their condition monitoring directly affect the safety of pipeline operation and maintenance. Currently, mainstream pipeline condition monitoring methods fall into two main categories: one is simulation methods based on physical models. While the mechanism is clear, this method is computationally expensive and struggles to respond in real-time to dynamic changes in actual operating conditions, failing to meet the timeliness requirements of pipeline operation and maintenance. The other is threshold alarm methods based on single sensors. These rely solely on data from a single type of sensor, providing limited information dimensions and failing to comprehensively reflect the complex operating conditions of the pipeline network. This can easily lead to false alarms and missed alarms, making it unsuitable for the refined operation and maintenance requirements of modern pipeline networks.

[0003] To enhance the comprehensiveness of monitoring, various types of sensors, including pressure, temperature, flow, acoustic, and distributed optical fibers, were deployed along the carbon dioxide pipeline network, forming multi-source heterogeneous monitoring data. However, the physical meaning, sampling frequency, and data structure of different types of data vary significantly. Simple splicing or weighted fusion methods are insufficient to uncover the inherent synergistic relationships between the data and cannot fully leverage the complementary value of multi-source data.

[0004] As a spatially continuous system with a well-defined topology, pipeline networks exhibit close spatiotemporal correlations among their monitoring points. Existing time-series analysis models or feature extraction methods often overlook the global spatiotemporal dependencies formed by the physical connections and fluid dynamics propagation within the pipeline network, resulting in a lack of holistic perspective in assessing the network's condition. Furthermore, pipeline network runtime data generally exhibits non-stationary characteristics, making it difficult for traditional signal decomposition or feature extraction techniques to accurately capture key information under different operating conditions, further limiting the reliability of condition monitoring. Therefore, a novel pipeline network condition sensing method is urgently needed that can deeply integrate multi-source heterogeneous data and collaboratively mine local causal relationships and global spatiotemporal dependencies. Summary of the Invention

[0005] The purpose of this invention is to propose a carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion, in order to solve the technical problem of low reliability of existing technologies for monitoring the status of carbon dioxide pipeline networks.

[0006] To address the above problems, the technical solution of the carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion proposed in this invention is as follows:

[0007] A method for sensing the status of carbon dioxide pipeline networks based on multi-source heterogeneous data fusion includes the following steps:

[0008] The system acquires pressure, temperature, flow, acoustic, and fiber optic sensing data from the carbon dioxide pipeline network and constructs multi-source heterogeneous time-series data. Based on the multi-source heterogeneous time-series data within a preset time window, it calculates the causal relationship strength matrix between each data source and uses a spectral clustering algorithm to divide the multi-source heterogeneous time-series data into different spatiotemporal causal clusters.

[0009] For each spatiotemporal causal cluster, the number of modes and penalty factor of variational mode decomposition are determined based on the characteristics of the data within the cluster. Variational mode decomposition is then performed on each multi-source heterogeneous time series data within the cluster to obtain a series of intrinsic mode function components.

[0010] A weighted graph is constructed with data sources within the cluster as nodes and causal relationship strength as edges. The intrinsic mode function components are aggregated using a graph attention network to generate cluster-level state feature vectors representing the states of each cluster.

[0011] All cluster-level state feature vectors are fused with location codes representing the physical topology of the carbon dioxide pipeline network. The fused feature sequence is then input into a Transformer encoder network to extract global spatiotemporal dependency features, and the current state parameters of the carbon dioxide pipeline network are calculated through a fully connected layer.

[0012] Furthermore, the calculation of the causal relationship strength matrix between the various data sources includes:

[0013] For any two data sequences X and Y from heterogeneous multi-source time series data, the strength of the causal relationship... It is determined by calculating the following KL divergence:

[0014] ;in, Let X be the state of the data sequence at the next time step; and Representing time respectively At that time, the past state vectors of data sequences X and Y, and For the embedding dimension, Represents a conditional probability distribution;

[0015] Traverse all data source pairs to construct the causal relationship strength matrix.

[0016] Furthermore, the step of using a spectral clustering algorithm to divide the multi-source heterogeneous time-series data into different spatiotemporal causal clusters includes:

[0017] The causal relationship strength matrix is ​​used as the adjacency matrix and input into the spectral clustering algorithm. By calculating the eigenvectors of the Laplacian matrix and performing K-Means clustering on the eigenvectors in a low-dimensional space, closely related data sources are grouped into the same spatiotemporal causal cluster.

[0018] Furthermore, determining the number of modes and penalty factor for variational mode decomposition based on the characteristics of the intra-cluster data includes:

[0019] The search range for the number of modes R is set to [2, 9], and the search range for the penalty factor α is set to [500, 2500].

[0020] The Grey Wolf optimization algorithm is adopted, which uses the minimum envelope entropy of all intrinsic mode function components after decomposition as the fitness function. The optimal combination of mode number R and penalty factor α is obtained through iterative optimization.

[0021] Furthermore, the aggregation of the intrinsic mode function components using a graph attention network includes:

[0022] The R intrinsic mode function components obtained from the decomposition of each data source are concatenated into a one-dimensional vector, which serves as the initial feature representation of the corresponding node of the data source in the weighted graph.

[0023] By using the self-attention mechanism in the graph attention network, different attention weights are assigned to the neighboring nodes of each target node, and the features of the neighboring nodes are weighted and summed to update the feature representation of each target node.

[0024] The cluster-level state feature vector is generated by performing graph pooling on all node feature representations updated by the graph attention network.

[0025] Furthermore, the graph pooling operation is global average graph pooling.

[0026] Furthermore, the process of fusing all cluster-level state feature vectors with location codes representing the physical topology of the carbon dioxide pipeline network includes:

[0027] Calculate the average physical mileage markers of all data source nodes in each spatiotemporal causal cluster on the carbon dioxide pipeline network, and use this as the representative position of the spatiotemporal causal cluster; map the representative position to a position encoding vector with the same dimension as the cluster-level state feature vector using a sine-cosine function; and fuse the position encoding vector with the cluster-level state feature vector by element-wise addition.

[0028] Furthermore, the step of inputting the fused feature sequence into the Transformer encoder network to extract global spatiotemporal dependency features includes:

[0029] The fused feature sequence is input into a network consisting of six stacked encoders. Each encoder layer contains an eight-head self-attention module and a feedforward neural network module. Residual connections and layer normalization operations are used between the two modules to detect long-distance dependencies between different cluster-level state feature vectors in the fused feature sequence.

[0030] Furthermore, the calculation of the current state parameters of the carbon dioxide pipeline network through the fully connected layer includes:

[0031] The global spatiotemporal dependency features output by the Transformer encoder network are flattened and input into a fully connected network containing two layers of neurons. The pressure and temperature values ​​of preset key sections along the pipeline are output through a linear activation function as the current state parameters.

[0032] The beneficial effects of this invention are as follows: This invention breaks through the limitation of traditional methods that rely solely on physical distance to group monitoring points. By mining the causal relationships between multi-source heterogeneous data, it divides data sources that may be physically dispersed but have close internal logical connections into the same spatiotemporal causal cluster. This clustering method based on data mechanism effectively eliminates general false associations, suppresses interference from irrelevant information, and provides a data foundation with a higher signal-to-noise ratio for subsequent feature extraction.

[0033] To address the non-stationary and heterogeneous characteristics of pipeline network signals, this invention introduces adaptive variational mode decomposition and graph attention network within a spatiotemporal causal cluster, achieving multi-scale time-frequency feature separation of complex signals. Simultaneously, it utilizes graph structures to efficiently aggregate data features from different modes such as pressure, temperature, and optical fiber. The combination of these two approaches enables refined processing from signal-level decomposition to feature-level fusion, improving the accuracy of local region state representation.

[0034] This invention embeds the physical topology of the pipeline network into the Transformer architecture. Compared to traditional recurrent neural networks, the Transformer's self-attention mechanism can capture the global spatiotemporal dependencies between monitoring points across regions and over long distances. This design not only considers numerical correlations at the data level but also incorporates the actual physical connection structure of the pipeline network, enabling a layer-by-layer in-depth analysis from local states to a global view. This allows for a more accurate identification of the overall operational status and potential risks of the pipeline network, providing reliable support for safe operation and maintenance. Attached Figure Description

[0035] Figure 1 This is a flowchart of the steps of the carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion of the present invention. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0037] Specific embodiments of the carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion proposed in this invention are as follows:

[0038] like Figure 1 As shown, the carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion includes steps S1 to S4:

[0039] S1. Acquire pressure data, temperature data, flow data, acoustic data, and fiber optic sensing data of the carbon dioxide pipeline network and construct multi-source heterogeneous time-series data; based on the multi-source heterogeneous time-series data within a preset time window, calculate the causal relationship strength matrix between each data source, and use a spectral clustering algorithm to divide the multi-source heterogeneous time-series data into different spatiotemporal causal clusters.

[0040] Pressure sensors, temperature sensors, flow meters, acoustic leak detectors, and distributed fiber optic sensors were deployed at different locations along the carbon dioxide pipeline network. A SCADA system was used to collect the measurements from each sensor at a uniform time frequency, forming a multi-source heterogeneous time-series data set. To improve the accuracy of subsequent analysis, the collected raw data underwent preprocessing. The preprocessing included filling missing values ​​using linear interpolation, aligning data from different sampling frequencies to a unified time base using resampling technology, and normalizing all data using Z-score standardization to eliminate the influence of different physical dimensions.

[0041] A sliding time window is defined, for example, the past 60 minutes. For any two data series from multi-source heterogeneous time series data within the sliding time window, the Granger causality test is used. A vector regression model is constructed to determine whether one data series can predict the change of another, and the F-statistic of the test result is used as the strength of the causal relationship between the two data sources. This test is performed pairwise on all data sources in the carbon dioxide pipeline network, constructing a symmetric causal relationship strength matrix. This causal relationship strength matrix is ​​used as the similarity matrix input for a spectral clustering algorithm. By calculating the eigenvectors of the Laplacian matrix, the original data points are mapped to a low-dimensional feature space. Then, the K-Means method is used to cluster in this space, thereby grouping sensors that are physically close and whose data changes have a strong causal correlation into the same spatiotemporal causal cluster.

[0042] In an optional embodiment, calculating the causal relationship strength matrix between the data sources includes:

[0043] For any two data sequences X and Y from heterogeneous multi-source time series data, the strength of the causal relationship... It is determined by calculating the following KL divergence:

[0044] ;in, Let X be the state of the data sequence at the next time step; and Representing time respectively At that time, the past state vectors of data sequences X and Y, and For the embedding dimension, Represents a conditional probability distribution;

[0045] Traverse all data source pairs to construct a causal relationship strength matrix.

[0046] Specifically, time-series data from two sources are selected from the carbon dioxide pipeline network as data sequence X and data sequence Y. For example, X is the pressure sequence of node A, and Y is the temperature sequence of node B. First, a past state vector is constructed using phase space reconstruction technology. The embedding dimension is then determined. Taking =3 as an example, at the time point Past state vector Composed of observations from the current and past moments, denoted as Subsequently, in order to quantify the information transmission from Y to X, it is necessary to estimate two key conditional probability distributions: first, the probability distribution of X's next state considering only X's own historical information. Secondly, considering the historical information of both X and Y, what is the probability distribution of the state of X at the next moment? The above probability distribution can be estimated using nonparametric estimation methods, such as the K-nearest neighbor algorithm.

[0047] After obtaining estimates of the two conditional probability distributions, the strength of the causal relationship from data sequence Y to data sequence X is determined by calculating the KL divergence between them. The calculated KL divergence value, or transitive entropy, quantifies the information gain brought by introducing additional historical information of Y to predict the future state of X, based on the known historical information of X itself. The larger the value, the stronger the causal driving force or predictive contribution of Y to the evolutionary trend of X. To capture the dynamic time-varying characteristics of the carbon dioxide pipeline network's operating state, a sliding time window is set, and the above calculation process is repeated within each window to detect changes in causal relationships. By traversing all N data sources in the carbon dioxide pipeline network and performing pairwise calculations on any two data sources, an N×N dimensional causal relationship strength matrix is ​​finally constructed. The element value in the i-th row and j-th column of this matrix represents the causal influence strength of data source j on data source i.

[0048] In an optional embodiment, the step of using a spectral clustering algorithm to divide the multi-source heterogeneous time-series data into different spatiotemporal causal clusters includes:

[0049] The causal relationship strength matrix is ​​used as the adjacency matrix and input into the spectral clustering algorithm. By calculating the eigenvectors of the Laplacian matrix and performing K-Means clustering on the eigenvectors in a low-dimensional space, closely related data sources are grouped into the same spatiotemporal causal cluster.

[0050] The generated N×N causal relationship strength matrix is ​​mapped to an adjacency matrix A of the graph. In this graph structure, each data source is considered a node, and the weight of the edge between node a and node b is directly determined by the corresponding element values ​​in the matrix, thus quantifying the strength of the causal association between nodes. Based on the adjacency matrix A, the degree matrix D is calculated. The degree matrix D is a diagonal matrix, and its main diagonal elements... It is the sum of all elements in the i-th row of the adjacency matrix A. To achieve better clustering convergence, the Laplacian matrix L of the graph is calculated, preferably using the symmetric normalized Laplacian matrix form, whose formula is: ,in Represents the identity matrix.

[0051] After calculating the Laplacian matrix L, eigenvalue decomposition is performed on this matrix to select the eigenvectors corresponding to the m smallest eigenvalues. These m eigenvectors constitute an N×m dimensional feature matrix, where each row of the feature matrix corresponds to an original data source node, representing the embedding representation of that node in the mapped m-dimensional low-dimensional space. This dimensionality reduction operation can effectively reveal the potential topological clustering structure between data points. Using these N m-dimensional feature vectors as input data, the K-Means clustering algorithm is applied for clustering. The K-Means algorithm assigns data points to a predetermined number of clusters, maximizing the similarity of vectors within the same cluster while minimizing the similarity between different clusters, thereby grouping data sources with close causal relationships into the same spatiotemporal causal cluster.

[0052] S2, for each spatiotemporal causal cluster, determine the number of modes and penalty factor of variational mode decomposition based on the characteristics of the data within the cluster, and perform variational mode decomposition on each multi-source heterogeneous time series data within the cluster to obtain a series of intrinsic mode function components.

[0053] For each spatiotemporal causal cluster, the time series with the largest variance within the cluster is selected as the representative signal of that cluster. A genetic algorithm is used to adaptively optimize the two key parameters of variational mode decomposition: the number of modes R and the penalty factor α. Using the minimum envelope entropy as the fitness function of the genetic algorithm, an optimal combination of R and α values ​​is found through iterative calculation, resulting in the sparsest and most concentrated features of each decomposed modal component. This optimal set of parameters is then uniformly applied to all time series data within the cluster, and variational mode decomposition is performed on each series, thereby obtaining R intrinsic mode function components arranged from high frequency to low frequency.

[0054] To find the optimal variational mode decomposition parameters for the data of each spatiotemporal causal cluster, in another optional embodiment, determining the number of modes and the penalty factor for variational mode decomposition based on the characteristics of the data within the cluster includes:

[0055] The search range for the number of modes R is set to [2, 9], and the search range for the penalty factor is set to [500, 2500].

[0056] The Grey Wolf optimization algorithm is adopted, which uses the minimum envelope entropy of all intrinsic mode function components after decomposition as the fitness function. The optimal combination of mode number R and penalty factor α is obtained through iterative optimization.

[0057] Specifically, for each defined spatiotemporal causal cluster, the arithmetic mean of the time series data from all data sources within that cluster is calculated and used as the representative signal for that cluster. The population for the gray wolf optimization algorithm is initialized, with the spatial location of each individual gray wolf directly mapped to a set of candidate parameter combinations, namely the number of modes R and the penalty factor α. These parameters are randomly generated within a preset search range; for example, the spatial location of an individual might correspond to a parameter combination where R=5 and α=1500.

[0058] For each individual gray wolf representing a parameter combination, variational mode decomposition is performed on the representative signal of the cluster using this set of parameters, yielding R intrinsic mode function (EMF) components. The envelope entropy of these R EMF components is calculated. Since envelope entropy effectively characterizes signal sparsity, a smaller entropy value indicates stronger signal periodicity and better mode separation. Therefore, the minimum value among the R envelope entropies is selected as the fitness value of the individual gray wolf. The gray wolf optimization algorithm simulates the social hierarchy and group predation behavior of the wolf pack, continuously updating the positions of all gray wolves in the population through multiple iterations, i.e., dynamically adjusting the combination of R and α, searching towards a better fitness value, i.e., a smaller minimum envelope entropy. The algorithm stops when the preset maximum number of iterations is reached or the fitness value converges. The parameter combination corresponding to the position of the alpha wolf at this point is considered the optimal number of modes and penalty factor for the data cluster.

[0059] S3. Construct a weighted graph with data sources within the cluster as nodes and causal relationship strength as edges, and use a graph attention network to aggregate the intrinsic mode function components to generate cluster-level state feature vectors representing the states of each cluster.

[0060] Within each constructed spatiotemporal causal cluster, each data source within the cluster is treated as a graph node, and the edge weights between nodes are defined using the corresponding values ​​in the previously calculated causal relationship strength matrix, thus constructing a weighted graph containing rich topological information. For each data source sequence within the cluster, variational mode decomposition is performed using the optimal parameters R and α determined in step S2, yielding R intrinsic mode function components. The values ​​of these R intrinsic mode function components at the current time are concatenated into a one-dimensional vector, serving as the initial feature representation of the corresponding node in the weighted graph. Graph attention networks are used to process the graph data. The sensor data source processing includes: calculating the attention weights between the target node and its neighboring nodes in the graph structure through a self-attention mechanism. These attention weights accurately characterize the importance of different neighboring nodes to the state update of the target node; secondly, the features of all neighboring nodes are weighted and summed based on the attention weights, thereby updating the feature representation of the target node. After deep feature extraction by a multi-layer graph attention network, graph pooling is performed on the feature vectors of all nodes in the cluster. For example, the average or maximum value of each element of the feature vector of all nodes is taken, and the scattered node-level features are aggregated into a high-dimensional and compact cluster-level state feature vector.

[0061] S4. All cluster-level state feature vectors are fused with the location codes representing the physical topology of the carbon dioxide pipeline network. The fused feature sequence is then input into the Transformer encoder network to extract global spatiotemporal dependency features, and the current state parameters of the carbon dioxide pipeline network are calculated through a fully connected layer.

[0062] In this step, a corresponding location code is generated for each spatiotemporal causal cluster. Specifically, the arithmetic mean of the physical mileage markers of all sensors within the spatiotemporal causal cluster on the pipeline is calculated, and this value is used as the representative physical location of the spatiotemporal causal cluster, mapped to a location code vector. This location code vector is then element-wise added to the cluster-level state feature vector of the corresponding cluster, forming a fused feature sequence that contains both local state features and global spatial location information. This fused feature sequence is input into a Transformer encoder. The multi-head self-attention mechanism within the Transformer encoder can calculate the correlation weights between any two cluster-level state feature vectors, effectively capturing and detecting long-distance, cross-regional state dependencies in the pipeline network. The sequence features output by the Transformer encoder are passed through a global average pooling layer and then input into a multilayer perceptron network with two hidden layers. The output layer of the network can be flexibly configured according to task requirements. For example, it can be set to a single neuron containing a sigmoid activation function to output the probability value of a pipeline leak; or it can be set to a classification layer containing a softmax activation function to accurately determine whether the pipeline is currently in a normal, blocked, or leaking operating state.

[0063] In an optional embodiment, fusing all cluster-level state feature vectors with location codes representing the physical topology of the carbon dioxide pipeline network includes:

[0064] Calculate the average physical mileage markers of all data source nodes in each spatiotemporal causal cluster on the carbon dioxide pipeline network, and use this as the representative position of the spatiotemporal causal cluster; map the representative position to a position encoding vector with the same dimension as the cluster-level state feature vector using a sine-cosine function; and fuse the position encoding vector with the cluster-level state feature vector by element-wise addition.

[0065] Specifically, the actual physical installation locations of all data source sensors within the cluster on the carbon dioxide pipeline network are collected, typically represented by kilometer markers, and the arithmetic mean of these kilometer markers is calculated. For example, if a cluster contains three sensors located at 10.2 km, 10.5 km, and 11.2 km, the representative location of the cluster is 10.63 km. After obtaining the representative location scalar value, this scalar value needs to be converted into a high-dimensional vector, i.e., a location encoding vector, for fusion with the cluster-level state feature vector. This conversion is accomplished using a set of sine and cosine functions of different frequencies, assuming the dimension of the cluster-level state feature vector is... Representative positions are Then the position encoding vector of the first Each element is calculated as , No. Each element is calculated as The generated vector has the same dimension as the cluster-level state feature vector and can uniquely represent the physical location of the cluster. The location encoding vector is then fused with the generated cluster-level state feature vector through element-wise addition, so that the feature vector simultaneously contains both the cluster's behavioral information and static spatial location information.

[0066] In an optional embodiment, the step of inputting the fused feature sequence into the Transformer encoder network to extract global spatiotemporal dependency features includes:

[0067] The fused feature sequence is input into a network consisting of six stacked encoders. Each encoder layer contains an eight-head self-attention module and a feedforward neural network module. Residual connections and layer normalization operations are used between the two modules to detect long-distance dependencies between different cluster-level state feature vectors in the fused feature sequence.

[0068] Specifically, an input feature sequence of length C is constructed, where each element corresponds to a cluster-level state feature vector fused with positional encoding, as previously generated. It is assumed that the sequence is divided into C clusters. This sequence is then fed into a deep network consisting of six identical encoder layers stacked in series. Within each encoder layer, the data passes through an 8-head self-attention module. This module allows the cluster-level state feature vector fused with positional encoding from each cluster in the sequence to interact with the cluster-level state feature vectors fused with positional encoding from all other clusters in the sequence. By calculating the correlation weights between them, information is aggregated from a global perspective to update its own feature representation. The 8-head mechanism refers to the fact that the above attention calculation process is executed in parallel eight times, with each head learning a different dependency pattern, thereby enhancing the model's ability to perceive and detect complex spatiotemporal features.

[0069] Following the self-attention module, data flows through a feedforward neural network module, which consists of two linear transformation layers and an activation function. This module aims to transform the feature representation of each cluster to enhance the model's feature representation capabilities. To ensure the stability and efficiency of deep network training, residual connections and layer normalization are applied after each module—the self-attention module and the feedforward neural network module. Residual connections directly superimpose the module's input onto the output, effectively preventing the vanishing gradient problem in deep networks; layer normalization standardizes the feature distribution of each sample, accelerating the model's convergence process. Through the layer-by-layer stacking of these six encoder layers, the model can gradually build a deep understanding of the long-distance, global spatiotemporal dependencies between all clusters in the entire network, outputting a set of highly abstract global spatiotemporal dependency features.

[0070] In an optional embodiment, calculating the current state parameters of the carbon dioxide pipeline network through the fully connected layer includes:

[0071] The global spatiotemporal dependency features output by the Transformer encoder network are flattened and input into a fully connected network containing two layers of neurons. The pressure and temperature values ​​of preset key sections along the pipeline are output through a linear activation function as the current state parameters.

[0072] The system receives the output feature sequence from the Transformer encoder network, where each element is a deep feature representation of the corresponding spatiotemporal causal cluster. For example, if there are Q clusters and each feature vector has a dimension of M, the output is a Q×M matrix. The matrix is ​​then flattened by concatenating all its elements sequentially into a single Q×M one-dimensional vector, which encapsulates the global spatiotemporal dependency information extracted from all clusters.

[0073] The flattened feature vector is input into a fully connected network consisting of two layers of neurons, also known as a multilayer perceptron. The first layer of this network is a hidden layer that performs linear transformations and activations on the input high-dimensional features, refining and integrating information. The second layer is the output layer, with the number of neurons set to the total number of state parameters to be predicted. For example, if the prediction target covers the pressure and temperature values ​​at 15 key sections along the pipeline network, the output layer will be configured with 30 neurons. This output layer uses a linear activation function because it allows the output value to be any real number, matching the task of predicting continuous physical quantities such as pressure and temperature. Ultimately, the 30 values ​​output by the fully connected network correspond to the current pressure and temperature prediction values ​​at the 15 key sections, thereby achieving an accurate calculation of the operating status of the carbon dioxide pipeline network.

[0074] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for sensing the status of carbon dioxide pipeline networks based on multi-source heterogeneous data fusion, characterized in that, Includes the following steps: Acquire pressure, temperature, flow, acoustic, and fiber optic sensing data of the carbon dioxide pipeline network and construct multi-source heterogeneous time-series data; Based on multi-source heterogeneous time-series data within a preset time window, the causal relationship strength matrix between each data source is calculated, including: For any two data sequences X and Y from heterogeneous multi-source time series data, the strength of the causal relationship... It is determined by calculating the following KL divergence: ; Let X be the state of the data sequence at the next time step; and Representing time respectively At that time, the past state vectors of data sequences X and Y, and For the embedding dimension, Represents a conditional probability distribution; Traverse all data source pairs to construct the causal relationship strength matrix; The multi-source heterogeneous time-series data are then divided into different spatiotemporal causal clusters using a spectral clustering algorithm, including: The causal relationship strength matrix is ​​used as the adjacency matrix and input into the spectral clustering algorithm. By calculating the eigenvectors of the Laplacian matrix and performing K-Means clustering on the eigenvectors in a low-dimensional space, closely related data sources are divided into the same spatiotemporal causal cluster. For each spatiotemporal causal cluster, the number of modes and penalty factor of variational mode decomposition are determined based on the characteristics of the data within the cluster. Variational mode decomposition is then performed on each multi-source heterogeneous time series data within the cluster to obtain a series of intrinsic mode function components. A weighted graph is constructed with data sources within the cluster as nodes and causal relationship strength as edges. The intrinsic mode function components are aggregated using a graph attention network to generate cluster-level state feature vectors representing the states of each cluster. All cluster-level state feature vectors are fused with location codes representing the physical topology of the carbon dioxide pipeline network. The fused feature sequence is then input into a Transformer encoder network to extract global spatiotemporal dependency features, and the current state parameters of the carbon dioxide pipeline network are calculated through a fully connected layer.

2. The carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The step of determining the number of modes and penalty factor for variational mode decomposition based on the characteristics of intra-cluster data includes: The search range for the number of modes R is set to [2, 9], and the search range for the penalty factor α is set to [500, 2500]. The Grey Wolf optimization algorithm is adopted, which uses the minimum envelope entropy of all intrinsic mode function components after decomposition as the fitness function. The optimal combination of mode number R and penalty factor α is obtained through iterative optimization.

3. The carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The aggregation of the intrinsic mode function components using a graph attention network includes: The R intrinsic mode function components obtained from the decomposition of each data source are concatenated into a one-dimensional vector, which serves as the initial feature representation of the corresponding node of the data source in the weighted graph. By using the self-attention mechanism in the graph attention network, different attention weights are assigned to the neighboring nodes of each target node, and the features of the neighboring nodes are weighted and summed to update the feature representation of each target node. The cluster-level state feature vector is generated by performing graph pooling on all node feature representations updated by the graph attention network.

4. The carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The graph pooling operation is global average graph pooling.

5. The carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The process of fusing all cluster-level state feature vectors with location codes representing the physical topology of the carbon dioxide pipeline network includes: Calculate the average physical mileage markers of all data source nodes in each spatiotemporal causal cluster on the carbon dioxide pipeline network, and use this as the representative position of the spatiotemporal causal cluster; map the representative position to a position encoding vector with the same dimension as the cluster-level state feature vector using a sine-cosine function; and fuse the position encoding vector and the cluster-level state feature vector by adding them element by element.

6. The carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The step of inputting the fused feature sequence into the Transformer encoder network to extract global spatiotemporal dependency features includes: The fused feature sequence is input into a network consisting of six stacked encoders. Each encoder layer contains an eight-head self-attention module and a feedforward neural network module. Residual connections and layer normalization operations are used between the two modules to detect long-distance dependencies between different cluster-level state feature vectors in the fused feature sequence.

7. The carbon dioxide pipeline network status sensing method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The calculation of the current state parameters of the carbon dioxide pipeline network through the fully connected layer includes: The global spatiotemporal dependency features output by the Transformer encoder network are flattened and input into a fully connected network containing two layers of neurons. The pressure and temperature values ​​of preset key sections along the pipeline are output through a linear activation function as the current state parameters.

Images