Carbon dioxide pipe network state sensing method based on multi-source heterogeneous data fusion

Abstract

The invention belongs to the technical field of data processing, and relates to a carbon dioxide pipe network state sensing method based on multi-source heterogeneous data fusion. The method comprises the following steps: firstly, acquiring multi-source heterogeneous time sequence data of operation of the carbon dioxide pipe network, calculating a causal relationship intensity matrix among data sources based on a preset time window, and dividing space-time causal clusters by utilizing a spectral clustering algorithm; then a grey wolf optimization algorithm is adopted to adaptively optimize variational mode decomposition parameters, multi-scale decomposition is carried out on data in clusters, then a weighted graph is constructed, and cluster-level state features are generated through graph attention network aggregation; and finally, fusing the cluster-level state characteristics with the physical position codes of the pipe network, capturing long-distance global space-time dependence through a Transform encoder, and resolving the current state of the pipe network. According to the method, layer-by-layer deep analysis from local causal association to global space-time dependence is realized, and the accuracy and real-time performance of carbon dioxide pipe network state sensing are effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to a carbon dioxide pipe network state perception method based on multi-source heterogeneous data fusion. BACKGROUND

[0002] During the operation of the carbon dioxide pipe network system, it is easily disturbed by various internal and external factors, and the accuracy and real-time performance of the state monitoring are directly related to the safety of the pipe network operation and maintenance. The current mainstream pipe network state monitoring methods mainly include two types: one is a simulation method based on a physical model, which has a clear mechanism, but has high calculation cost and is difficult to respond to the dynamic changes of the real working conditions in real time, and cannot meet the timeliness requirements of the pipe network operation and maintenance. The other is a threshold alarm method based on a single sensor, which only relies on single type sensor data for monitoring, has single information dimension, and is difficult to fully reflect the complex operation state of the pipe network, and is prone to false alarm and missed alarm problems, and cannot adapt to the fine operation and maintenance requirements of the modern pipe network.

[0003] In order to improve the monitoring comprehensiveness, various types of sensors such as pressure, temperature, flow, sound wave and distributed optical fiber are arranged along the carbon dioxide pipe network to form multi-source heterogeneous monitoring data. However, the physical meaning, sampling frequency and data structure of different types of data are significantly different, and simple splicing or weighted fusion method cannot excavate the internal cooperative relationship between the data, and cannot fully play the complementary value of multi-source data.

[0004] The pipe network is a spatial continuous system with a clear topological structure, and there is a close spatio-temporal correlation between the monitoring points. The existing time series analysis model or feature extraction method often ignores the global spatio-temporal dependence based on the physical connection and fluid dynamics propagation of the pipe network, resulting in a lack of overall perspective for the judgment of the pipe network state. In addition, the pipe network operation time series data generally shows non-stationary characteristics, and the traditional signal decomposition or feature extraction technology cannot accurately capture the key information under different working conditions, which further restricts the reliability of the state monitoring. Therefore, a new pipe network state perception method is needed, which can deeply fuse multi-source heterogeneous data, cooperatively excavate local causal correlation and global spatio-temporal dependence. SUMMARY

[0005] The purpose of the present application is to provide a carbon dioxide pipe network state perception method based on multi-source heterogeneous data fusion, which solves the technical problem of low reliability of the existing state monitoring of the carbon dioxide pipe network.

[0006] To solve the above problems, the technical scheme of the carbon dioxide pipe network state perception method based on multi-source heterogeneous data fusion provided by the present application is as follows: The carbon dioxide pipe network state perception method based on multi-source heterogeneous data fusion comprises the following steps: acquire pressure data, temperature data, flow data, acoustic wave data and optical fiber sensing data of a 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 a causal relationship strength matrix between each data source, and divide the multi-source heterogeneous time series data into different spatio-temporal causal clusters by using a spectral clustering algorithm; For each spatio-temporal causal cluster, determine the number of modalities and the penalty factor of variational modality decomposition according to the characteristics of the data in the cluster, and perform variational modality decomposition on each multi-source heterogeneous time series data in the cluster to obtain a series of intrinsic modal function components; Construct a weighted graph with the data sources in the cluster as nodes and the causal relationship strength as edges, and aggregate the intrinsic modal function components by using a graph attention network to generate a cluster-level state feature vector representing the state of each cluster; Fuse all cluster-level state feature vectors with location encodings representing the physical topology of the carbon dioxide pipeline network, input the fused feature sequence into a Transformer encoder network to extract global spatio-temporal dependency features, and solve the current state parameters of the carbon dioxide pipeline network through a fully connected layer.

[0007] Further, the calculation of the causal relationship strength matrix between each data source includes: For any two data sequences X and Y corresponding to the data sources in the multi-source heterogeneous time series data, the causal relationship strength Determine by calculating the KL divergence as follows: ; Wherein, is the state of data sequence X at the next time; and respectively represent the past state vectors of data sequences X and Y at time , and and are embedding dimensions, represents the conditional probability distribution; Iterate through all data source pairs to form the causal relationship strength matrix.

[0008] Further, the division of the multi-source heterogeneous time series data into different spatio-temporal causal clusters by using a spectral clustering algorithm includes: Input the causal relationship strength matrix as an adjacency matrix into the spectral clustering algorithm, calculate the eigenvectors of the Laplacian matrix, and perform K-Means clustering in a low-dimensional space to divide the data sources with close relationships into the same spatio-temporal causal cluster.

[0009] Further, the determination of the number of modalities and the penalty factor of variational modality decomposition according to the characteristics of the data in the cluster 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.

[0010] Furthermore, 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.

[0011] Furthermore, the graph pooling operation is global average graph pooling.

[0012] 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: 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; use a sine-cosine function to map the representative position into a position encoding vector with the same dimension as the cluster-level state feature vector; and fuse the position encoding vector and the cluster-level state feature vector by adding them element by element.

[0013] Furthermore, 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.

[0014] Furthermore, the calculation of the current state parameters of the carbon dioxide pipeline network through the fully connected layer includes: The global space-time dependent features output by the Transformer encoder network are flattened and input into a fully connected network containing two layers of neurons, and the pressure values and temperature values of the preset key sections along the pipeline network are output by a linear activation function as the current state parameters.

[0015] The beneficial effects of the present application are: the present application breaks the limitation of traditional methods that only rely on physical distance for monitoring point grouping, and by mining the causal relationship between multiple source heterogeneous data, data sources with possible dispersion in physical location but close internal logical connection are divided into the same space-time causal cluster, this data mechanism-based clustering method effectively eliminates general false associations and suppresses irrelevant information interference, providing a higher signal-to-noise ratio data basis for subsequent feature extraction.

[0016] In view of the non-stationary and multi-source heterogeneous characteristics of the pipeline network signal, the present application introduces adaptive variational mode decomposition and graph attention network in the space-time causal cluster, realizes multi-scale time-frequency feature separation of complex signals, and simultaneously efficiently aggregates data features of different modalities such as pressure, temperature and optical fiber by using graph structure, and the combination of the two realizes fine processing from signal level decomposition to feature level fusion, improving the accuracy of local area state representation.

[0017] The present application embeds the physical topology of the pipeline network into the Transformer architecture, and compared with the traditional recurrent neural network, the self-attention mechanism of the Transformer can capture the global space-time dependency relationship between monitoring points across regions and long distances. This design not only considers the numerical correlation at the data level, but also combines the actual physical connection structure of the pipeline network, realizes layer-by-layer in-depth analysis from local state to global view, and thus can more accurately identify the overall operation state and potential risks of the pipeline network, providing reliable support for safe operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The step flowchart of the carbon dioxide pipeline network state perception method based on multi-source heterogeneous data fusion of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0020] The specific embodiments of the carbon dioxide pipeline network state perception method based on multi-source heterogeneous data fusion proposed by the present application are as follows: As shown in Figure 1 The carbon dioxide pipeline network state perception method based on multi-source heterogeneous data fusion includes steps S1 to S4: S1, acquire pressure data, temperature data, flow data, sound wave data and optical fiber sensing data of carbon dioxide pipeline network operation and construct multi-source heterogeneous time series data; based on the multi-source heterogeneous time series data in a preset time window, calculate the causal relationship strength matrix between each data source, and divide the multi-source heterogeneous time series data into different spatio-temporal causal clusters by using a spectral clustering algorithm.

[0021] Pressure sensors, temperature sensors, flow meters, sound wave leak detectors and distributed optical fiber sensors are arranged at different positions along the carbon dioxide pipeline network, and the measurement values of each sensor, i.e. each data source, are collected by the SCADA system at a unified time frequency to form multi-source heterogeneous time series data. In order to improve the accuracy of subsequent analysis, the collected raw data is preprocessed. The preprocessing process includes filling missing values using linear interpolation method, aligning data of different sampling frequencies to a unified time reference through resampling technology, and normalizing all data by using Z-score standardization method to eliminate the influence of different physical dimensions.

[0022] A sliding time window is set, for example, the past 60 minutes. For any two data sources in the multi-source heterogeneous time series data in the sliding time window, the Granger causality test method is used to judge whether one data sequence can predict the change of another data sequence by constructing a vector regression model, and the F statistic of the test result is taken as the causal relationship strength between the two data sources. The test is performed on all data sources in the carbon dioxide pipeline network in pairs to construct a symmetric causal relationship strength matrix. The causal relationship strength matrix is input as the similarity matrix of the spectral clustering algorithm, the original data points are mapped to a low-dimensional feature space by calculating the eigenvectors of the Laplacian matrix, and then the K-Means method is used for clustering in the space, so that sensors with strong causal relationship between physical location and data change are divided into the same spatio-temporal causal cluster.

[0023] In an optional embodiment, the calculation of the causal relationship strength matrix between each data source includes: For any two data sources corresponding to data sequences X and Y in the multi-source heterogeneous time series data, the causal relationship strength is determined by calculating the KL divergence as follows: ; wherein, is the state of data sequence X at the next time; and respectively represent the past state vectors of data sequences X and Y at time , and are embedding dimensions, represents the conditional probability distribution; Traverse all data source pairs to form a causality strength matrix.

[0024] Specifically, time series data of two data 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 by using a phase space reconstruction technique. Taking an embedding dimension m = 3 as an example, a past state vector X t-3 of a time point t-3 is composed of observation values at the current time point and the past two time points, and is expressed as X t-3 = [X t-3, X t-4, X t-5 ]. Subsequently, in order to quantify the information transmission of Y to X, two key conditional probability distributions need to be estimated: one is the probability distribution of the next time state of X only considering the historical information of X itself P (X t+1 | X t, X t-1, X t-2 ). The other is the probability distribution of the next time state of X considering the historical information of both X and Y P (X t+1 | X t, X t-1, X t-2, Y t, Y t-1, Y t-2 ). The above probability distributions can be estimated by a non-parametric estimation method, such as a K-nearest neighbor algorithm.

[0025] After obtaining the estimates of the two conditional probability distributions, the causality strength from data sequence Y to data sequence X is determined by calculating the KL divergence between them. The calculated KL divergence value, which is the transfer entropy, quantifies the information gain brought by the introduction of the historical information of Y to predict the future state of X on the basis of the historical information of X itself; the greater the value, the stronger the causal driving force or prediction contribution of Y to the evolution trend of X. In order to capture the dynamic time-varying characteristics of the operation state of the carbon dioxide pipeline network, by setting a sliding time window and repeatedly executing the above calculation process in each window, the change of the causality can be detected. Traverse all N data sources in the carbon dioxide pipeline network, and perform pairwise calculation on any two data sources, and finally form an N × N dimensional causality strength matrix. The element value in the i-th row and j-th column of the matrix represents the causality influence strength of data source j on data source i.

[0026] In an optional embodiment, the dividing the multi-source heterogeneous time series data into different spatiotemporal causal clusters by using a spectral clustering algorithm comprises: inputting the causality strength matrix as an adjacency matrix into the spectral clustering algorithm, calculating the eigenvectors of the Laplacian matrix, and performing K-Means clustering on the eigenvectors in a low-dimensional space, so as to divide the data sources with close relationships into the same spatiotemporal causal cluster.

[0027] ​​​The N*N dimensional causal relationship strength matrix generated above is mapped to an adjacency matrix A of a graph. In this graph structure, each data source is regarded as a node in the graph, and the weight of the edge between node a and node b is directly determined by the corresponding element value in the matrix, thereby quantifying the causal correlation strength between nodes. Based on the adjacency matrix A, a degree matrix D is calculated, which is a diagonal matrix whose elements on the main diagonal are the cumulative sum of all elements of the ith row of the adjacency matrix A. A Laplacian matrix L of the graph is calculated for better clustering convergence effect, and a symmetric normalized Laplacian matrix is preferably used, and the calculation formula is where represents the unit matrix.

[0028] After the Laplacian matrix L is calculated, the matrix is subjected to eigenvalue decomposition, and the eigenvectors corresponding to the first m smallest eigenvalues are screened out. The m eigenvectors form an N*m dimensional feature matrix, wherein each row vector of the feature matrix corresponds to an original data source node, representing the embedding representation of the node in the mapped m dimensional low dimensional space. This dimensionality reduction operation can effectively reveal the potential topological clustering structure between data points. The N m dimensional eigenvectors are taken as input data, and a K-Means clustering algorithm is applied for clustering division. The K-Means algorithm assigns data points to a predetermined number of clusters, maximizes the similarity of vectors within the same cluster, and minimizes the similarity between different clusters, thereby realizing the division of data sources with close causal relationship into the same spatiotemporal causal cluster.

[0029] S2, for each spatiotemporal causal cluster, the number of modes and the penalty factor of the variational mode decomposition are determined according to the characteristics of the data in the cluster, and the variational mode decomposition is performed on each multi-source heterogeneous time series data in the cluster to obtain a series of intrinsic mode function components.

[0030] For each spatiotemporal causal cluster divided, a time series with the largest variance in the cluster is selected as the representative signal of the cluster. Genetic algorithm is used to adaptively optimize two key parameters of the variational mode decomposition, i.e. the number of modes R and the penalty factor a. The minimum envelope entropy is taken as the fitness function of the genetic algorithm, and through iterative calculation, an optimal R value and a value combination of a value are found, which make the characteristics of the decomposed mode components most sparse and the energy most concentrated. The optimal parameter combination is uniformly applied to all time series data contained in the cluster, and the variational mode decomposition is performed on each sequence to obtain R intrinsic mode function components arranged from high frequency to low frequency.

[0031] In order to find the optimal variational mode decomposition parameters for the data of each spatiotemporal causal cluster, in another optional embodiment, the determination of the number of modes and the penalty factor of the variational mode decomposition according to the characteristics of the data in the cluster comprises: The search range of the modal number R is set as [2, 9], and the search range of the penalty factor is [500, 2500]; The grey wolf optimization algorithm is adopted, the minimum envelope entropy of all the intrinsic mode function components after decomposition is taken as the fitness function, and the optimal modal number R and penalty factor α combination is obtained through iterative optimization.

[0032] Specifically, for each determined spatio-temporal causal cluster, the arithmetic mean of all data source time series in the cluster is calculated as the representative signal of the cluster. The population of the grey wolf optimization algorithm is initialized, and the spatial position of each grey wolf individual is directly mapped to a set of candidate parameter combinations, i.e., the modal number R and the penalty factor α, which are randomly generated within the preset search range, for example, the spatial position of a certain individual corresponds to the parameter combination of R = 5 and α = 1500.

[0033] For each parameter combination represented by a grey wolf individual, the representative signal of the cluster is subjected to variational modal decomposition using the parameter set, and R intrinsic mode function components are obtained. The envelope entropy of the R intrinsic mode function components is calculated. Since the envelope entropy can effectively represent the sparsity of the signal, the smaller the entropy value is, the stronger the periodicity of the signal is and the better the modal separation effect is, therefore the minimum value of the R envelope entropies is selected as the fitness value of the grey wolf individual. The grey wolf optimization algorithm simulates the social hierarchy and group hunting behavior of the wolf pack, and continuously updates the positions of all grey wolves in the population in multiple iterations, i.e., dynamically adjusts the combination of R and α, and searches in the direction of better fitness value, i.e., smaller minimum envelope entropy. When the preset maximum number of iterations or the fitness value converges, the algorithm stops, and the parameter combination corresponding to the position of the leader wolf at this time, i.e., the optimal modal number and penalty factor of the data cluster.

[0034] S3, a weighted graph is constructed with the data sources in the cluster as nodes and the causal relationship strength as edges, and a graph attention network is used to aggregate the intrinsic mode function components to generate a cluster-level state feature vector representing the state of each cluster.

[0035] Within each constructed spatio-temporal causal cluster, each data source within the cluster is regarded as a graph node, and the edge weights between nodes are defined by the corresponding values in the previously calculated causal strength matrix, thereby constructing a weighted graph containing rich topological information. For each data source sequence within the cluster, the optimal parameters R and a determined in step S2 are used to perform variational modal decomposition, obtaining R intrinsic modal function components. The values of the R intrinsic modal function components at the current time are spliced into a one-dimensional vector as the initial feature representation of the corresponding node in the weighted graph. The graph attention network is used to process the graph data, and the sensor data source processing process includes: calculating the attention weight between the target node and its neighbor nodes in the graph structure through the self-attention mechanism, which accurately represents the importance of different neighbor nodes to the state update of the target node; secondly, the features of all neighbor nodes are weighted and summed based on the attention weight, thereby updating the feature representation of the target node. After depth feature extraction by multiple layers of graph attention network, graph pooling operation is performed on the feature vectors of all nodes in the cluster, such as element-wise averaging or maximum value of all node feature vectors, to aggregate the dispersed node-level features into a high-dimensional and compact cluster-level state feature vector.

[0036] S4, fuse all cluster-level state feature vectors with position encoding representing the physical topology of the carbon dioxide pipeline network, and input the fused feature sequence into the Transformer encoder network to extract global spatio-temporal dependence features, and calculate the current state parameters of the carbon dioxide pipeline network through the fully connected layer.

[0037] In this step, a corresponding position encoding is generated for each spatio-temporal causal cluster. Specifically, the arithmetic mean of the physical milepost numbers of all sensors in the spatio-temporal causal cluster on the pipeline is calculated, which is taken as the representative physical position of the spatio-temporal causal cluster and is mapped into a position encoding vector. The position encoding vector is added element-wise to the cluster-level state feature vector of the corresponding cluster, thereby forming a fused feature sequence containing both local state features and global spatial position information. The fused feature sequence is input into the Transformer encoder, and the multi-head self-attention mechanism inside the Transformer encoder can calculate the correlation weight between any two cluster-level state feature vectors, thereby effectively capturing and detecting long-distance, cross-regional state dependence relationships in the pipeline network. The sequence features output by the Transformer encoder are input into a multi-layer perceptron network containing two hidden layers after a global average pooling layer, and the output layer of the network is flexibly set according to task requirements. For example, it can be set as a single neuron containing a Sigmoid activation function, which is used to output the probability value of pipeline leakage; or it can be set as a classification layer containing a Softmax activation function, which is used to accurately determine the specific operating state of the pipeline network, which is normal, blocked or leaking.

[0038] In an optional embodiment, the fusing of all cluster-level state feature vectors with the location encoding representing the physical topology of the carbon dioxide pipeline network comprises: calculating the average of the physical milepost numbers of all data source nodes within each spatiotemporal causal cluster on the carbon dioxide pipeline network as the representative location of the spatiotemporal causal cluster; mapping the representative location into a location encoding vector with the same dimension as the cluster-level state feature vector using the sine-cosine function; fusing the location encoding vector with the cluster-level state feature vector by element-wise addition.

[0039] Specifically, the arithmetic mean of the milepost numbers is calculated by collecting the actual physical installation locations of all data source sensors within the cluster on the carbon dioxide pipeline network, usually represented by milepost numbers. For example, if a cluster contains three sensors, the locations of the three sensors are 10.2 kilometers, 10.5 kilometers, and 11.2 kilometers, respectively, then the representative location of the cluster is 10.63 kilometers. After obtaining the representative location scalar value, the scalar value needs to be converted into a high-dimensional vector, i.e., a location encoding vector, in order to be fused with the cluster-level state feature vector. This conversion is completed using a set of sine and cosine functions with different frequencies. Assuming that the dimension of the cluster-level state feature vector is , and the representative location is , the th element of the location encoding vector is calculated as , and the th 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 element-wise added to the generated cluster-level state feature vector described above to realize the fusion of the two, so that the feature vector contains both the behavior information and the static spatial location information of the cluster.

[0040] In an optional embodiment, the inputting of the fused feature sequence into the Transformer encoder network to extract global spatiotemporal dependency features comprises: inputting the fused feature sequence into a network stacked by 6 layers of encoders, each layer of encoder containing an 8-head self-attention module and a feedforward neural network module, both modules using residual connection and layer normalization operation between them, detecting the long-distance dependency relationship between different cluster-level state feature vectors in the fused feature sequence.

[0041] Specifically, an input feature sequence of length C is constructed, where each element in the sequence corresponds to a cluster-level state feature vector fused with position encoding, assuming that the whole pipe network is divided into C clusters. The sequence is input into a deep network stacked by 6 encoder layers with the same structure. Inside each encoder layer, the data passes through an 8-head self-attention module, which allows each cluster-level state feature vector fused with position encoding in the sequence to interact with all other cluster-level state feature vectors fused with position encoding in the sequence, aggregates information from a global perspective to update its own feature representation by calculating the correlation weight between them. The 8-head mechanism means that the above attention calculation process is executed 8 times in parallel, and each head learns a different dependency pattern, thereby enhancing the model's ability to perceive and detect complex spatio-temporal features.

[0042] After the self-attention module, the data flows through a feedforward neural network module composed of two linear transformation layers and an activation function, which aims to transform the feature representation of each cluster to enhance the model's feature expression ability. To ensure the stability and efficiency of deep network training, residual connection and layer normalization are applied after each module, i.e., the self-attention module and the feedforward neural network module. Residual connection directly adds the input to the output of the module, effectively preventing the gradient vanishing 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 the 6 encoder layers, the model can gradually build a deep understanding of the long-distance, global spatio-temporal dependencies between all clusters in the entire pipe network, and output a set of highly abstract global spatio-temporal dependency features.

[0043] In an optional embodiment, the current state parameters of the carbon dioxide pipe network are calculated by a fully connected layer, including: The global spatio-temporal dependency features output by the Transformer encoder network are flattened and input into a fully connected network containing two neurons, and the pressure and temperature values of the preset key sections along the pipe network are output by a linear activation function as the current state parameters.

[0044] The output feature sequence from the Transformer encoder network is received, where each element in the output feature sequence is a deep feature representation of a spatio-temporal causal cluster. For example, if there are Q clusters and each feature vector has a dimension of M, the output is a QxM matrix. A flattening operation is performed on the matrix, i.e., all elements of the matrix are sequentially connected into a one-dimensional long vector with a length of QxM, which collects global spatio-temporal dependency information extracted from all clusters.

[0045] The flattened feature vectors are input into a fully connected network consisting of two layers of neurons, also known as a multi-layer perceptron. The first layer of the network is a hidden layer that performs linear transformation and activation 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 of 15 key sections along the pipe network, the output layer will be configured with 30 neurons. The output layer uses a linear activation function, as it allows the output values to be any real number, matching the prediction task of continuous physical quantities such as pressure and temperature. Finally, the 30 values output by the fully connected network correspond to the current pressure and temperature prediction values of the 15 key sections, respectively, thus achieving accurate calculation of the operating state of the carbon dioxide pipe network.

[0046] While the present specification has shown and described a number of embodiments of the present application, it is to be understood that such embodiments are merely illustrative of and not restrictive on the present application. Many modifications, changes, and substitutions can now occur to one skilled in the art without departing from the spirit and scope of the present application.

Claims

1. A method for sensing the state of a carbon dioxide pipe network based on multi-source heterogeneous data fusion, characterized in that, The method comprises the following steps: obtaining pressure data, temperature data, flow data, sound wave data and optical fiber sensing data of a carbon dioxide pipeline network and constructing multi-source heterogeneous time series data; based on the multi-source heterogeneous time series data within a preset time window, calculating a causal relationship strength matrix between each data source, and dividing the multi-source heterogeneous time series data into different spatio-temporal causal clusters by using a spectral clustering algorithm; for each spatio-temporal causal cluster, determining the mode number and penalty factor of variational mode decomposition according to the characteristics of the data in the cluster, and performing variational mode decomposition on each multi-source heterogeneous time series data in the cluster to obtain a series of intrinsic mode function components; constructing a weighted graph with the data sources in the cluster as nodes and the causal relationship strength as edges, and aggregating the intrinsic mode function components by using a graph attention network to generate a cluster-level state feature vector representing the state of each cluster; fusing all the cluster-level state feature vectors with position encoding representing the physical topology of the carbon dioxide pipeline network, inputting the fused feature sequence into a Transformer encoder network to extract global spatio-temporal dependence features, and solving the current state parameters of the carbon dioxide pipeline network through a fully connected layer.

2. The method for carbon dioxide pipe network state perception based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The calculation of the causal relationship strength matrix between each data source comprises: For any two data sequences X and Y corresponding to any two data sources in the multi-source heterogeneous time series data, the causality strength The causality strength is determined by calculating the KL divergence as follows: ; where, is the state of the data sequence X at the next time instant; and denote the past state vectors of the data sequences X and Y at time instant and and is the embedding dimension, denotes the conditional probability distribution; traversing all data source pairs to form the causal relationship strength matrix.

3. The method according to claim 1, wherein, The division of the multi-source heterogeneous time series data into different spatio-temporal causal clusters by using the spectral clustering algorithm comprises: inputting the causal relationship strength matrix as an adjacency matrix into the spectral clustering algorithm, calculating the eigenvectors of the Laplacian matrix, and performing K-Means clustering on the eigenvectors in a low-dimensional space, so as to divide the data sources with close relationships into the same spatio-temporal causal cluster.

4. The method for carbon dioxide pipe network state perception based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The determination of the mode number and penalty factor of variational mode decomposition according to the characteristics of the data in the cluster comprises: setting the search range of the mode number R to [2, 9] and the search range of the penalty factor a to [500, 2500]; using the grey wolf optimization algorithm, taking the minimum envelope entropy of all intrinsic mode function components after decomposition as the fitness function, and obtaining the optimal mode number R and penalty factor a combination through iterative optimization.

5. The method for carbon dioxide pipe network state perception based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The aggregation of the intrinsic mode function components by using the graph attention network comprises: concatenating the R intrinsic mode function components obtained by decomposing each data source into a one-dimensional vector as the initial feature representation of the corresponding node of the data source in the weighted graph; allocating different attention weights to the neighbor nodes of each target node through the self-attention mechanism in the graph attention network, and performing weighted summation on the features of the neighbor nodes to update the feature representation of each target node; performing graph pooling operation on all the node feature representations updated by the graph attention network to aggregate and generate the cluster-level state feature vector.

6. The method for carbon dioxide pipe network state perception based on multi-source heterogeneous data fusion according to claim 5, characterized in that, The graph pooling operation is global average graph pooling.

7. The method for carbon dioxide pipe network state perception based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The fusion of all the cluster-level state feature vectors with the position encoding representing the physical topology of the carbon dioxide pipeline network comprises: The average of physical milepost numbers of all data source nodes in each spatiotemporal causal cluster on the carbon dioxide pipeline network is calculated as a representative position of the spatiotemporal causal cluster; the representative position is mapped into a position encoding vector with the same dimension as the cluster-level state feature vector by using a sine-cosine function; and the position encoding vector and the cluster-level state feature vector are fused by element-wise addition.

8. The method for carbon dioxide pipe network state perception based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The fused feature sequence is input into a Transformer encoder network to extract global spatiotemporal dependency features, including: The fused feature sequence is input into a network stacked by six layers of encoders, each layer of encoder including an 8-head self-attention module and a feedforward neural network module, and residual connection and layer normalization operations being used between the two modules to detect long-distance dependency relationships between different cluster-level state feature vectors in the fused feature sequence.

9. The method for carbon dioxide pipe network state perception based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The current state parameters of the carbon dioxide pipeline network are calculated by a fully connected layer, including: Global spatiotemporal dependency features output by the Transformer encoder network are flattened and input into a fully connected network including two layers of neurons, and pressure values and temperature values of preset key sections along the pipeline are output by a linear activation function as the current state parameters.

Images