Underground pipe network defect intelligent identification and early warning method and system based on deep learning

Abstract

The invention provides an underground pipe network defect intelligent identification and early warning method and system based on deep learning, and relates to the technical field of pipe network detection, and the method comprises the steps: employing a pipeline detection robot to obtain multi-source data, and carrying out the preprocessing; mapping the features to a manifold space, constructing a similarity matrix, and carrying out geometric difference weighted fusion; constructing an incidence matrix to calculate a spatial distance and mechanical strength, and carrying out iterative propagation to generate a defect characteristic spectrum; and acquiring an evolution sequence by adopting a self-adaptive sliding window, determining influence factors based on anti-fact intervention, and calculating a state transition probability to determine an optimal maintenance scheme. According to the method, the pipe network defect identification accuracy is improved, and pipeline state prediction and maintenance decision intelligence are realized.

Description

Technical Field

[0001] The present invention relates to the field of pipe network detection technology, and in particular to a method and system for intelligent identification and early warning of underground pipe network defects based on deep learning. Background Art

[0002] Underground pipeline networks are a vital component of urban infrastructure, providing multiple functions such as water supply, drainage, gas, and electricity. Their safe operation is crucial to the normal functioning of cities. With urban development and the aging of pipeline networks, underground pipeline defects are becoming increasingly prominent. Damage such as pipe wall corrosion, cracks, and misaligned interfaces is increasing, posing a serious threat to urban safety and residents' lives. Traditional pipeline inspections rely primarily on manual inspections and simple mechanical equipment, which are inefficient and difficult to fully identify hidden defects. In recent years, pipeline inspection robot technology has gradually been applied to underground pipeline inspections. Equipped with various sensors, they can collect multi-source data such as internal pipeline images, acoustics, and lasers, providing technical support for pipeline status assessment.

[0003] With the development of deep learning technology, pipeline defect detection methods based on image recognition have made significant progress. Existing technologies typically use algorithms such as convolutional neural networks to analyze pipeline images and identify surface defects such as cracks and corrosion. Furthermore, some research is attempting to combine acoustic detection with laser scanning to build multimodal data fusion frameworks to improve the accuracy and comprehensiveness of pipeline defect detection.

[0004] Existing underground pipeline defect identification and early warning technologies still have obvious shortcomings. Existing methods have limited processing capabilities for multi-source heterogeneous data, making it difficult to effectively integrate detection information from different modalities, resulting in data silos and affecting the comprehensiveness and accuracy of defect identification. Traditional defect identification algorithms lack the ability to dynamically analyze the evolution of pipeline defects and often only focus on the static characteristics of defects, unable to accurately predict defect development trends and potential risks. Existing early warning methods lack a systematic pipeline maintenance decision-making support mechanism, making it difficult to formulate scientific and reasonable maintenance plans based on comprehensive information such as defect severity, development trends and influencing factors, resulting in irrational allocation of maintenance resources and difficulty in achieving precise and intelligent pipeline maintenance. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for intelligent identification and early warning of underground pipeline defects based on deep learning, which can solve the problems in the existing technology.

[0006] A first aspect of an embodiment of the present invention provides a method for intelligently identifying and warning underground pipeline network defects based on deep learning, comprising: Use pipeline inspection robots to obtain multi-source inspection data of underground pipeline networks and perform preprocessing to obtain initial feature data; The initial feature data is mapped to the manifold space through the feature mapping module to obtain the manifold feature data. A similarity matrix is ​​constructed in the manifold space to divide the feature subspace. The feature subspace is weightedly fused by geometric differences to obtain the pipeline defect representation vector. The pipeline defect characterization vector is constructed as an association matrix. The spatial distance association value and mechanical interaction strength between pipeline defect features are calculated to obtain a connection relationship matrix. Iterative propagation is performed to obtain a pipeline state sequence. A pipeline state evolution diagram is constructed and the propagation direction is optimized to generate a pipeline defect feature map. An adaptive sliding window is used to dynamically sample the pipeline defect feature map to obtain the pipeline defect evolution sequence. The influencing factor set is determined based on the counterfactual intervention strategy analysis. The pipeline state transition probability is calculated based on the influencing factor set, and the optimal pipeline maintenance plan is determined based on the multi-objective decision function. When the pipeline state transition probability exceeds the preset probability threshold, a pipeline network defect warning information is generated; Send pipeline network defect warning information and optimal pipeline maintenance plan to the monitoring terminal.

[0007] In an optional embodiment, the initial feature data is mapped to the manifold space through a feature mapping module to obtain manifold feature data, a similarity matrix is ​​constructed in the manifold space to divide the feature subspace, and geometric difference weighted fusion is performed on the feature subspace to obtain the pipeline defect characterization vector, including: Inputting the initial feature data into a feature mapping module, the feature mapping module calculating the neighbor relationship of sample points in the initial feature data to construct an adjacency graph, calculating the similarity relationship between the sample points in the adjacency graph to obtain a weight matrix, constructing a diagonal matrix based on the weight matrix and calculating a difference matrix, obtaining a projection matrix based on the difference matrix, and using the projection matrix to map the initial feature data to a manifold space to obtain manifold feature data; Constructing a local tangent plane in the manifold space where the manifold feature data is located, converting the global nonlinear space of the underground pipe network defects into a local linear space, calculating the shortest path between sample points in the manifold feature data in the local linear space as a geodesic distance, and converting it into a similarity score, and constructing a similarity matrix based on the similarity score; Based on the similarity matrix, the manifold feature data is divided into multiple feature subspaces. The local distribution characteristics of pipeline defects are calculated in each feature subspace to obtain a feature distribution matrix. A metric tensor is constructed based on the feature distribution matrix. The geometric difference values ​​between the multiple feature subspaces are calculated according to the metric tensor, and the geometric difference values ​​are converted into corresponding feature subspace weight coefficients. The feature distribution matrices of the multiple feature subspaces are multiplied by the corresponding weight coefficients and superimposed to obtain a defect representation vector.

[0008] In an optional embodiment, dividing the manifold feature data into multiple feature subspaces based on the similarity matrix includes: In the local linear space where the manifold feature data is located, a search radius is constructed for each sample point of the underground pipe network defect state, the number of neighboring sample points within the search radius is counted to obtain an initial local density value of the sample point, and the density ratio of the sample point to the corresponding neighboring sample point is calculated based on the initial local density value to obtain a local density weight; Multiplying the similarity value between each sample point in the similarity matrix by the corresponding local density weight to determine a density weighting coefficient, and using the density weighting coefficient to update the similarity matrix to obtain a density-aware similarity matrix; The manifold feature data is grouped and divided according to the similarity relationship between the sample points in the density-aware similarity matrix. When the density difference between any two sample points is less than a preset density threshold and the corresponding similarity relationship in the density-aware similarity matrix is ​​greater than the preset similarity threshold, the two sample points are divided into the same feature subspace to obtain multiple feature subspaces.

[0009] In an optional embodiment, pipeline defect characterization vectors are constructed as an association matrix, spatial distance association values ​​and mechanical interaction strengths between pipeline defect features are calculated to obtain a connection relationship matrix, iterative propagation is performed to obtain a pipeline state sequence, a pipeline state evolution diagram is constructed, and the propagation direction is optimized to generate a pipeline defect feature map, including: The pipeline defect characterization vector is constructed as an association matrix, the spatial distance association value between each group of pipeline defect features is calculated based on the association matrix, and normalized to obtain an initial connection matrix; Calculating the mechanical interaction strength between pipeline defect features based on the spatial distance correlation value in the initial connection matrix, and converting the mechanical interaction strength into a connection weight to obtain a connection relationship matrix; Constructing a pipeline state transfer matrix based on the connection relationship matrix, performing iterative propagation on pipeline defect features, fusing state information of adjacent pipeline features in each iteration process to update the current pipeline state, and obtaining a pipeline state sequence; Constructing a state evolution graph based on the pipeline state sequence, calculating the transition probability between nodes in the state evolution graph to obtain a propagation direction, and filtering the propagation direction according to a preset pipeline structure constraint condition to obtain an optimized direction; The pipeline status is layered based on the optimization direction, a hierarchical pipeline association structure is constructed, the importance index of the feature nodes in each layer is calculated, the feature nodes are screened based on the importance index, and a pipeline defect feature map is generated.

[0010] In an optional embodiment, a pipeline state transition matrix is ​​constructed based on the connection relationship matrix, and iterative propagation is performed on pipeline defect features. During each iteration, state information of adjacent pipeline features is integrated to update the current pipeline state. The pipeline state sequence obtained includes: Calculating the eigenvalue decomposition of the connection relationship matrix to obtain an eigenvector matrix, using the eigenvector matrix to construct frequency response functions of different scales, and performing a convolution operation on the connection relationship matrix and the corresponding frequency response function to obtain pipeline state transfer matrices of multiple scales; Pipeline defect feature propagation is performed on the pipeline state transfer matrix of each scale to obtain multiple sets of pipeline state sequences; the degree of difference between adjacent states in each set of pipeline state sequences is calculated to obtain a propagation response value, and attention weights are determined based on the propagation response values; the multiple sets of pipeline state sequences are fused using the attention weights to obtain a fused pipeline state; The deviation between the fused pipeline state and the initial pipeline state is calculated to obtain a state residual, which is fed back to the pipeline defect feature propagation process. The attention weight is dynamically adjusted based on the state residual to adaptively adjust the intensity of pipeline defect feature propagation. The feature propagation update is repeatedly performed until the state residual is less than a preset residual threshold, thereby obtaining an optimized pipeline state sequence.

[0011] In an optional embodiment, an adaptive sliding window is used to dynamically sample the pipeline defect feature map to obtain a pipeline defect evolution sequence, a set of influencing factors is determined based on counterfactual intervention strategy analysis, a pipeline state transition probability is calculated based on the set of influencing factors, and an optimal pipeline maintenance plan is determined based on a multi-objective decision function, including: An adaptive sliding window is used to dynamically sample the pipeline defect feature map. The size of the sliding window is dynamically adjusted according to the frequency of change of the pipeline defect feature. The statistics and change trend characteristics of the pipeline features within the sliding window are calculated to obtain a local feature sequence. The local feature sequence is constructed into a temporal pyramid structure, pipeline feature representation is extracted at each level of the temporal pyramid structure, an inter-level feature propagation network is constructed to calculate the co-evolution relationship between pipeline features, and cross-level feature recursive update is performed based on the co-evolution relationship to obtain a pipeline defect evolution sequence; Constructing a counterfactual intervention strategy based on the pipeline defect evolution sequence to generate a control experiment scenario, and performing a deviation analysis between the control experiment scenario and the actual scenario to obtain a set of influencing factors; Constructing a pipeline defect state descriptor using the influencing factor set, calculating a state transition matrix under different pipeline maintenance plans based on the pipeline defect state descriptor, and performing probability distribution estimation on the state transition matrix to obtain a pipeline state transition probability; The pipeline state transition probability is constructed into a multi-objective decision function along with pipeline maintenance cost, maintenance result, and maintenance risk. An optimization solution is performed based on the multi-objective decision function to obtain a set of pipeline maintenance plans. The reliability score of each pipeline maintenance plan set is calculated, and the pipeline maintenance plan with the highest reliability score is selected to determine the optimal pipeline maintenance plan.

[0012] In an optional embodiment, a counterfactual intervention strategy is constructed based on the pipeline defect evolution sequence to generate a control experiment scenario, and a deviation analysis is performed between the control experiment scenario and the actual scenario to obtain a set of influencing factors including: Calculate the time series nodes in the pipeline defect evolution sequence where the feature change rate exceeds the preset change rate threshold, adjust the intensity of the pipeline defect status characteristics at the time series nodes, and generate counterfactual intervention strategies; Execute pipeline defect state simulation according to the counterfactual intervention strategy and pipeline constraint conditions, perform evolution calculation on the pipeline defect state under each intervention intensity, and generate multiple groups of control experimental scenarios; Calculating feature deviations between the control experimental scenario and the actual pipeline scenario, performing cluster analysis on the feature deviations to obtain feature deviation clusters, calculating the variance contribution rate of the pipeline features in each feature deviation cluster, and selecting features whose variance contribution rate exceeds a preset variance threshold as dominant factors; Calculating a time series correlation coefficient based on the dominant factors, determining a contribution weight of each dominant factor according to the time series correlation coefficient, and selecting a dominant factor whose contribution weight is greater than a preset contribution weight threshold; The correlation degree between the dominant factors is calculated, the dominant factors are ranked and sorted, and a set of influencing factors is generated.

[0013] A second aspect of an embodiment of the present invention provides an intelligent identification and early warning system for underground pipeline network defects based on deep learning, comprising: The first unit is used to obtain multi-source detection data of the underground pipeline network using a pipeline detection robot and perform preprocessing to obtain initial feature data; The second unit is used to map the initial feature data to the manifold space through the feature mapping module to obtain manifold feature data, construct a similarity matrix in the manifold space to divide the feature subspace, and perform geometric difference weighted fusion on the feature subspace to obtain the pipeline defect representation vector; The third unit is used to construct the pipeline defect representation vector into an association matrix, calculate the spatial distance association value and mechanical action strength between pipeline defect features to obtain a connection relationship matrix, perform iterative propagation to obtain a pipeline state sequence, construct a pipeline state evolution diagram and optimize the propagation direction to generate a pipeline defect feature map; The fourth unit is used to dynamically sample the pipeline defect feature map using an adaptive sliding window to obtain the pipeline defect evolution sequence, determine the influencing factor set based on the counterfactual intervention strategy analysis, calculate the pipeline state transition probability based on the influencing factor set, and determine the optimal pipeline maintenance plan based on the multi-objective decision function; The fifth unit is used to generate pipeline network defect warning information when the pipeline state transition probability exceeds a preset probability threshold; The sixth unit is used to send pipeline network defect warning information and the optimal pipeline maintenance plan to the monitoring terminal.

[0014] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In an embodiment of the present invention, through manifold space feature mapping and geometric difference weighted fusion, the characterization characteristics of pipeline defects are accurately extracted and enhanced, effectively solving the problem of insufficient recognition accuracy of traditional methods in complex environments, and significantly improving the accuracy and robustness of pipeline defect feature extraction; using iterative propagation and pipeline state evolution graph construction technology, deep mining of the correlation between pipeline defect features is achieved, and a complete pipeline defect feature map is established, enabling the system to fully grasp the laws of pipeline state evolution and provide reliable data support for early warning decision-making; combining counterfactual intervention strategies and multi-objective decision functions, the intelligent generation of pipeline defect early warning and maintenance plans is achieved, while ensuring the safe operation of the pipeline network, optimizing the allocation of maintenance resources, greatly improving the maintenance efficiency and economic benefits of underground pipeline networks, and having significant engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the process of the method for intelligent identification and early warning of underground pipeline network defects based on deep learning according to an embodiment of the present invention; Figure 2 This is the pipeline defect feature propagation and adaptive adjustment flow chart. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0019] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0020] Figure 1 FIG is a flow chart of an intelligent identification and early warning method for underground pipe network defects based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes: Use pipeline inspection robots to obtain multi-source inspection data of underground pipeline networks and perform preprocessing to obtain initial feature data; The initial feature data is mapped to the manifold space through the feature mapping module to obtain the manifold feature data. A similarity matrix is ​​constructed in the manifold space to divide the feature subspace. The feature subspace is weightedly fused by geometric differences to obtain the pipeline defect representation vector. The pipeline defect characterization vector is constructed as an association matrix. The spatial distance association value and mechanical interaction strength between pipeline defect features are calculated to obtain a connection relationship matrix. Iterative propagation is performed to obtain a pipeline state sequence. A pipeline state evolution diagram is constructed and the propagation direction is optimized to generate a pipeline defect feature map. An adaptive sliding window is used to dynamically sample the pipeline defect feature map to obtain the pipeline defect evolution sequence. The influencing factor set is determined based on the counterfactual intervention strategy analysis. The pipeline state transition probability is calculated based on the influencing factor set, and the optimal pipeline maintenance plan is determined based on the multi-objective decision function. When the pipeline state transition probability exceeds the preset probability threshold, a pipeline network defect warning information is generated; Send pipeline network defect warning information and optimal pipeline maintenance plan to the monitoring terminal.

[0021] In one specific embodiment, the multi-source inspection data acquired by the pipeline inspection robot primarily comes from a variety of sensor devices. Equipped with visual sensors such as high-definition cameras and infrared thermal imagers, the pipeline inspection robot can capture visible light and thermal images of the pipeline's inner wall. This image data is used to identify visual defects such as surface cracks, corrosion spots, and misaligned joints. The robot is also equipped with ultrasonic sensors to measure pipe wall thickness and internal structural integrity, enabling the detection of non-surface defects such as wall thinning, internal voids, and material degradation.

[0022] The inspection robot also features a laser scanning system, which uses laser ranging and 3D imaging technology to obtain precise internal geometric dimensions and deformation data for detecting pipe deformation, displacement, and geometric anomalies. Acoustic sensors, including acoustic emission and sound wave propagation detection equipment, collect acoustic signals from the pipe structure to identify microcrack propagation, leaks, and structural stress changes. For metal pipes, the robot is also equipped with electromagnetic sensors such as eddy current probes and magnetic flux detectors to detect changes in conductivity and magnetic anomalies, thereby identifying weak areas of corrosion and material fatigue.

[0023] To comprehensively assess the pipeline's environmental conditions, the inspection robot is equipped with environmental parameter sensors that measure internal parameters such as temperature, humidity, pH, and gas concentration, providing data for defect formation mechanism analysis. The robot's inertial measurement unit records precise position and attitude information for defect location and spatial mapping.

[0024] The specific data acquired include high-definition image sequences and thermal imaging data of the inner wall of the pipeline; pipe wall thickness measurement data and material sound velocity characteristics with an accuracy of ±0.1mm; three-dimensional point cloud data and geometric contour information of the interior of the pipeline with an accuracy of ±0.5mm; acoustic waveform data, spectrum characteristics and acoustic emission event records; electromagnetic characteristic parameters of metal pipelines and abnormal area distribution maps; environmental parameter time series data covering the temperature range of -10°C to 100°C, the humidity range of 0 to 100%, and the pH range of 0 to 14; and the three-dimensional coordinates and motion trajectory data of the inspection robot in the pipeline.

[0025] This multi-source data is acquired in real time by pipeline inspection robots and processed synchronously in time and space to form a complete pipeline status data set, providing a comprehensive data foundation for subsequent defect identification and early warning. Through the fusion and analysis of this multi-source, heterogeneous data, the system can more comprehensively and accurately identify pipeline defects, improving the reliability of underground pipeline network safety operations.

[0026] In an optional embodiment, the initial feature data is mapped to the manifold space through a feature mapping module to obtain manifold feature data, a similarity matrix is ​​constructed in the manifold space to divide the feature subspace, and geometric difference weighted fusion is performed on the feature subspace to obtain the pipeline defect representation vector, including: Inputting the initial feature data into a feature mapping module, the feature mapping module calculating the neighbor relationship of sample points in the initial feature data to construct an adjacency graph, calculating the similarity relationship between the sample points in the adjacency graph to obtain a weight matrix, constructing a diagonal matrix based on the weight matrix and calculating a difference matrix, obtaining a projection matrix based on the difference matrix, and using the projection matrix to map the initial feature data to a manifold space to obtain manifold feature data; Constructing a local tangent plane in the manifold space where the manifold feature data is located, converting the global nonlinear space of the underground pipe network defects into a local linear space, calculating the shortest path between sample points in the manifold feature data in the local linear space as a geodesic distance, and converting it into a similarity score, and constructing a similarity matrix based on the similarity score; Based on the similarity matrix, the manifold feature data is divided into multiple feature subspaces. The local distribution characteristics of pipeline defects are calculated in each feature subspace to obtain a feature distribution matrix. A metric tensor is constructed based on the feature distribution matrix. The geometric difference values ​​between the multiple feature subspaces are calculated according to the metric tensor, and the geometric difference values ​​are converted into corresponding feature subspace weight coefficients. The feature distribution matrices of the multiple feature subspaces are multiplied by the corresponding weight coefficients and superimposed to obtain a defect representation vector.

[0027] In one specific implementation, in the underground pipeline defect identification scenario, the initial feature data typically includes multi-source heterogeneous data, including pipeline inner wall image features, acoustic detection features, and dimensional measurement features. For example, for a section of urban underground pipeline network, the image features collected by the inspection robot may include surface texture, color changes, and morphological characteristics to identify corrosion and cracks; acoustic features include echo intensity and spectral distribution to detect changes in pipe wall thickness; and dimensional measurement features reflect pipe deformation and interface misalignment. Mapping these multi-source features into a unified manifold space is a key step in accurately identifying complex underground pipeline defects.

[0028] The acquired initial pipeline defect feature data is input into the feature mapping module. The feature mapping module constructs an adjacency graph by calculating the nearest neighbor relationships between sample points in the initial feature data. In underground pipeline network scenarios, sample points represent defect characteristics at different locations in the pipeline. For example, the data collected by the inspection robot along the pipeline includes information such as wall thickness variation, crack characteristics, and interface status. These features together constitute the initial feature data for underground pipeline network defects. When constructing the adjacency graph, pipeline network defect points with similar spatial locations and characteristics are connected to form a topological structure that characterizes the distribution characteristics of underground pipeline network defects. Specifically, for each sample point, K nearest neighbors (for example, K = 10) are selected and these nearest neighbors are determined by calculating the Euclidean distance. When sample points i and j are neighbors, a connection is established in the adjacency graph. Next, the similarity relationships between the sample points in the adjacency graph are calculated to obtain a weight matrix. The similarity relationships are calculated using a Gaussian kernel function. If the Euclidean distance between sample points i and j is 2.5 and the variance parameter is set to 0.8, the similarity weight between them is approximately 0.04. If the sample points are not adjacent, the corresponding weight is 0.

[0029] Based on the weight matrix, a diagonal matrix is ​​constructed, where the diagonal elements are the sum of all elements in the corresponding row of the weight matrix. For example, if the sum of the weights of a sample point and all its neighbors is 5.3, the value of that position in the diagonal matrix is ​​5.3. Then calculate the difference matrix, which is the diagonal matrix minus the weight matrix. Use the difference matrix to solve the generalized eigenvalue problem and obtain the projection matrix. The column vector of the projection matrix is ​​the eigenvector of the difference matrix, corresponding to the smallest non-zero eigenvalue. Select the first d eigenvectors to form the projection matrix, for example d=50, for dimensionality reduction. Use this projection matrix to map the initial feature data to the manifold space to obtain the manifold feature data.

[0030] In the manifold space where the manifold feature data resides, a local tangent plane must be constructed to transform the global nonlinear space of underground pipeline defects into a local linear space. For each sample point on the manifold, m neighboring points (for example, m = 15) are selected to form a local region within which the tangent plane is approximated. The tangent plane is determined using local principal component analysis, which calculates the covariance matrix of the sample points within the local region and extracts its principal directions as the basis vectors of the tangent plane.

[0031] In the local linear space, the shortest path between the sample points in the manifold feature data is calculated as the geodesic distance. Specifically, a distance graph is first established, in which the nodes are sample points and the weights of the edges are the Euclidean distances of the corresponding sample points in the local linear space. If two points are not in the same local area, there is no direct connection between them. The shortest path distance between any two points is calculated using the Dijkstra algorithm. For example, the geodesic distance between sample A and sample B is calculated to be 7.2. The geodesic distance is then converted to a similarity score using a negative exponential transformation: similarity score = negative exponential function (negative geodesic distance divided by the adjustment parameter). If the geodesic distance is 7.2 and the adjustment parameter is 3.5, the similarity score is approximately 0.13. Based on the similarity scores between all pairs of sample points, a complete similarity matrix is ​​constructed.

[0032] Based on the similarity matrix, the spectral clustering algorithm is used to divide the manifold feature data into multiple feature subspaces. The similarity matrix is ​​standardized, its eigenvalues ​​and eigenvectors are calculated, and the first k (for example, k=8) eigenvectors are selected to form a new feature representation. These new features are clustered using the K-means algorithm to divide the samples into different feature subspaces.

[0033] Within each characteristic subspace, the local distribution characteristics of pipeline defects are calculated to obtain a characteristic distribution matrix. These local distribution characteristics are obtained by calculating statistics of the samples within the subspace, including the mean vector, covariance matrix, and higher-order moments such as skewness and kurtosis. For example, for a subspace containing 150 sample points, the statistical characteristics such as the mean and variance of these samples in each dimension are calculated to form the characteristic distribution matrix for that subspace.

[0034] Based on the feature distribution matrix, a metric tensor is constructed to characterize the local geometric structure of the subspace. The metric tensor is obtained by calculating the local dispersion of sample points within the subspace, reflecting the directional sensitivity of the region. Using the covariance information in the feature distribution matrix, a Riemann metric tensor is constructed to represent the local deformation in different directions.

[0035] Using the metric tensor, calculate the geometric differences between multiple feature subspaces. The geometric difference between two subspaces is measured by the logarithmic Euclidean distance between their metric tensors. If the logarithmic Euclidean distance between the metric tensors of subspace 1 and subspace 2 is 2.8, it indicates that they have a large geometric difference.

[0036] Convert the geometric dissimilarity values ​​into corresponding feature subspace weight coefficients. Use softmax normalization to ensure that the sum of the weight coefficients is 1, with subspaces with greater geometric dissimilarity receiving higher weights. For example, if there are three subspaces with geometric dissimilarity values ​​of 2.8, 1.5, and 0.9, the converted weight coefficients might be 0.5, 0.3, and 0.2.

[0037] The feature distribution matrices of multiple feature subspaces are multiplied by their corresponding weight coefficients, and the results are superimposed to obtain the final defect representation vector. This representation vector integrates the feature distribution information of different subspaces while taking into account the geometric differences between each subspace. It can more comprehensively characterize the characteristics of pipeline defects, which is beneficial for subsequent defect classification and identification tasks.

[0038] In an optional embodiment, dividing the manifold feature data into multiple feature subspaces based on the similarity matrix includes: In the local linear space where the manifold feature data is located, a search radius is constructed for each sample point of the underground pipe network defect state, the number of neighboring sample points within the search radius is counted to obtain an initial local density value of the sample point, and the density ratio of the sample point to the corresponding neighboring sample point is calculated based on the initial local density value to obtain a local density weight; Multiplying the similarity value between each sample point in the similarity matrix by the corresponding local density weight to determine a density weighting coefficient, and using the density weighting coefficient to update the similarity matrix to obtain a density-aware similarity matrix; The manifold feature data is grouped and divided according to the similarity relationship between the sample points in the density-aware similarity matrix. When the density difference between any two sample points is less than a preset density threshold and the corresponding similarity relationship in the density-aware similarity matrix is ​​greater than the preset similarity threshold, the two sample points are divided into the same feature subspace to obtain multiple feature subspaces.

[0039] In one specific embodiment, a manifold feature data set is acquired to identify underground pipeline network defect states. This manifold feature data is derived from multi-source data collected during underground pipeline network inspections, including visual images, acoustic signals, and laser scanning. In underground pipeline networks, different types of defects, such as wall scaling, pipe corrosion, and loose joints, form complex distribution patterns in the feature space. Especially for older, long-operated pipelines, defect densities and types vary significantly across different regions. These differences are reflected in the manifold feature data, facilitating subsequent defect identification and classification. This data set contains multiple sample points, each representing a specific underground pipeline network defect state, such as a pipe rupture, corrosion severity, or loose joints. These sample points are distributed in a high-dimensional feature space, exhibiting specific manifold structural characteristics. The density distribution characteristics of manifold feature data are particularly important in underground pipeline network defect analysis. Due to the complexity of the underground environment, pipeline network defects often exhibit clustered distribution patterns. For example, corrosion areas are often concentrated in pipe sections with specific environmental conditions, and cracks often occur at stress concentrations. By constructing a density-aware similarity matrix, the spatial distribution pattern of such underground pipeline network defects can be captured more accurately, thereby improving the accuracy of subspace division and providing a more reliable data basis for subsequent defect assessment and early warning.

[0040] For the acquired manifold feature data, a search radius is constructed in the local linear space to determine the neighbor relationship of each sample point. Specifically, for each sample point in the sample set, a fixed search radius is set. For example, the search radius can be set to 30% of the average distance between all sample points. The number of neighboring sample points of each sample point within the search radius is counted to obtain the initial local density value of the sample point. For example, for sample point P1, there are 8 neighboring sample points within the search radius, so its initial local density value is 8; for sample point P2, there are 5 neighboring sample points within the search radius, so its initial local density value is 5.

[0041] Based on the calculated initial local density value, the density ratio between the sample point and its nearest neighboring sample points is calculated to obtain the local density weight. For each sample point, its initial local density value is divided by the initial local density values ​​of all its nearest neighboring sample points within the search radius to calculate the density ratio. For example, if the initial local density value of sample point P1 is 8, and the initial local density value of one of its nearest neighboring sample points, P3, is 4, then the density ratio between P1 and P3 is 8 / 4=2. The local density weight of the sample point is obtained by averaging the density ratios of the sample point and all its nearest neighboring sample points. If the average density ratio of the sample point P1 to its eight nearest neighboring sample points is 1.5, then the local density weight of P1 is 1.5.

[0042] Update the similarity matrix to generate a density-aware similarity matrix. Assume that in the original similarity matrix S, the similarity value between sample points P1 and P2 is 0.8, the local density weight of P1 is 1.5, and the local density weight of P2 is 1.2. The density weighting coefficient is calculated as 1.5 × 1.2 = 1.8. Multiplying the original similarity value of 0.8 by the density weighting coefficient of 1.8 yields a new similarity value between P1 and P2 in the density-aware similarity matrix of 0.8 × 1.8 = 1.44. In this manner, the similarity values ​​between each pair of sample points in the similarity matrix are updated to obtain the density-aware similarity matrix DS.

[0043] After obtaining the density-aware similarity matrix DS, the manifold feature data is grouped and divided based on the similarity relationships between the sample points. Two preset thresholds are set: a preset density threshold DT and a preset similarity threshold ST. The preset density threshold DT can be set to the 25th percentile of the initial local density difference of all sample points, for example, DT = 3; the preset similarity threshold ST can be set to the 75th percentile of all similarity values ​​in the density-aware similarity matrix DS, for example, ST = 1.2.

[0044] For any two sample points, if the absolute value of the difference between their initial local density values ​​is less than the preset density threshold DT, and their corresponding similarity relationship in the density-aware similarity matrix DS is greater than the preset similarity threshold ST, then the two sample points are classified into the same feature subspace. For example, the initial local density values ​​of sample points P1 and P2 are 8 and 6, respectively. The difference between their initial local density values ​​is |8-6|=2, which is less than the preset density threshold DT=3. At the same time, their similarity value in the density-aware similarity matrix DS is 1.44, which is greater than the preset similarity threshold ST=1.2. Therefore, P1 and P2 are classified into the same feature subspace.

[0045] By iteratively processing all sample point pairs, the entire manifold feature data set is ultimately divided into multiple feature subspaces. For example, in underground pipeline defect status data containing 100 sample points, it may be divided into five feature subspaces, representing defect states of different types or severities, such as the slight corrosion subspace (containing 23 sample points), the moderate corrosion subspace (containing 18 sample points), the severe corrosion subspace (containing 15 sample points), the slight rupture subspace (containing 25 sample points), and the severe rupture subspace (containing 19 sample points).

[0046] This manifold feature data partitioning method based on density-aware similarity matrix fully considers the density distribution characteristics of sample points in local areas, and can more accurately identify the inherent correlation of underground pipeline network defect status, providing more precise data support for subsequent defect analysis and prediction, thereby improving the efficiency of underground pipeline network maintenance and management.

[0047] In an optional embodiment, the pipeline defect characterization vector is constructed as an association matrix, the spatial distance association value and mechanical action strength between pipeline defect features are calculated to obtain a connection relationship matrix, iterative propagation is performed to obtain a pipeline state sequence, a pipeline state evolution diagram is constructed, and the propagation direction is optimized. The generation of the pipeline defect feature map includes: The pipeline defect characterization vector is constructed as an association matrix, the spatial distance association value between each group of pipeline defect features is calculated based on the association matrix, and normalized to obtain an initial connection matrix; Calculating the mechanical interaction strength between pipeline defect features based on the spatial distance correlation value in the initial connection matrix, and converting the mechanical interaction strength into a connection weight to obtain a connection relationship matrix; Constructing a pipeline state transfer matrix based on the connection relationship matrix, performing iterative propagation on pipeline defect features, fusing state information of adjacent pipeline features in each iteration process to update the current pipeline state, and obtaining a pipeline state sequence; Constructing a state evolution graph based on the pipeline state sequence, calculating the transition probability between nodes in the state evolution graph to obtain a propagation direction, and filtering the propagation direction according to a preset pipeline structure constraint condition to obtain an optimized direction; The pipeline status is layered based on the optimization direction, a hierarchical pipeline association structure is constructed, the importance index of the feature nodes in each layer is calculated, the feature nodes are screened based on the importance index, and a pipeline defect feature map is generated.

[0048] In a specific embodiment, the defect status of the underground pipeline network has obvious spatiotemporal correlation and evolution patterns. For example, corrosion defects in pipelines will expand over time and may induce structural damage in the surrounding area, forming a defect chain. By constructing an association matrix based on defect characterization vectors, the mutual influence relationship between defects at different locations in the underground pipeline network can be quantified, revealing the propagation path and mechanism of the defects. This feature map construction method based on deep learning can upgrade traditional isolated defect detection to a systematic pipeline health status assessment. The collected pipeline defect data includes feature information such as defect type, location, size, and shape. For the collected pipeline defect data, features are extracted and a characterization vector is constructed. The characterization vector includes spatial position coordinates (x, y, z), defect type (such as corrosion, crack, deformation, etc.), defect size (length, width, depth), material parameters and other information. For example, a pipeline defect feature vector can be represented as: [102.5, 35.6, 0.8, 1, 2.3, 1.5, 0.5, 7.8], where the first three values ​​represent the spatial coordinates, the fourth value represents the defect type, the next three values ​​represent the defect size, and the last value represents the material parameters. In underground pipeline inspection scenarios, this defect data typically comes from inspection robots equipped with multiple sensors, capable of identifying various defect types such as pipe wall deformation, cracks, and corrosion. The unique environment of underground pipelines results in defects exhibiting diverse characteristics, such as pipe wall corrosion depth, crack length and width, and the degree of interface misalignment. These data together constitute the pipeline defect representation vector, which is used for subsequent feature map construction and defect assessment.

[0049] All pipeline defect representation vectors are combined to construct an association matrix. Assuming there are f pipeline defects, each represented by a g-dimensional feature vector, the dimension of the association matrix is ​​f × g. Based on the association matrix, the spatial distance association values ​​between each set of pipeline defect features are calculated. Spatial distance association values ​​are calculated using Euclidean distance. For defect points 1 and 2, the spatial distance is the Euclidean distance between the spatial position coordinates of the two point representation vectors. For example, if the position coordinates of defect point 1 are (102.5, 35.6, 0.8) and the position coordinates of defect point 2 are (105.2, 36.4, 0.7), the spatial distance between the two points is 3.01. The results of the spatial distance calculations for all pairs of defect points are combined into a spatial distance matrix with a dimension of f × f.

[0050] Normalize the spatial distance matrix to obtain the initial connectivity matrix. This normalization method uses a Gaussian kernel function to convert spatial distances into correlation values ​​between 0 and 1. The smaller the spatial distance, the greater the correlation. For example, the spatial distance between defect points 1 and 2 is 3.01, and the normalized correlation value is 0.82. Each element in the initial connectivity matrix represents the spatial correlation between two defect points.

[0051] The mechanical interaction strength between pipeline defect features is calculated based on the spatial distance correlation values ​​in the initial connectivity matrix. The mechanical interaction strength takes into account factors such as defect type, size, and material parameters. For defect points 1 and 2, if the spatial correlation exceeds a threshold of 0.7 and the defect type similarity is greater than 0.6, the mechanical interaction strength is calculated; otherwise, the mechanical interaction strength is 0. For example, if the spatial correlation between defect points 1 and 2 is 0.82 and the defect type similarity is 0.85, the calculated mechanical interaction strength is 0.76.

[0052] The mechanical action strengths are converted into connection weights to obtain a connection relationship matrix. This conversion method uses a sigmoid function to map the mechanical action strengths to the interval [0, 1], which serve as connection weights. For example, if the mechanical action strength between defects 1 and 2 is 0.76, the converted connection weight is 0.82. The connection relationship matrix has dimensions f × f, and each element in the connection relationship matrix represents the connection weight between two defects.

[0053] The pipeline state transition matrix is ​​constructed based on the connection relationship matrix. Elements in the state transition matrix represent the probability of transitioning from one defect point state to another. Each row of the connection relationship matrix is ​​normalized so that the sum of the elements in each row is 1. This yields the state transition matrix. For example, the connection weights from defect point 1 to other defect points are [0, 0.82, 0.56, 0.43, 0], which after normalization are [0, 0.45, 0.31, 0.24, 0].

[0054] Iterative propagation is performed on the pipeline defect features, and the number of iterations is set to T = 10. The pipeline state vector S0 is initialized, and the initial state value of each defect point is 1. In each iteration, the state transfer matrix and the current state vector S t Multiply them together to get the updated state vector S t+1 For example, after the first iteration, the state value of defect point 1 changes from 1 to 0.88, and the state value of defect point 2 changes from 1 to 1.24. After T iterations, the final state sequence {S0, S1, ..., S T}.

[0055] A state evolution graph is constructed based on the pipeline state sequence. Nodes in the state evolution graph represent defect points, edges represent connections between defect points, and edge weights are the corresponding values ​​in the connection matrix. Transition probabilities between nodes in the state evolution graph are calculated to determine the propagation direction. For example, the transition probability from defect point 1 to defect point 2 is 0.45, indicating the probability of the defect propagating from point 1 to point 2.

[0056] The propagation direction is filtered based on preset pipeline structural constraints to obtain the optimized direction. These constraints include: the propagation direction must conform to the pipeline flow direction; the propagation distance must not exceed a preset threshold; and the propagation path must avoid areas with special structures. For example, if the propagation from defect point 3 to point 4 is opposite to the pipeline flow direction, this propagation direction is filtered out.

[0057] Pipeline status is layered based on the optimization direction, creating a hierarchical pipeline association structure. Starting from the source, defect points are divided into different layers based on the propagation direction. For example, defect point 1 is the source (layer 0), and its directly connected defect points 2, 3, and 4 are layer 1. Other defect points directly connected to layer 1 are layer 2, and so on.

[0058] The importance index of each feature node is calculated. The importance index comprehensively considers the node's degree centrality, closeness centrality, and betweenness centrality. For example, defect point 2 has a degree centrality of 3, a closeness centrality of 0.68, and a betweenness centrality of 0.45, resulting in an importance index of 0.72. Feature nodes are screened based on the importance index, and nodes with an importance index greater than a threshold of 0.5 are retained to generate a pipeline defect feature map. The resulting pipeline defect feature map contains the hierarchical relationships, connectivity, and importance information between defect points, visually demonstrating the distribution characteristics and evolution trends of pipeline defects.

[0059] In an optional embodiment, a pipeline state transition matrix is ​​constructed based on the connection relationship matrix, and iterative propagation is performed on the pipeline defect features. In each iteration, the state information of the adjacent pipeline features is integrated to update the current pipeline state. The pipeline state sequence obtained includes: Calculating the eigenvalue decomposition of the connection relationship matrix to obtain an eigenvector matrix, using the eigenvector matrix to construct frequency response functions of different scales, and performing a convolution operation on the connection relationship matrix and the corresponding frequency response function to obtain pipeline state transfer matrices of multiple scales; Pipeline defect feature propagation is performed on the pipeline state transfer matrix of each scale to obtain multiple sets of pipeline state sequences; the degree of difference between adjacent states in each set of pipeline state sequences is calculated to obtain a propagation response value, and attention weights are determined based on the propagation response values; the multiple sets of pipeline state sequences are fused using the attention weights to obtain a fused pipeline state; The deviation between the fused pipeline state and the initial pipeline state is calculated to obtain a state residual, which is fed back to the pipeline defect feature propagation process. The attention weight is dynamically adjusted based on the state residual to adaptively adjust the intensity of pipeline defect feature propagation. The feature propagation update is repeatedly performed until the state residual is less than a preset residual threshold, thereby obtaining an optimized pipeline state sequence.

[0060] In one specific embodiment, the propagation of defects in underground pipeline networks typically follows specific physical laws and topological constraints. Unlike traditional single-point defect analysis, this method, through eigenvalue decomposition and multi-scale frequency response functions, comprehensively captures the propagation characteristics of defect signals in underground pipeline networks at different spatial scales. For example, the local scale (parameter 0.5) primarily reflects defect correlations between adjacent pipe segments and is suitable for identifying microcrack propagation; the intermediate scale (parameter 1.0) captures interactions between groups of pipe segments and is suitable for analyzing the spread of corroded areas; and the global scale (parameter 2.0) focuses on changes in the structural integrity of the entire pipeline network system, helping to assess the impact of large-scale subsidence or geological changes on the pipeline network. The eigenvalue decomposition process of the connectivity matrix iteratively calculates the eigenvalues ​​and corresponding eigenvectors of the connectivity matrix. For an example underground pipeline network with 200 pipeline nodes and a 200×200 connectivity matrix, eigenvalue decomposition is performed on this matrix to obtain 200 eigenvalues ​​and their corresponding eigenvectors, which form the eigenvector matrix. Each column of the eigenvector matrix represents an eigenvector, representing the topological information of the pipeline network structure.

[0061] When constructing frequency response functions at different scales based on the eigenvector matrix, three different scale parameters were set: 0.5, 1.0, and 2.0, corresponding to the frequency responses at the local, medium, and global scales. By combining the eigenvector matrix with the corresponding scale parameters, three sets of frequency response functions were constructed. Each set of frequency response functions can capture the signal propagation characteristics of the pipeline network structure at different scales. The frequency response function with a scale parameter of 0.5 focuses on the local connectivity characteristics between adjacent nodes in the pipeline network; the frequency response function with a scale parameter of 1.0 can capture the connectivity patterns within the medium range; and the frequency response function with a scale parameter of 2.0 focuses on the expression of the global connectivity structure.

[0062] When performing the convolution operation between the connectivity matrix and the frequency response function, the frequency response function at each scale is convolved with the connectivity matrix separately. Specifically, for each scale parameter, the connectivity matrix is ​​dot-producted with the corresponding frequency response function, and the resulting matrix is ​​normalized. Ultimately, three pipeline state transition matrices at different scales are obtained. These transition matrices represent the state transition probabilities between pipeline network nodes at different scales.

[0063] During the pipeline defect feature propagation phase, the feature propagation process is performed for each scale's pipeline state transition matrix. Using the initial pipeline state vector as input, the pipeline defect features are propagated along the path defined by the transition matrix through iterative calculations. In implementation, the number of propagation steps is set to 10. For each propagation step, the current state vector is multiplied by the state transition matrix to obtain the state vector for the next moment. After completing 10 propagation steps, three pipeline state sequences (including the initial state) are obtained, each containing 11 state vectors.

[0064] When calculating the propagation response value, the degree of difference between adjacent state vectors in each pipeline state sequence is analyzed. Specifically, the Euclidean distance between adjacent state vectors is calculated and normalized to obtain the propagation response value. For the three state sequences in this example, 10 propagation response values ​​are obtained for each. Larger propagation response values ​​indicate more significant pipeline state changes during that propagation step, potentially indicating the presence of more important defect information.

[0065] When determining attention weights based on propagation response values, a softmax function is applied to convert them into attention weights. For each scale state sequence, the corresponding 10 propagation response values ​​are converted into weight coefficients summing to 1 using the softmax function. This gives higher attention weights to propagation steps with significant changes, while lower weights are given to steps with smaller changes.

[0066] During the fusion phase of multiple pipeline state sequences, the calculated attention weights are used to perform a weighted fusion of the three state sequences. Specifically, for each pipeline node, its weighted state value at three scales is calculated and summed to obtain the fused pipeline state. This fusion process considers the importance of defect information at different scales, providing a more comprehensive representation of the actual pipeline state.

[0067] During the state residual calculation process, the fused pipeline state is compared with the initial pipeline state, and the difference between the two is calculated to obtain the state residual vector. The state residual reflects the degree of correction to the initial state through the feature propagation process. In this example, the initial residual threshold is set to 0.01. If the norm of the state residual is less than this threshold, the state is considered to have converged; otherwise, further optimization is required.

[0068] During the state residual feedback and adaptive adjustment phase, the state residual information is fed back into the feature propagation process to dynamically adjust the attention weights and propagation strength. The propagation step sizes of the three scales are adjusted proportionally based on the size of the state residual. The propagation step size is increased for regions with large residuals and decreased for regions with small residuals. This adaptive adjustment mechanism enables more precise feature propagation optimization in problem areas.

[0069] The process of feature propagation, state fusion, and residual calculation is repeated until the state residual falls below the preset threshold of 0.01 or the maximum number of iterations, 50, is reached. In this example, after 12 iterations, the state residual drops to 0.0092, which is below the threshold, so iteration is stopped, resulting in the final optimized pipeline state sequence.

[0070] Conventional methods for identifying underground pipeline defects rely primarily on manual inspection or simple signal processing techniques, which are unable to effectively address the propagation characteristics of defects in complex pipeline structures. While existing deep learning methods incorporate neural network models, most employ fixed network structures and parameters, making them difficult to adapt to pipelines of varying sizes and complexities. To address these issues, the method of this embodiment proposes an adaptive approach based on eigenvalue decomposition and multi-scale feature propagation. A multi-scale frequency response function is introduced to capture defect characteristics across different ranges, and adaptive adjustment is achieved through an attention mechanism and residual feedback.

[0071] like Figure 2 As shown, the pipeline defect feature propagation and adaptive adjustment flow chart is presented.

[0072] In an optional embodiment, an adaptive sliding window is used to dynamically sample the pipeline defect feature map to obtain a pipeline defect evolution sequence, a set of influencing factors is determined based on a counterfactual intervention strategy analysis, a pipeline state transition probability is calculated based on the set of influencing factors, and an optimal pipeline maintenance plan is determined based on a multi-objective decision function, including: An adaptive sliding window is used to dynamically sample the pipeline defect feature map. The size of the sliding window is dynamically adjusted according to the frequency of change of the pipeline defect feature. The statistics and change trend characteristics of the pipeline features within the sliding window are calculated to obtain a local feature sequence. The local feature sequence is constructed into a temporal pyramid structure, pipeline feature representation is extracted at each level of the temporal pyramid structure, an inter-level feature propagation network is constructed to calculate the co-evolution relationship between pipeline features, and cross-level feature recursive update is performed based on the co-evolution relationship to obtain a pipeline defect evolution sequence; Constructing a counterfactual intervention strategy based on the pipeline defect evolution sequence to generate a control experiment scenario, and performing a deviation analysis between the control experiment scenario and the actual scenario to obtain a set of influencing factors; Constructing a pipeline defect state descriptor using the influencing factor set, calculating a state transition matrix under different pipeline maintenance plans based on the pipeline defect state descriptor, and performing probability distribution estimation on the state transition matrix to obtain a pipeline state transition probability; The pipeline state transition probability is constructed into a multi-objective decision function along with pipeline maintenance cost, maintenance result, and maintenance risk. An optimization solution is performed based on the multi-objective decision function to obtain a set of pipeline maintenance plans. The reliability score of each pipeline maintenance plan set is calculated, and the pipeline maintenance plan with the highest reliability score is selected to determine the optimal pipeline maintenance plan.

[0073] In one specific embodiment, in an underground pipeline network scenario, the defect signature map contains information on the distribution of structural damage along the pipeline. Underground pipelines are subject to long-term environmental erosion and internal media effects, resulting in varying degrees of damage. Adaptive sliding window sampling dynamically adjusts the sampling interval based on defect density, providing more detailed data collection in areas with significant defect variation, thereby accurately characterizing the spatial distribution characteristics and temporal evolution trends of underground pipeline defects. After acquiring the pipeline defect signature map, dynamic sampling is performed using an adaptive sliding window. The initial sliding window size is set to 30 days of data and dynamically adjusted based on the frequency of pipeline defect signature changes. Specifically, when the feature change rate for five consecutive sampling points exceeds a preset threshold of 10%, the window size is reduced to 80% of the current window size. When the feature change rate for ten consecutive sampling points falls below a preset threshold of 5%, the window size is increased to 120% of the current window size. For each sliding window, statistics of the pipeline features within the window are calculated, including the mean, standard deviation, kurtosis, and skewness. Trend characteristics, including the mean of the first-order differences, the mean of the second-order differences, and the trend slope, are also calculated to obtain a local feature sequence. In practical applications, for example, a 500-meter oil pipeline section was inspected. The average corrosion depth within a sliding window was 2.5 mm, with a standard deviation of 0.8 mm and a first-order difference mean of 0.15 mm / day, indicating accelerating corrosion in this section. Defect evolution in underground pipelines often exhibits nonlinear, multi-stage characteristics. Adaptive sliding window technology dynamically adjusts the sampling frequency based on the rate of defect change in different pipeline sections. This is crucial for accurately capturing both sudden defects (such as cracks) and progressive defects (such as slowly accumulating corrosion) in underground pipelines. By constructing a temporal pyramid structure, defect evolution patterns at different time scales can be simultaneously analyzed, such as short-term pipe wall stress changes, medium-term material degradation processes, and long-term structural stability trends, providing data support for the full lifecycle management of underground pipelines.

[0074] The local feature sequence is constructed as a temporal pyramid structure consisting of three levels. The first level retains all local features at the original sampling frequency; the second level is formed by aggregating features from every two adjacent windows, with the sampling frequency halved; the third level is formed by aggregating features from every four adjacent windows, with the sampling frequency reduced to 1 / 4 of the original. At each level, a feature extraction module is used to extract pipeline feature representations. The feature extraction module adopts a three-layer convolutional network structure with convolution kernel sizes of 3×3, 5×5, and 7×7, and the number of channels is 16, 32, and 64, respectively. The inter-level feature propagation network calculates the association weights of features at different levels through a self-attention mechanism. The weight threshold is set to 0.3, and associations below this threshold are pruned. Based on the co-evolution relationship, cross-level feature recursive updates are performed with an update step size of 0.05 and 50 iterations, ultimately resulting in a pipeline defect evolution sequence. For example, the analysis of pipeline wall thinning showed that the initial tiny cracks were identified at the first level, the crack expansion trend was captured at the second level, and the long-term evolution results were predicted at the third level, fully depicting the evolution process from tiny defects to potential leaks.

[0075] A counterfactual intervention strategy was constructed based on the pipeline defect evolution sequence. First, five control intervention scenarios were set, namely: an increase of 10°C in ambient temperature, an increase of 15% in ambient humidity, an increase of 20% in pipeline pressure, a change of ±1.5 pH in fluid pH, and an increase of 2 years in pipeline service life. For each intervention scenario, an isomorphic simulator was used to generate pipeline defect evolution data, and a deviation analysis was performed with the actual scenario data. The root mean square error (RMSE) was used for deviation calculation, and the deviation threshold was set to 15%. When the deviation in a certain intervention scenario exceeded the threshold, the corresponding factor was included in the influencing factor set. In a case analysis, the deviation caused by the change in ambient temperature was 17.3%, the deviation caused by the change in pipeline pressure was 19.8%, and the deviation caused by the change in ambient humidity was 12.5%. Therefore, the ambient temperature and pipeline pressure were determined to be the key influencing factors.

[0076] A pipeline defect state descriptor was constructed using a set of influencing factors. This descriptor consists of eight dimensions: defect depth, defect area, defect growth rate, ambient temperature, pipeline pressure, fluid corrosivity, pipeline material type, and age. Each dimension was discretized, with continuous values ​​divided into five levels. Based on the pipeline defect state descriptor, the state transition matrix for different pipeline repair scenarios was calculated. Repair scenarios included no repair, partial repair, partial replacement, and full segment replacement. For each repair scenario, an initial state transition matrix of 125×125 (the eighth power of five state levels) was constructed using historical data and expert knowledge. Monte Carlo simulation was performed with 10,000 samplings to estimate the probability distribution of the state transition matrix and obtain the pipeline state transition probabilities. For example, for the partial repair scenario, the probability of transitioning from state [3, 2, 4, 3, 4, 2, 1, 3] to state [2, 1, 3, 3, 4, 2, 1, 3] was 0.73, indicating that this repair scenario effectively reduced defect depth and area.

[0077] A multi-objective decision function was constructed by combining pipeline state transition probabilities with pipeline maintenance costs, maintenance outcomes, and maintenance risks. Maintenance costs include material costs, labor costs, and downtime losses; maintenance outcome assessments include expected service life extension and performance recovery; and maintenance risk assessments include construction risk and secondary damage risk. Weights of 0.4, 0.35, and 0.25 were assigned to these three dimensions, respectively. The multi-objective decision function employed a weighted summation approach. A simulated annealing algorithm was used to optimize the solution, with an initial temperature of 100, a cooling coefficient of 0.95, and 1000 iterations. A set of pipeline maintenance plans was obtained. A reliability score was calculated for each maintenance plan, taking into account three factors: historical success rate (weight 0.5), technological maturity (weight 0.3), and adaptability (weight 0.2). Ultimately, the pipeline maintenance plan with the highest reliability score was selected as the optimal plan. In a real-world case, partial repair scored 0.86, partial replacement scored 0.92, and full-section replacement scored 0.78, making partial replacement the optimal pipeline maintenance plan.

[0078] In an optional embodiment, a counterfactual intervention strategy is constructed based on the pipeline defect evolution sequence to generate a control experiment scenario, and a deviation analysis is performed between the control experiment scenario and the actual scenario to obtain an influencing factor set including: Calculate the time series nodes in the pipeline defect evolution sequence where the feature change rate exceeds the preset change rate threshold, adjust the intensity of the pipeline defect status characteristics at the time series nodes, and generate counterfactual intervention strategies; Execute pipeline defect state simulation according to the counterfactual intervention strategy and pipeline constraint conditions, perform evolution calculation on the pipeline defect state under each intervention intensity, and generate multiple groups of control experimental scenarios; Calculating feature deviations between the control experimental scenario and the actual pipeline scenario, performing cluster analysis on the feature deviations to obtain feature deviation clusters, calculating the variance contribution rate of the pipeline features in each feature deviation cluster, and selecting features whose variance contribution rate exceeds a preset variance threshold as dominant factors; Calculating a time series correlation coefficient based on the dominant factors, determining a contribution weight of each dominant factor according to the time series correlation coefficient, and selecting a dominant factor whose contribution weight is greater than a preset contribution weight threshold; The correlation degree between the dominant factors is calculated, the dominant factors are ranked and sorted, and a set of influencing factors is generated.

[0079] In one specific embodiment, a pipeline defect evolution sequence typically contains defect status feature data at multiple time points. In underground pipeline network monitoring applications, the defect evolution sequence reflects the accumulation of pipeline damage over time. During long-term operation, underground pipeline networks are affected by various internal and external factors, and their defect status exhibits certain evolution patterns. By recording the temporal changes of key parameters (such as wall thickness reduction rate and crack propagation rate), a complete defect evolution sequence is constructed. Counterfactual intervention analysis can identify key factors influencing the development of underground pipeline defects, such as ambient humidity and internal pressure changes, providing a scientific basis for pipeline maintenance decisions. The system reads a pipeline defect evolution sequence dataset containing various pipeline status parameters at different time points, such as wall thickness, corrosion depth, and crack length. For the evolution sequence data for a particular pipeline, the system calculates the rate of change of each feature between adjacent time points. For example, for the corrosion depth feature, if the value at time 1 is 2.5 mm and the value at time 2 is 3.2 mm, the rate of change is 28%. The system compares the rate of change of each feature with a preset rate of change threshold (typically set to 15%) and marks the time points where the rate of change exceeds the threshold. In actual applications, the wall thickness thinning rate of a certain pipeline section reached 17% and 22% on the 45th and 78th days respectively, exceeding the preset threshold and was therefore marked as a critical timing node.

[0080] For marked time series nodes, the system designs counterfactual intervention strategies to adjust the intensity of pipeline defect status characteristics. Intensity adjustments typically use three adjustment factors: 0.8x, 1.0x, and 1.2x, representing weakening, maintaining, and enhancing intervention, respectively. For example, for a node with a 22% wall thickness reduction rate, three counterfactual intervention scenarios are generated: 17.6% (0.8 x 22%), 22%, and 26.4% (1.2 x 22%).

[0081] Defect state simulations were performed based on counterfactual intervention strategies and pipeline constraints. Pipeline constraints included material properties (e.g., elastic modulus of 210 GPa, Poisson's ratio of 0.3), operating parameters (e.g., internal pressure of 8 MPa, temperature of 65°C), and geometric constraints. Finite element analysis was used to calculate the evolution of pipeline defect states under each intervention intensity. For example, for wall thickness reduction rates of 17.6%, 22%, and 26.4%, defect development trends over the next 30 days were calculated, yielding data for three control experimental scenarios. In actual calculations, a pipeline section exhibited a maximum deformation of 2.1 mm after 30 days under an intervention rate of 26.4%, compared to only 1.7 mm under the standard scenario.

[0082] Calculate the feature deviations between the control experimental scenario and the actual pipeline scenario. This feature deviation calculation involves multiple feature dimensions, such as stress distribution, deformation, and energy release rate. For each feature, the deviation value is calculated and stored in a deviation matrix. For example, for the deformation feature, the intervention scenario value is 2.1 mm, while the actual scenario value is 1.7 mm, resulting in a deviation of 0.4 mm and a deviation rate of approximately 23.5%.

[0083] Cluster analysis is performed on the feature deviations, using the K-means clustering algorithm to group similar deviation patterns together. In practical applications, the number of clusters is typically set to 3-5. After clustering, the variance contribution of each pipeline feature is calculated for each feature deviation cluster. The variance contribution calculation is based on the principal component analysis method, which determines the contribution of each feature to the overall variance. For example, in one analysis, the variance contributions of wall thickness thinning, crack length, and material hardness were 42%, 31%, and 15%, respectively. Features with variance contributions exceeding a preset variance threshold (typically 20%) are selected as dominant factors. In this example, wall thickness thinning and crack length were identified as dominant factors.

[0084] For the identified dominant factors, their temporal correlation coefficients are calculated, and a lagged correlation analysis method is used to assess the degree of correlation between each factor over time. The temporal correlation calculation is based on the changing trends of each dominant factor in the historical data. For example, the temporal correlation coefficient between wall thinning and crack length is 0.76, indicating that the two have a strong temporal correlation. The contribution weight of each dominant factor is determined based on the temporal correlation coefficient, usually obtained through normalization. In this example, the contribution weights of wall thinning and crack length are 0.58 and 0.42, respectively. The dominant factors with contribution weights greater than the preset contribution weight threshold (usually 0.3) are selected; in this example, both factors are retained.

[0085] The correlation between the dominant factors is calculated to construct a factor correlation network. This correlation is based on conditional mutual information and partial correlation analysis. For example, the analysis shows a correlation coefficient of 0.65 between wall thickness reduction and crack length, indicating a moderate mutual influence. The system ranks the dominant factors based on their correlation to generate a final set of influencing factors. In this example, the final set of influencing factors is: wall thickness reduction (a primary factor with a weight of 0.58) and crack length (a secondary factor with a weight of 0.42).

[0086] Through the above method, the key influencing factors in the evolution of pipeline defects were successfully identified, providing data support for subsequent preventive maintenance and risk management.

[0087] The deep learning-based intelligent identification and early warning system for underground pipeline network defects in an embodiment of the present invention includes: The first unit is used to obtain multi-source detection data of the underground pipeline network using a pipeline detection robot and perform preprocessing to obtain initial feature data; The second unit is used to map the initial feature data to the manifold space through the feature mapping module to obtain manifold feature data, construct a similarity matrix in the manifold space to divide the feature subspace, and perform geometric difference weighted fusion on the feature subspace to obtain the pipeline defect representation vector; The third unit is used to construct the pipeline defect representation vector into an association matrix, calculate the spatial distance association value and mechanical action strength between pipeline defect features to obtain a connection relationship matrix, perform iterative propagation to obtain a pipeline state sequence, construct a pipeline state evolution diagram and optimize the propagation direction to generate a pipeline defect feature map; The fourth unit is used to dynamically sample the pipeline defect feature map using an adaptive sliding window to obtain the pipeline defect evolution sequence, determine the influencing factor set based on the counterfactual intervention strategy analysis, calculate the pipeline state transition probability based on the influencing factor set, and determine the optimal pipeline maintenance plan based on the multi-objective decision function; The fifth unit is used to generate pipeline network defect warning information when the pipeline state transition probability exceeds a preset probability threshold; The sixth unit is used to send pipeline network defect warning information and the optimal pipeline maintenance plan to the monitoring terminal.

[0088] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0089] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0090] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The intelligent identification and early warning method of underground pipeline network defects based on deep learning is characterized by: include: Use pipeline inspection robots to obtain multi-source inspection data of underground pipeline networks and perform preprocessing to obtain initial feature data; The initial feature data is mapped to the manifold space through the feature mapping module to obtain the manifold feature data. A similarity matrix is ​​constructed in the manifold space to divide the feature subspace. The feature subspace is weightedly fused by geometric differences to obtain the pipeline defect representation vector. The pipeline defect characterization vector is constructed as an association matrix. The spatial distance association value and mechanical interaction strength between pipeline defect features are calculated to obtain a connection relationship matrix. Iterative propagation is performed to obtain a pipeline state sequence. A pipeline state evolution diagram is constructed and the propagation direction is optimized to generate a pipeline defect feature map. An adaptive sliding window is used to dynamically sample the pipeline defect feature map to obtain the pipeline defect evolution sequence. The influencing factor set is determined based on the counterfactual intervention strategy analysis. The pipeline state transition probability is calculated based on the influencing factor set, and the optimal pipeline maintenance plan is determined based on the multi-objective decision function. When the pipeline state transition probability exceeds the preset probability threshold, a pipeline network defect warning information is generated; Send pipeline network defect warning information and optimal pipeline maintenance plan to the monitoring terminal.

2. The method according to claim 1, characterized in that The initial feature data is mapped to the manifold space through the feature mapping module to obtain the manifold feature data. A similarity matrix is ​​constructed in the manifold space to divide the feature subspace. The geometric difference weighted fusion of the feature subspace is performed to obtain the pipeline defect representation vector including: Inputting the initial feature data into a feature mapping module, the feature mapping module calculating the neighbor relationship of sample points in the initial feature data to construct an adjacency graph, calculating the similarity relationship between the sample points in the adjacency graph to obtain a weight matrix, constructing a diagonal matrix based on the weight matrix and calculating a difference matrix, obtaining a projection matrix based on the difference matrix, and using the projection matrix to map the initial feature data to a manifold space to obtain manifold feature data; Constructing a local tangent plane in the manifold space where the manifold feature data is located, converting the global nonlinear space of the underground pipe network defects into a local linear space, calculating the shortest path between sample points in the manifold feature data in the local linear space as a geodesic distance, and converting it into a similarity score, and constructing a similarity matrix based on the similarity score; Based on the similarity matrix, the manifold feature data is divided into multiple feature subspaces. The local distribution characteristics of pipeline defects are calculated in each feature subspace to obtain a feature distribution matrix. A metric tensor is constructed based on the feature distribution matrix. The geometric difference values ​​between the multiple feature subspaces are calculated according to the metric tensor, and the geometric difference values ​​are converted into corresponding feature subspace weight coefficients. The feature distribution matrices of the multiple feature subspaces are multiplied by the corresponding weight coefficients and superimposed to obtain a defect representation vector.

3. The method according to claim 2, characterized in that Based on the similarity matrix, the manifold feature data is divided into multiple feature subspaces including: In the local linear space where the manifold feature data is located, a search radius is constructed for each sample point of the underground pipe network defect state, the number of neighboring sample points within the search radius is counted to obtain an initial local density value of the sample point, and the density ratio of the sample point to the corresponding neighboring sample point is calculated based on the initial local density value to obtain a local density weight; Multiplying the similarity value between each sample point in the similarity matrix by the corresponding local density weight to determine a density weighting coefficient, and using the density weighting coefficient to update the similarity matrix to obtain a density-aware similarity matrix; The manifold feature data is grouped and divided according to the similarity relationship between the sample points in the density-aware similarity matrix. When the density difference between any two sample points is less than a preset density threshold and the corresponding similarity relationship in the density-aware similarity matrix is ​​greater than the preset similarity threshold, the two sample points are divided into the same feature subspace to obtain multiple feature subspaces.

4. The method according to claim 1, wherein The pipeline defect characterization vector is constructed as an association matrix. The spatial distance association value and mechanical interaction strength between pipeline defect features are calculated to obtain a connection relationship matrix. Iterative propagation is performed to obtain a pipeline state sequence. A pipeline state evolution diagram is constructed and the propagation direction is optimized. The pipeline defect feature map is generated, including: The pipeline defect characterization vector is constructed as an association matrix, the spatial distance association value between each group of pipeline defect features is calculated based on the association matrix, and normalized to obtain an initial connection matrix; Calculating the mechanical interaction strength between pipeline defect features based on the spatial distance correlation value in the initial connection matrix, and converting the mechanical interaction strength into a connection weight to obtain a connection relationship matrix; Constructing a pipeline state transfer matrix based on the connection relationship matrix, performing iterative propagation on pipeline defect features, fusing state information of adjacent pipeline features in each iteration process to update the current pipeline state, and obtaining a pipeline state sequence; Constructing a state evolution graph based on the pipeline state sequence, calculating the transition probability between nodes in the state evolution graph to obtain a propagation direction, and filtering the propagation direction according to a preset pipeline structure constraint condition to obtain an optimized direction; The pipeline status is layered based on the optimization direction, a hierarchical pipeline association structure is constructed, the importance index of the feature nodes in each layer is calculated, the feature nodes are screened based on the importance index, and a pipeline defect feature map is generated.

5. The method according to claim 4, characterized in that Based on the connection relationship matrix, a pipeline state transfer matrix is ​​constructed, and iterative propagation is performed on the pipeline defect features. During each iteration, the state information of the adjacent pipeline features is integrated to update the current pipeline state. The pipeline state sequence obtained includes: Calculating the eigenvalue decomposition of the connection relationship matrix to obtain an eigenvector matrix, using the eigenvector matrix to construct frequency response functions of different scales, and performing a convolution operation on the connection relationship matrix and the corresponding frequency response function to obtain pipeline state transfer matrices of multiple scales; Pipeline defect feature propagation is performed on the pipeline state transfer matrix of each scale to obtain multiple sets of pipeline state sequences; the degree of difference between adjacent states in each set of pipeline state sequences is calculated to obtain a propagation response value, and attention weights are determined based on the propagation response values; the multiple sets of pipeline state sequences are fused using the attention weights to obtain a fused pipeline state; The deviation between the fused pipeline state and the initial pipeline state is calculated to obtain a state residual, which is fed back to the pipeline defect feature propagation process. The attention weight is dynamically adjusted based on the state residual to adaptively adjust the intensity of pipeline defect feature propagation. The feature propagation update is repeatedly performed until the state residual is less than a preset residual threshold, thereby obtaining an optimized pipeline state sequence.

6. The method according to claim 1, characterized in that The pipeline defect feature map is dynamically sampled using an adaptive sliding window to obtain the pipeline defect evolution sequence. The influencing factor set is determined based on the counterfactual intervention strategy analysis. The pipeline state transition probability is calculated based on the influencing factor set. The optimal pipeline maintenance plan is determined based on the multi-objective decision function, including: An adaptive sliding window is used to dynamically sample the pipeline defect feature map. The size of the sliding window is dynamically adjusted according to the frequency of change of the pipeline defect feature. The statistics and change trend characteristics of the pipeline features within the sliding window are calculated to obtain a local feature sequence. The local feature sequence is constructed into a temporal pyramid structure, pipeline feature representation is extracted at each level of the temporal pyramid structure, an inter-level feature propagation network is constructed to calculate the co-evolution relationship between pipeline features, and cross-level feature recursive update is performed based on the co-evolution relationship to obtain a pipeline defect evolution sequence; Constructing a counterfactual intervention strategy based on the pipeline defect evolution sequence to generate a control experiment scenario, and performing a deviation analysis between the control experiment scenario and the actual scenario to obtain a set of influencing factors; Constructing a pipeline defect state descriptor using the influencing factor set, calculating a state transition matrix under different pipeline maintenance plans based on the pipeline defect state descriptor, and performing probability distribution estimation on the state transition matrix to obtain a pipeline state transition probability; The pipeline state transition probability is constructed into a multi-objective decision function along with pipeline maintenance cost, maintenance result, and maintenance risk. An optimization solution is performed based on the multi-objective decision function to obtain a set of pipeline maintenance plans. The reliability score of each pipeline maintenance plan set is calculated, and the pipeline maintenance plan with the highest reliability score is selected to determine the optimal pipeline maintenance plan.

7. The method according to claim 6, characterized in that Based on the pipeline defect evolution sequence, a counterfactual intervention strategy is constructed to generate a control experiment scenario. Deviation analysis is performed between the control experiment scenario and the actual scenario to obtain a set of influencing factors including: Calculate the time series nodes in the pipeline defect evolution sequence where the feature change rate exceeds the preset change rate threshold, adjust the intensity of the pipeline defect status characteristics at the time series nodes, and generate counterfactual intervention strategies; Execute pipeline defect state simulation according to the counterfactual intervention strategy and pipeline constraint conditions, perform evolution calculation on the pipeline defect state under each intervention intensity, and generate multiple groups of control experimental scenarios; Calculating feature deviations between the control experimental scenario and the actual pipeline scenario, performing cluster analysis on the feature deviations to obtain feature deviation clusters, calculating the variance contribution rate of the pipeline features in each feature deviation cluster, and selecting features whose variance contribution rate exceeds a preset variance threshold as dominant factors; Calculating a time series correlation coefficient based on the dominant factors, determining a contribution weight of each dominant factor according to the time series correlation coefficient, and selecting a dominant factor whose contribution weight is greater than a preset contribution weight threshold; The correlation degree between the dominant factors is calculated, the dominant factors are ranked and sorted, and a set of influencing factors is generated.

8. An intelligent identification and early warning system for underground pipe network defects based on deep learning, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain multi-source detection data of the underground pipeline network using a pipeline detection robot and perform preprocessing to obtain initial feature data; The second unit is used to map the initial feature data to the manifold space through the feature mapping module to obtain manifold feature data, construct a similarity matrix in the manifold space to divide the feature subspace, and perform geometric difference weighted fusion on the feature subspace to obtain the pipeline defect representation vector; The third unit is used to construct the pipeline defect representation vector into an association matrix, calculate the spatial distance association value and mechanical action strength between pipeline defect features to obtain a connection relationship matrix, perform iterative propagation to obtain a pipeline state sequence, construct a pipeline state evolution diagram and optimize the propagation direction to generate a pipeline defect feature map; The fourth unit is used to dynamically sample the pipeline defect feature map using an adaptive sliding window to obtain the pipeline defect evolution sequence, determine the influencing factor set based on the counterfactual intervention strategy analysis, calculate the pipeline state transition probability based on the influencing factor set, and determine the optimal pipeline maintenance plan based on the multi-objective decision function; The fifth unit is used to generate pipeline network defect warning information when the pipeline state transition probability exceeds a preset probability threshold; The sixth unit is used to send pipeline network defect warning information and the optimal pipeline maintenance plan to the monitoring terminal.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images