Oil pipeline monitoring method and system based on space-time modeling and multi-dimensional analysis

Abstract

The invention discloses an oil pipeline monitoring method and system based on spatio-temporal modeling and multi-dimensional analysis, and belongs to the technical field of oil and gas pipeline safety monitoring, data processing and intelligent prediction.The method comprises the steps that multi-source spatio-temporal data is collected to construct a probability generation model; performing inversion on the model by using the observation value to reconstruct four-dimensional posterior probability distribution; performing causal inference on the posterior probability distribution to generate a dynamic causal information flow map; and analyzing topological evolution of the atlas to generate a stability monitoring report. According to the method, a technical path of combining probability modeling based on an information field theory and causal dynamics inversion is adopted, and accurate prediction of a systematic risk critical transition precursor can be realized by reconstructing a pipeline holographic state field and analyzing topological evolution of a causal network of the pipeline holographic state field; and the operation safety and the intelligent monitoring level of the long-distance oil pipeline are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas pipeline safety monitoring, data processing and intelligent prediction, in particular to an oil pipeline monitoring method and system based on space-time modeling and multi-dimensional analysis. BACKGROUND

[0002] Long-distance oil pipelines are lifelines of national energy strategy, and their safe and stable operation is of great importance. To ensure pipeline integrity, existing pipeline monitoring usually relies on data collection by pressure, flow and temperature sensors along the line, combined with regular line inspection and maintenance, to monitor the basic operating state of the pipeline, which plays a fundamental role in preventing major safety accidents.

[0003] However, as the service life of the pipeline increases and the external environment becomes more complex, the existing monitoring method has certain limitations. Most methods focus on threshold alarm or post-detection of specific failures of a single physical quantity, lack the ability to analyze multi-source data in space-time coupling to identify early evolution characteristics of systemic risks, and are difficult to achieve the leap from "fault diagnosis" to "failure prognosis". SUMMARY

[0004] To solve the above problems, the present application provides an oil pipeline monitoring method and system based on space-time modeling and multi-dimensional analysis, which adopts a technical path combining probabilistic modeling based on information field theory and causal dynamics inversion, and can realize accurate prediction of systemic risk critical transition precursors by reconstructing the holographic state field of the pipeline and analyzing the topological evolution of its causal network, significantly improving the operation safety and intelligent monitoring level of long-distance oil pipelines.

[0005] The above object can be achieved by the following scheme:

[0006] The oil pipeline monitoring method based on space-time modeling and multi-dimensional analysis comprises collecting multi-source space-time data by sensors arranged along the oil pipeline, fitting the statistical distribution of the multi-source space-time data, constructing and generating a probabilistic generative model, using the multi-source space-time data as observation values, and inverting the probabilistic generative model using a preset inference algorithm to calculate and reconstruct a four-dimensional posterior probability distribution, performing causal inference analysis on the four-dimensional posterior probability distribution to generate a dynamic causal information flow graph, and analyzing the topological structure evolution of the dynamic causal information flow graph to generate a stability monitoring report.

[0007] Optionally, the constructing and generating a probabilistic generative model comprises: performing space-time alignment processing on the multi-source space-time data to generate a standardized space-time data set, calculating space-time mutual covariance based on the standardized space-time data set to generate a multi-dimensional state coupling tensor, and constructing and generating a probabilistic generative model based on the multi-dimensional state coupling tensor.

[0008] Optionally, the constructing and generating the probabilistic generative model based on the multi-dimensional state coupling tensor comprises: performing dimension reduction and feature extraction on the multi-dimensional state coupling tensor to generate a low-dimensional core tensor; performing a mapping process based on the low-dimensional core tensor to construct and generate a Gaussian process covariance kernel function; and performing process regression modeling by using the Gaussian process covariance kernel function to obtain the probabilistic generative model.

[0009] Optionally, the calculating and reconstructing the four-dimensional posterior probability distribution comprises: globally optimizing and solving the probabilistic generative model to generate a global approximate distribution; quantifying uncertainty of the global approximate distribution to identify and extract a high-variance key area; performing local fine sampling in the high-variance key area to obtain a sampling result, and fusing the sampling result with the global approximate distribution to obtain the four-dimensional posterior probability distribution.

[0010] Optionally, the method further comprises: performing topological structure analysis on a spatiotemporal distribution characteristic of the high-variance key area to generate a local complexity adjustment instruction; performing local parameterization reconstruction on the Gaussian process covariance kernel function based on the local complexity adjustment instruction to generate an adaptive covariance kernel function; and performing online updating on the probabilistic generative model by using the adaptive covariance kernel function to obtain an updated probabilistic generative model.

[0011] Optionally, the generating the dynamic causal information flow graph comprises: performing information transmission quantization on a time series of the four-dimensional posterior probability distribution to generate an initial causal relationship matrix; performing iterative pruning and removing indirect association on the initial causal relationship matrix to obtain a sparse causal network skeleton; and performing path quantization on information transmission intensity based on the sparse causal network skeleton to generate a dynamic causal information flow graph.

[0012] Optionally, the method further comprises: calculating and generating a spatiotemporal causal distance matrix based on paths and weights of the dynamic causal information flow graph; fusing the spatiotemporal causal distance matrix with a preset standard spatial distance matrix to construct a hybrid distance metric; and reconstructing the Gaussian process covariance kernel function by using the hybrid distance metric.

[0013] Optionally, the generating the stability monitoring report comprises: generating a network spectral gap time series by graph analysis calculation based on an adjacency matrix of the dynamic causal information flow graph; constructing and generating an uncertainty spatial weight field based on a spatial distribution of the high-variance key area; weighting the network spectral gap time series by applying the uncertainty spatial weight field to calculate a weighted vulnerability index; and comparing a change rate of the weighted vulnerability index with a preset instability alarm threshold to generate the stability monitoring report.

[0014] Optionally, the method further comprises: quantifying the uncertainty in the four-dimensional posterior probability distribution as a cognitive uncertainty field; calculating and generating a causal network centrality field based on the dynamic causal information flow graph; coupling the cognitive uncertainty field and the causal network centrality field to construct and generate an information value graph; and generating an optimal sensor deployment strategy through an optimization solving process based on the information value graph.

[0015] Based on the same inventive concept, the application also provides an oil pipeline monitoring system based on spatiotemporal modeling and multidimensional analysis, which comprises: a probability generating model construction module, which is used to collect multi-source spatiotemporal data from sensors arranged along an oil pipeline, and to perform statistical distribution fitting on the multi-source spatiotemporal data to construct and generate a probability generating model; a posterior probability inversion module, which is used to take the multi-source spatiotemporal data as observation values, and to perform inversion on the probability generating model by using a preset inference algorithm to calculate and reconstruct a four-dimensional posterior probability distribution; a causal information flow analysis module, which is used to perform causal inference analysis on the four-dimensional posterior probability distribution to generate a dynamic causal information flow graph; and a stability evaluation module, which is used to analyze the topological structure evolution of the dynamic causal information flow graph to generate a stability monitoring report.

[0016] Compared with the prior art, the application has the following advantages:

[0017] 1. The application reconstructs a complete and continuous pipeline holographic state field by constructing a probability generating model and performing inversion on multi-source data, and improves the monitoring method from relying on a large number of sensors along the line to measure local physical quantities to reconstructing a complete and continuous pipeline holographic state field through sparse observation. This mode can reveal the state correlation across regions and at the system level that cannot be perceived by traditional methods, and fundamentally improves the depth and breadth of the overall state cognition of the pipeline;

[0018] 2. The application analyzes the topological structure evolution of the dynamic causal information flow graph, rather than only evaluating the static risk. It can quantify the dynamic recovery ability of the system at the information transmission level, so as to identify the critical precursor of the system approaching instability in advance before an explicit failure such as a physical leakage or failure occurs, and realize the “prognosis” rather than “diagnosis” of the systematic risk of the pipeline.

[0019] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0021] Figure 1 FIG. 1 is a flowchart of a pipeline monitoring method based on spatiotemporal modeling and multidimensional analysis according to an embodiment of the present application.

[0022] Figure 2 FIG. 2 is a pipeline state posterior probability distribution diagram according to an embodiment of the present application.

[0023] Figure 3 FIG. 3 is a dynamic causal information flow diagram according to an embodiment of the present application.

[0024] Figure 4 FIG. 4 is a system weighted vulnerability index evolution and early warning diagram according to an embodiment of the present application.

[0025] Figure 5 FIG. 5 is a structural diagram of a pipeline monitoring system based on spatiotemporal modeling and multidimensional analysis according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0027] With reference to Figure 1 , one embodiment of the present application proposes a pipeline monitoring method and system based on spatiotemporal modeling and multidimensional analysis, adopts a technical path combining probability modeling based on information field theory and causal dynamics inversion, and can realize accurate prediction of precursors of critical transition of systematic risks by reconstructing a holographic state field of a pipeline and analyzing topological evolution of a causal network, thereby significantly improving operation safety and intelligent monitoring level of long-distance oil pipelines.

[0028] The method according to the embodiment specifically includes:

[0029] A sensor arranged along an oil pipeline collects multi-source spatiotemporal data, and performs statistical distribution fitting on the multi-source spatiotemporal data to construct and generate a probability generation model;

[0030] The multi-source spatio-temporal data are taken as observation values, and a preset inference algorithm is used to inverse the probability generating model, to calculate and reconstruct a four-dimensional posterior probability distribution;

[0031] The four-dimensional posterior probability distribution is analyzed for causal inference, to generate a dynamic causal information flow atlas;

[0032] The topological structure evolution of the dynamic causal information flow atlas is analyzed, to generate a stability monitoring report.

[0033] The technical path combining probability modeling based on information field theory and causal dynamics inversion can realize accurate prediction of the precursor of critical transition of systemic risk by reconstructing the holographic state field of the pipeline and analyzing the topological evolution of the causal network, and significantly improve the operation safety and intelligent monitoring level of the long-distance oil pipeline.

[0034] Optionally, the constructing and generating a probability generating model comprises:

[0035] The multi-source spatio-temporal data are processed for spatio-temporal alignment, to generate a standardized spatio-temporal data set;

[0036] Specifically, this step aims to unify the sensor data from different types and different layout positions to a standard spatio-temporal reference. A data preprocessing process first synchronously corrects the time stamps of all sensors, for example, using the second pulse signal of the Network Time Protocol (NTP) or Global Positioning System (GPS) to eliminate the deviation of the time reference. Subsequently, the physical position coordinates of all sensors are mapped to a one-dimensional mileage coordinate based on the pipeline centerline as the reference through a spatial interpolation algorithm. Finally, all spatio-temporally aligned data are standardized, for example, using the Z-score standardization method to eliminate the influence of different physical dimensions and numerical ranges, so that different types of features are comparable. All data processed in this way jointly constitute a standardized spatio-temporal data set.

[0037] Based on the standardized spatio-temporal data set, a spatio-temporal cross-covariance is calculated, to generate a multi-dimensional state coupling tensor;

[0038] Specifically, this step aims to learn and quantify the strength of the intrinsic correlations between different physical quantities, between different spatial locations from the data. A covariance computation process will compute the cross-covariance between any two types of sensor data, for example, the pressure signal at one location and the strain signal at another location, at different time delays. By computing the cross-covariance values for all possible pairs of sensor types, all possible pairs of spatial distances, and all possible time delays, a high-dimensional array can be constructed. Each element of this array precisely quantifies the degree of correlation between two state variables within the pipeline system at a specific spatio-temporal scale. This high-dimensional array is the final generated multi-dimensional state coupling tensor.

[0039] Based on the multi-dimensional state coupling tensor, a probabilistic generative model is constructed and generated.

[0040] Specifically, this step aims to utilize the high-dimensional tensor generated in the previous step, which contains the intrinsic correlation laws, to construct a probabilistic model that can describe the overall behavior. A non-parametric Bayesian method, such as Gaussian Process (GP), is adopted. In this method, the multi-dimensional state coupling tensor generated in the previous step is used as the core basis for constructing the Gaussian Process covariance kernel function. This data-driven, structured kernel function defines the correlation between the states of any two spatio-temporal points on the pipeline. Finally, the Gaussian Process completely defined by this covariance kernel function constitutes the probabilistic generative model. This model can give a complete probabilistic prediction, including the mean and variance, for the state of any unmeasured location on the pipeline.

[0041] Optionally, the step of constructing and generating a probabilistic generative model based on the multi-dimensional state coupling tensor comprises:

[0042] Dimensionality reduction and feature extraction are performed on the multi-dimensional state coupling tensor to generate a low-dimensional core tensor.

[0043] Specifically, this step aims to extract the most core and representative structural features of the intrinsic laws from the multi-dimensional state coupling tensor, which contains a large amount of correlation information, has extremely high dimensions, and is difficult to directly process. In this step, a tensor decomposition algorithm, such as Tucker decomposition or tensor sequential decomposition, is adopted. Through this algorithm, the original high-dimensional tensor can be decomposed into the product of a core tensor and multiple factor matrices. This core tensor is the low-dimensional core tensor, which retains the most important correlation patterns and structural information in the original tensor in a compressed form, greatly reducing the computational complexity of subsequent modeling.

[0044] Based on the low-dimensional core tensor, a mapping process is performed to construct and generate a Gaussian Process covariance kernel function.

[0045] Specifically, in the mapping process based on the low-dimensional core tensor, in the step of constructing and generating the Gaussian process covariance kernel function, the step is to convert the core structure features extracted in the previous step into a mathematically complete kernel function that can define a Gaussian process. A mapping process will use the feature scale and direction information contained in the low-dimensional core tensor to parameterize a composite kernel function. For example, a composite kernel function can be the product of a periodic kernel and a radial basis function (RBF) kernel, where the anisotropic length scale parameter of the radial basis function kernel is determined by the principal component of the low-dimensional core tensor. An exemplary parameterized covariance kernel function can be represented by the formula:

[0046]

[0047] where k(x i ,x j ) is the covariance between any two spatiotemporal points x i and x j ; D is the feature dimension; l d is the length scale hyperparameter of the dth dimension, which is obtained by mapping the low-dimensional core tensor. Through this step, a Gaussian process covariance kernel function that accurately reflects the complex spatiotemporal coupling characteristics is constructed and generated.

[0048] Using the Gaussian process covariance kernel function to perform process regression modeling to obtain a probabilistic generative model.

[0049] Specifically, this step is the final step of defining the probabilistic generative model. A Gaussian process (GP) is completely defined by a mean function and a covariance function. In this step, the Gaussian process covariance kernel function generated in the previous step is combined with a pre-set mean function to define a complete Gaussian process regression model. This model is the final Gaussian process model obtained as a probabilistic generative model. Based on any given observation data, the model can give a complete probabilistic prediction of the state of any unmeasured point in the entire spatiotemporal field, which conforms to the Gaussian distribution and includes the mean and variance.

[0050] Optionally, the calculating and reconstructing the four-dimensional posterior probability distribution comprises:

[0051] Solving the probabilistic generative model globally to generate a global approximate distribution;

[0052] Specifically, since the complete posterior inference of the obtained Gaussian process model has a huge amount of calculation, it is difficult to meet the real-time requirement, and a variational inference (VI) algorithm is used in this step. The algorithm aims to find a parameterized, simpler probability distribution that approximates the real, complex posterior probability distribution as much as possible. This process is achieved by maximizing an objective function called "evidence lower bound (ELBO)", so as to quickly obtain a global approximate distribution that covers the entire pipeline space-time field within an acceptable calculation cost.

[0053] The uncertainty of the global approximate distribution is quantified, and a high-variance key area is identified and extracted;

[0054] Specifically, the purpose of this step is to automatically identify the area with the highest information value and the highest uncertainty in the global approximate result. The uncertainty is quantified by calculating the difference between the global approximate distribution and the prior distribution, and a commonly used measure is the point-by-point Kullback-Leibler divergence (KL divergence). By calculating the KL divergence value of all spatial positions, the area with the largest difference between the model cognition and the prior assumption can be identified, which is the high-variance key area. A KL divergence calculation can be represented by the formula:

[0055]

[0056] Where D KL is the KL divergence value, the larger the value, the higher the uncertainty; Q(z) is the global approximate posterior distribution of the point obtained by variational inference; P(z) is the prior distribution of the point.

[0057] In the high-variance key area, local refined sampling is performed to obtain a sampling result, and the sampling result is fused with the global approximate distribution to obtain a four-dimensional posterior probability distribution.

[0058] Specifically, in the step of local fine sampling in the high-variance focus area and fusing the sampling result with the global approximate distribution, to improve the prediction accuracy in the high-risk area, a Markov Chain Monte Carlo (MCMC) method such as the Metropolis-Hastings algorithm is used in this step. The algorithm only in the high-variance focus area identified in the previous step, through thousands of iterations of sampling, generates a large number of sample points conforming to the real posterior distribution. Subsequently, through a data fusion process such as importance sampling, these high-precision local sampling results obtained in the high-variance area are used to correct or "enhance" the global approximate distribution generated in the first step. Through this hybrid inference process, a final four-dimensional posterior probability distribution is finally reconstructed and generated, which has both global coverage and high accuracy and high reliability in key areas, as shown in Figure 2 As shown, the state field along a certain profile of the pipeline reconstructed by the application is shown, which not only gives the most likely mean prediction, but also quantifies the "uncertainty" or "degree of confidence" of the prediction result at each location through the confidence band.

[0059] Optionally, the method further comprises:

[0060] topological structure analysis is performed on the spatio-temporal distribution characteristics of the high-variance focus area to generate a local complexity adjustment instruction;

[0061] Specifically, this step aims to mine the potential structural reasons that cause the model to be "confused" from the uncertainty distribution of the model. A topological analysis algorithm will first perform geometric morphological analysis on the extracted high-variance focus area, for example, calculate its extension direction, branching condition or aggregation density and other topological structure characteristics. By comparing these characteristics with a pre-set rule base, the possible reasons for the insufficient prediction ability of the model in this area can be inferred, for example, it may be because the physical field change in this area is too drastic, which exceeds the expression ability of the original kernel function. Based on this inference, a local complexity adjustment instruction is generated, which contains suggestions on what adjustments need to be made to the model structure, such as "increase the local flexibility of the kernel function in this area" or "introduce a new kernel function component that can describe short-range changes".

[0062] based on the local complexity adjustment instruction, the Gaussian process covariance kernel function is locally parameterized and reconstructed to generate an adaptive covariance kernel function;

[0063] Specifically, this step is the core of performing model structure adaptive optimization. A kernel function reconstruction process receives the local complexity adjustment instruction generated in the previous step. According to the instruction, a new adaptive covariance kernel function is generated by modifying the constructed original Gaussian process covariance kernel function. An exemplary reconstruction method is to add the original global kernel function and a local kernel function that is activated only in the high-variance key area, as shown in the formula:

[0064] k adaptive (x i ,x j )=k global (x i ,x j )+w(x i ,x j )*k local (x i ,x j ),

[0065] where k adaptive (x i ,x j ) is the reconstructed adaptive covariance kernel function; k global (x i ,x j ) is the original global kernel function; k local (x i ,x j ) is a kernel function with stronger local fitting ability, such as a rational quadratic kernel; w(x i ,x j ) is a weight function, which has a significant non-zero value only when x i and x j are located in the high-variance key area.

[0066] Using the adaptive covariance kernel function, the probability generation model is updated online to obtain an updated probability generation model.

[0067] Specifically, this step is the last link of completing the entire model optimization closed loop. The adaptive covariance kernel function with more complex structure generated in the previous step is replaced with the original kernel function in the obtained probability generation model. This replacement operation completes the online update of the probability generation model. The updated probability generation model finally obtained has stronger expression ability and fitting flexibility of the kernel function in the key area, so that more accurate and reliable probabilistic prediction of the state of these complex areas can be made in the subsequent inference process.

[0068] Optionally, the generating dynamic causal information flow map includes:

[0069] performing information transfer quantification on the time series of the four-dimensional posterior probability distribution to generate an initial causal relationship matrix;

[0070] Specifically, to achieve information transfer quantification on the time series of the four-dimensional posterior probability distribution to generate an initial causal relationship matrix, this step aims to preliminarily quantify the information flow intensity that may exist between any two spatiotemporal points on the pipeline. An information quantification process will extract the state time series of any two spatial position points from the obtained four-dimensional posterior probability distribution. Subsequently, a nonlinear, asymmetric metric method, such as transfer entropy, is used to calculate the information transfer amount from the time series of one point to the time series of another point. A transfer entropy calculation can be represented by the formula:

[0071]

[0072] where T j→i (t) represents the information transfer amount from sequence j to sequence i at time t; x i,t+1 is the state of sequence i at the next time; and are the historical states of sequences i and j at the past k and l time steps, respectively. By calculating the transfer entropy between all position pairs, a fully connected, directed initial causal relationship matrix can be constructed. Iterative pruning of the initial causal relationship matrix removes indirect associations to obtain a sparse causal network skeleton;

[0073] performing iterative pruning of the initial causal relationship matrix to remove indirect associations to obtain a sparse causal network skeleton;

[0074] Specifically, this step aims to refine the true, direct causal conduction path from the relationship matrix obtained in the previous step, which contains a large amount of redundant information. A network pruning process uses a causal discovery algorithm based on conditional independence testing, such as the PC algorithm or the FCI algorithm. The algorithm will test each causal connection in the matrix. For example, when testing the connection from node X to node Z, the algorithm will traverse all possible intermediate nodes Y and test whether the original information transfer amount between X and Z significantly decreases or disappears after "controlling" or "given" the information of the intermediate node Y. If it disappears, it is determined that the connection from X to Z is an indirect association due to transmission through Y, and should be pruned. Through this iterative testing and pruning of all connections, a sparse causal network skeleton containing only direct causal relationships is finally obtained.

[0075] performing path quantification of information transfer intensity based on the sparse causal network skeleton to generate a dynamic causal information flow map.

[0076] Specifically, to realize path quantification of information transmission intensity based on the sparse causal network skeleton, a dynamic causal information flow graph is generated. This step is responsible for the final weighting and visualization on the purified network skeleton. A path quantification process assigns a weight to each remaining edge in the sparse causal network skeleton, which can be determined by the corresponding transfer entropy value calculated in the first step. The collection of all these weighted nodes and edges constitutes the final dynamic causal information flow graph. The graph intuitively and quantitatively shows the fundamental propagation path and intensity of abnormal or state changes in the pipeline system, as shown in Figure 3 The graph shows the causal information flow between monitoring points at different positions on the pipeline in the form of a network graph. The thickness and color depth of the edges represent the intensity of information transmission, clearly revealing the key propagation path of abnormal states in the system.

[0077] Optionally, the method further comprises:

[0078] Based on the paths and weights of the dynamic causal information flow graph, a spatiotemporal causal distance matrix is calculated and generated;

[0079] Specifically, this step aims to quantify the causal conduction relationship contained in the graph into a new distance metric. A graph algorithm, such as Dijkstra's algorithm or Floyd-Warshall's algorithm, is applied to the generated dynamic causal information flow graph. The algorithm calculates the shortest weighted path length between any two nodes in the graph. In this calculation, the weight of the edge is defined as the inverse of the information transmission intensity, so the stronger the information transmission, the shorter the "causal distance". Through this algorithm, the shortest causal path length between all node pairs can be calculated, which together constitutes a spatiotemporal causal distance matrix.

[0080] The spatiotemporal causal distance matrix is fused with a preset standard spatial distance matrix to construct a hybrid distance metric;

[0081] Specifically, the purpose of this step is to create a new metric that reflects both physical proximity and causal correlation. Through a weighted fusion function, the spatiotemporal causal distance matrix generated in the previous step is weighted and summed with a traditional Euclidean spatial distance matrix calculated based on the three-dimensional geometric coordinates of the pipeline. The fusion process is controlled by a preset fusion weight factor, which adjusts the importance of causal distance in the final metric, thereby constructing the final hybrid distance metric.

[0082] The hybrid distance metric is used to reconstruct the Gaussian process covariance kernel function.

[0083] Specifically, to realize the reconstruction of the Gaussian process covariance kernel function by using the hybrid distance metric, this step is to use the more insightful distance metric generated in the previous step to optimize the core of the constructed probabilistic generative model. A kernel function reconstruction process will replace all distance calculation links in the Gaussian process covariance kernel function with the hybrid distance metric generated in the previous step. Through this reconstruction, the original probabilistic generative model is replaced by a new version of the causal enhanced Gaussian process model that can perceive causal transmission paths, so that the subsequent prediction conforms to the internal dynamic law.

[0084] Optionally, the generating the stability monitoring report comprises:

[0085] Based on the adjacency matrix of the dynamic causal information flow graph, a network spectral gap time series is generated through graph analysis calculation;

[0086] Specifically, this step aims to quantify the global connectivity and dynamic response capability of the entire information flow network. First, from the generated dynamic causal information flow graph, a graph Laplacian matrix corresponding to the graph is constructed. Then, by performing eigenvalue decomposition on the graph Laplacian matrix, the second smallest eigenvalue is calculated, which is defined as the spectral gap of the network in network science. Since the dynamic causal information flow graph evolves over time, repeating this calculation process can obtain a time series composed of a series of spectral gap values, i.e. the network spectral gap time series. This sequence can reflect the macroscopic trend of the efficiency of information transmission.

[0087] Based on the spatial distribution of the high-variance key area, an uncertainty spatial weight field is constructed and generated;

[0088] Specifically, this step aims to quantify the spatial distribution of the uncertainty recognized by the model itself. A weight field generation process will spatially map the identified high-variance key area. For each spatial position on the pipeline, if it belongs to the high-variance key area, a higher weight value is assigned; otherwise, a lower weight value is assigned. Through this process, an uncertainty spatial weight field that reflects the reliability difference of the prediction ability of the model in different regions is constructed and generated.

[0089] Apply the uncertainty spatial weight field to the network spectral gap time series for weighting, and calculate a weighted vulnerability index;

[0090] Specifically, this step creatively couples the "causal network stability" with the "model cognitive uncertainty". A weighted calculation process will weight each spectral gap value in the network spectral gap time series with its corresponding weight value in the uncertainty space weight field at the same time. In this way, the contribution of a change in the causal network topology to the final index will be amplified if it occurs in an area with high model cognitive uncertainty. An exemplary weighted vulnerability index calculation can be represented by the formula:

[0091]

[0092] where I vuln (t) is the weighted vulnerability index at time t; λ2(t) is the network spectral gap value at this time; W uncertain (t) is the comprehensive weight value in the uncertainty space weight field corresponding to the state at this time.

[0093] The rate of change of the weighted vulnerability index is compared with a preset instability alarm threshold to generate a stability monitoring report.

[0094] Specifically, to realize the comparison of the rate of change of the weighted vulnerability index with the preset instability alarm threshold to generate a stability monitoring report, this step is the final risk determination link. A trend analysis process will continuously calculate the first derivative of the weighted vulnerability index time series, i.e. its rate of change. According to complex system theory, when approaching the critical instability point, the state recovery will become abnormally slow, which will be reflected in the sharp growth of the weighted vulnerability index. When the rate of change of the index exceeds a preset instability alarm threshold according to historical data or safety specifications, it is determined that there is a precursor to the imminent risk. At this time, a stability monitoring report containing information such as risk level, possible instability area and key causal path will be automatically generated, as shown in Figure 4 which shows the evolution of the system vulnerability index over time, which integrates "network stability" and "model uncertainty", and shows how the system triggers corresponding alarms when the index exceeds different levels of warning thresholds.

[0095] Optionally, the method further comprises:

[0096] quantifying the uncertainty in the four-dimensional posterior probability distribution as a cognitive uncertainty field;

[0097] Specifically, this step aims to generate a quantitative map that characterizes the model's "cognitive grasp" of the state of the pipeline at each point. An uncertainty quantification process will calculate the variance or information entropy of the four-dimensional posterior probability distribution at each spatiotemporal grid cell. The higher the variance or entropy value of a region, the greater the model's uncertainty in predicting it, and the more ambiguous its understanding of the region's state. The collection of uncertainty quantification values for all grid cells collectively constitutes the cognitive uncertainty field. The generation of this field marks the first step in "metacognition" from pure state prediction to examining the model's own cognitive abilities.

[0098] Based on the dynamic causal information flow map, calculate and generate a causal network centrality field;

[0099] Specifically, this step aims to identify key spatiotemporal nodes in the complex causal relationship network that play a pivotal role in information transmission and risk propagation. A network topology analysis process employs one or more network centrality algorithms, such as PageRank, Betweenness Centrality, or Eigenvector Centrality, to calculate the generated dynamic causal information flow map. Through these algorithms, a quantitative centrality score can be calculated for each spatiotemporal node in the map. The higher the score, the greater the node's influence in the entire information flow network, making it a key path node for potential abnormal propagation. The collection of centrality scores for all nodes constitutes the causal network centrality field.

[0100] Couple the cognitive uncertainty field and the causal network centrality field to construct and generate an information value map;

[0101] Specifically, this step is a core, non-obvious innovation of the present invention, and its purpose is to identify those "neither uncertain nor important" spatiotemporal positions that are most valuable for further observation. An information value calculation process will perform a point-by-point nonlinear coupling operation on the two independent fields generated in the previous two steps. An exemplary coupling function can be represented by the formula, which generates information value by calculating the geometric mean of the two fields and amplifying it by a synergy gain term:

[0102]

[0103] where V I (v) is the information value at spatiotemporal cell v; U(v) is the value of the cognitive uncertainty field corresponding to the cell; C(v) is the value of the causal network centrality field corresponding to the cell; λ is a synergy weight coefficient; corr(U,C) vis the spatial correlation of two fields in the local neighborhood of the point, which is used to give extra value gain to the area where both uncertainty and importance present high values. Through this operation, an information value map that quantifies the degree of information value of adding a new observation at any location to reduce the global uncertainty is constructed.

[0104] Based on the information value map, an optimal sensor deployment strategy is generated by an optimization solving process.

[0105] Specifically, this step aims to transform the quantification result of information value into an executable engineering decision. An optimization solving process takes the information value map generated in the previous step as input, and its optimization goal is to select a set of deployment locations that maximize the sum of information value of these locations under the constraint of limited sensor budget cost. This problem can be modeled as a constrained combinatorial optimization problem. By solving it with algorithms such as the maximum weight cover algorithm in graph theory or heuristic greedy algorithm, a solution containing a set of recommended deployment location three-dimensional coordinates can be obtained. This solution is the final generated optimal sensor deployment strategy, which can be used to guide the placement of new sensors in the next round of survey or to plan the optimal inspection path for mobile detection equipment, so as to accurately invest valuable detection resources in the most valuable place for cognition improvement.

[0106] Embodiment 1:

[0107] In order to verify the feasibility of the application in implementation, the application is applied to a long-distance, high-pressure oil pipeline that passes through mountainous areas and densely populated areas. Due to frequent geological activities and frequent third-party construction interference, the traditional monitoring method that relies on local pressure and flow threshold alarms is difficult to provide early and accurate warning for the progressive risk caused by multi-factor coupling.

[0108] In order to verify the beneficial effects of the application, a 200-kilometer section of the pipeline containing multiple high-consequence areas is selected for a six-month comparative test. The control group uses the traditional monitoring method based on real-time data threshold alarm of the SCADA system (Supervisory Control and Data Acquisition). The experimental group deploys the method and system of the application, which collects multi-source spatio-temporal data through distributed fiber optic temperature sensors (DTS), distributed acoustic sensors (DAS), strain gauges, and high-frequency pressure sensors placed along the line.

[0109] In this embodiment, the data collected by all sensors is fused, and combined with historical operation data and geographic information of the pipeline to construct a probabilistic generative model. The core of the model is a Gaussian process covariance kernel function defined by a "multi-dimensional state coupled tensor", which can represent the complex spatio-temporal correlation between any two points on the pipeline in different physical fields such as pressure, temperature and strain.

[0110] Subsequently, based on the probabilistic generative model, a hybrid inference algorithm reversely reconstructs the four-dimensional posterior probability distribution of the entire pipeline. The distribution not only gives the state mean value of each spatio-temporal point, but more importantly, quantifies the "uncertainty" of the model prediction. On this basis, a self-optimizing closed loop is started, which identifies the "high variance key areas" with the highest prediction uncertainty of the model, and automatically performs local parameterization reconstruction on the kernel function of the probabilistic generative model in these areas, so that the cognition of these key areas is more accurate in the next inference.

[0111] Then, through causal inference analysis based on transfer entropy of the four-dimensional posterior probability distribution, and removal of indirect correlation, a dynamic causal information flow atlas is constructed. The atlas reveals the most likely propagation path of abnormal disturbances within the pipeline system. At the same time, a meta-learning feedback closed loop is triggered, which calculates the "spatio-temporal causal distance" according to the atlas, and uses this distance metric with more physical meaning to reconstruct the kernel function of the underlying probabilistic generative model, so that it learns and understands the internal causal conduction rules.

[0112] Finally, by performing spectral analysis on the Laplacian matrix of the dynamic causal information flow atlas, a "weighted vulnerability index" representing the global stability can be calculated. At the same time, by coupling the "cognitive uncertainty field" of the model itself with the "causal centrality field" of the network, an "information value atlas" is also generated to guide the optimal deployment location of future new sensors.

[0113] After a 6-month comparative test, the invention shows significant technical advantages in early warning capability, fault attribution accuracy and adaptability. See Tables 1, 2 and 3 for specific data.

[0114] Table 1 Comparison of early warning capability

[0115] Test Group Identified risks / alarms Average warning lead time Key risk underreporting rate control group 8 2.1 20.00％ Experimental group 15 72.5 0.00％ Performance improvements +87.5％ +3352％ -100％

[0116] Table 2 Comparison of abnormal event attribution accuracy

[0117] Test Group Number of incidents requiring root cause analysis First-time attribution accuracy control group 8 37.50％ Experimental group 15 92.30％ Performance improvements - +146.1％

[0118] Table 3 Verification table of model uncertainty self-adaptive optimization effect

[0119] Testing cycle Model average prediction uncertainty Month 1 0.89 Month 3 0.52 6th month 0.34 Optimization effect -61.80％

[0120] The above Tables 1 to 3 record the comparative data of the application method in the actual application of long-distance oil pipeline monitoring, which details the excellent performance of the application in early warning, root cause analysis and model adaptive optimization.

[0121] In Table 1, the core difference between the two methods in the early warning ability is shown. The data shows that during the 6-month test period, the control group had a total of 8 threshold alarms that required manual emergency intervention, with an average early warning period of only 2.1 hours, and 2 times failed to successfully warn due to unclear signal characteristics. The experimental group successfully identified 15 potential risk evolution trends, with an average early warning period of 72.5 hours, and a zero key risk false negative rate. This clearly proves that the application can foresee potential risks earlier and more reliably than traditional threshold alarm methods by analyzing the evolution trend of the state.

[0122] In Table 2, the accuracy in diagnosing the root cause of the problem is quantified. Among all the identified abnormal events, the matching accuracy of the "causal source" indicated by the dynamic causal information flow diagram generated by the application with the actual fault root cause verified by subsequent engineering is as high as 92.3%. The attribution accuracy of the control group is only 37.5% due to the lack of causal analysis capability, and it is highly dependent on manual investigation. This shows that the causal inference capability of the application can provide accurate guidance for the maintenance and repair of the pipeline.

[0123] In Table 3, the effectiveness of the unique online self-optimization mechanism of the application is tested. The data shows that at the beginning of the test, the average prediction uncertainty of the model for the entire pipeline state is high. But with the continuous operation of the feedback loop, the model learns from the areas it "doesn't know" and the causal relationships it identifies, and its average prediction uncertainty decreases by 61.8% in 6 months. This strongly proves that the meta-learning mechanism included in the application can enable the monitoring system to evolve over time, and the prediction ability "becomes more accurate with use".

[0124] Based on the same inventive concept, the application also provides an oil pipeline monitoring system based on spatiotemporal modeling and multi-dimensional analysis, as shown in Figure 5 The system comprises:

[0125] A probabilistic generative model construction module is configured to collect multi-source spatiotemporal data from sensors arranged along the oil pipeline, perform statistical distribution fitting on the multi-source spatiotemporal data, and construct and generate a probabilistic generative model;

[0126] A posterior probability inversion module is configured to use the multi-source spatiotemporal data as observation values, and use a preset inference algorithm to invert the probabilistic generative model, calculate and reconstruct a four-dimensional posterior probability distribution;

[0127] a causal information flow analysis module for performing causal inference analysis on the four-dimensional posterior probability distribution to generate a dynamic causal information flow graph;

[0128] a stability evaluation module for analyzing the topological structure evolution of the dynamic causal information flow graph to generate a stability monitoring report.

[0129] It should be noted that the function division and information interaction between the above-mentioned various modules are logical, and can be integrated in the same software platform or distributed in physical implementation. The connection between them represents the data flow and control flow, and aims to cooperatively achieve the building energy consumption dynamic optimization goal of the present application. The above only describes exemplary embodiments of the present application, and cannot limit the protection scope of the present application.

Claims

1. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis is characterized by: The method comprises: Sensors deployed along the oil pipeline collect multi-source spatiotemporal data, perform statistical distribution fitting on the multi-source spatiotemporal data, and construct and generate a probability generation model; The multi-source spatiotemporal data are used as observations, and a preset inference algorithm is used to invert the probability generation model to calculate and reconstruct a four-dimensional posterior probability distribution; Performing causal inference analysis on the four-dimensional posterior probability distribution to generate a dynamic causal information flow graph; Analyze the topological structure evolution of the dynamic causal information flow graph and generate a stability monitoring report.

2. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis according to claim 1 is characterized in that: The constructing and generating the probability generation model includes: Performing spatiotemporal alignment processing on the multi-source spatiotemporal data to generate a standardized spatiotemporal dataset; Calculating spatiotemporal cross-covariance based on the standardized spatiotemporal data set to generate a multi-dimensional state coupling tensor; Based on the multi-dimensional state coupling tensor, a probability generation model is constructed and generated.

3. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis according to claim 2 is characterized in that: The constructing and generating a probability generation model based on the multi-dimensional state coupling tensor includes: Performing dimensionality reduction and feature extraction on the multi-dimensional state coupling tensor to generate a low-dimensional core tensor; Performing a mapping process based on the low-dimensional core tensor to construct and generate a Gaussian process covariance kernel function; The Gaussian process covariance kernel function is used to perform process regression modeling to obtain a probability generation model.

4. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis according to claim 3 is characterized in that: Calculating and reconstructing the four-dimensional posterior probability distribution includes: Performing global optimization on the probability generation model to generate a global approximate distribution; quantifying uncertainty of the global approximate distribution, identifying and extracting key areas with high variance; In the high variance key area, local refined sampling is performed to obtain sampling results, and the sampling results are fused with the global approximate distribution to obtain a four-dimensional posterior probability distribution.

5. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis according to claim 4 is characterized in that: The method further comprises: Performing a topological structure analysis on the spatiotemporal distribution characteristics of the high-variance key area to generate a local complexity adjustment instruction; Based on the local complexity adjustment instruction, locally parameterize and reconstruct the Gaussian process covariance kernel function to generate an adaptive covariance kernel function; The probability generation model is updated online using the adaptive covariance kernel function to obtain an updated probability generation model.

6. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis according to claim 1 is characterized in that: Generating a dynamic causal information flow graph includes: Performing information transfer quantization on the time series of the four-dimensional posterior probability distribution to generate an initial causal relationship matrix; Iteratively pruning the initial causal relationship matrix to remove indirect relationships to obtain a sparse causal network skeleton; Based on the sparse causal network skeleton, the information transmission intensity is path-quantified to generate a dynamic causal information flow graph.

7. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis according to claim 3 is characterized in that: The method further comprises: Based on the paths and weights of the dynamic causal information flow graph, a spatiotemporal causal distance matrix is ​​calculated and generated; The spatiotemporal causal distance matrix is ​​fused with a preset standard spatial distance matrix to construct a hybrid distance metric; The Gaussian process covariance kernel function is reconstructed using the hybrid distance metric.

8. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis according to claim 5 is characterized in that: Generating the stability monitoring report includes: Based on the adjacency matrix of the dynamic causal information flow graph, a network spectrum gap time series is generated through graph analysis and calculation; Based on the spatial distribution of the high variance key areas, an uncertainty spatial weight field is constructed and generated; Applying the uncertainty spatial weight field to the network spectrum slot time series for weighting, and calculating a weighted vulnerability index; The rate of change of the weighted vulnerability index is compared with a preset instability alarm threshold to generate a stability monitoring report.

9. The oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis according to claim 1 is characterized in that: The method further comprises: quantifying the uncertainty in the four-dimensional posterior probability distribution as an epistemic uncertainty field; Based on the dynamic causal information flow graph, calculating and generating a causal network centrality field; Couple the epistemic uncertainty field with the causal network centrality field to construct and generate an information value map; Based on the information value map, an optimal sensor deployment strategy is generated through an optimization solution process.

10. An oil pipeline monitoring system based on spatiotemporal modeling and multi-dimensional analysis, applied to an oil pipeline monitoring method based on spatiotemporal modeling and multi-dimensional analysis as claimed in any one of claims 1 to 9, characterized in that: The system comprises: A probability generation model building module is used to collect multi-source spatiotemporal data from sensors deployed along the oil pipeline, perform statistical distribution fitting on the multi-source spatiotemporal data, and build and generate a probability generation model; A posterior probability inversion module is used to use the multi-source spatiotemporal data as observation values ​​and use a preset inference algorithm to invert the probability generation model to calculate and reconstruct a four-dimensional posterior probability distribution; A causal information flow analysis module is used to perform causal inference analysis on the four-dimensional posterior probability distribution and generate a dynamic causal information flow graph; The stability assessment module is used to analyze the topological structure evolution of the dynamic causal information flow graph and generate a stability monitoring report.

Images