Intelligent identification and early warning method and system for underground pipe network defects based on deep learning

Abstract

The application provides a kind of underground pipe network defect intelligent identification and early warning method and system based on deep learning, it is related to pipe network detection technical field, including using pipeline detection robot to obtain multi-source data and pre-processing;Through feature mapping to manifold space, construct similarity matrix and carry out geometric difference weighted fusion;Correlation matrix is constructed to calculate spatial distance and mechanical strength, and defect feature atlas is generated by iterative propagation;Adaptive sliding window is used to obtain evolution sequence, and influence factors are determined based on counterfactual intervention, and the optimal maintenance scheme is determined by calculating state transition probability.The application improves the pipe network defect identification accuracy, realizes the intelligentization of pipe state prediction and maintenance decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pipe network detection, and in particular to an underground pipe network defect intelligent identification and early warning method and system based on deep learning. BACKGROUND

[0002] Underground pipe networks are an important part of urban infrastructure, bearing functions such as water supply, drainage, gas supply, and power supply, and their safe operation is crucial to the normal operation of the city. With the development of cities and the aging of pipe networks, underground pipe network defect problems are increasingly prominent, such as increasing damage to pipe wall corrosion, cracks, and interface misplacement, which seriously threaten the safety of the city and the lives of residents. Traditional pipe network detection mainly relies on manual inspection and simple mechanical equipment, which is inefficient and difficult to fully identify hidden defects. In recent years, pipe detection robot technology has gradually been applied to the field of underground pipe network detection, which can collect internal pipe images, acoustic, laser, and other multi-source data by carrying various sensors, providing technical support for pipe condition assessment.

[0003] With the development of deep learning technology, pipe defect detection methods based on image recognition have made some progress. Existing technologies usually use convolutional neural networks and other algorithms to analyze pipe images, enabling the identification of surface defects such as cracks and corrosion. Meanwhile, some research has also attempted to combine acoustic detection and laser scanning methods to build a multi-modal data fusion framework, improving the accuracy and comprehensiveness of pipe defect detection.

[0004] Existing underground pipe network 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 dynamic analysis capabilities for the evolution of pipe defects, often only focusing on static features of defects, and are unable to accurately predict defect development trends and potential risks. Existing early warning methods lack a systematic pipe repair decision support mechanism, making it difficult to develop scientific and reasonable repair plans based on comprehensive information such as defect severity, development trends, and influencing factors, leading to unreasonable allocation of repair resources and difficulty in achieving precision and intelligent pipe network maintenance. SUMMARY

[0005] The present application provides an underground pipe network defect intelligent identification and early warning method and system based on deep learning, which can solve the problems in the prior art.

[0006] In a first aspect of the present application, an underground pipe network defect intelligent identification and early warning method based on deep learning is provided, comprising:

[0007] A pipe detection robot is used to obtain multi-source detection data of the underground pipe network and perform preprocessing to obtain initial feature data.

[0008] mapping the initial feature data to a manifold space through a feature mapping module to obtain manifold feature data, constructing a similarity matrix in the manifold space to divide a feature subspace, and performing geometric difference weighted fusion on the feature subspace to obtain a pipeline defect representation vector;

[0009] constructing the pipeline defect representation vector into a correlation matrix, calculating spatial distance correlation values and mechanical action intensities between pipeline defect features to obtain a connection relationship matrix, performing iterative propagation to obtain a pipeline state sequence, constructing a pipeline state evolution graph and optimizing a propagation direction, and generating a pipeline defect feature map;

[0010] performing dynamic sampling on the pipeline defect feature map using an adaptive sliding window to obtain a pipeline defect evolution sequence, determining a set of influence factors based on a counterfactual intervention strategy analysis, calculating a pipeline state transition probability according to the set of influence factors, and determining an optimal pipeline maintenance scheme based on a multi-objective decision function;

[0011] when the pipeline state transition probability exceeds a preset probability threshold, generating pipeline network defect early warning information;

[0012] sending the pipeline network defect early warning information and the optimal pipeline maintenance scheme to a monitoring terminal.

[0013] In an alternative embodiment, mapping the initial feature data to a manifold space through a feature mapping module to obtain manifold feature data, constructing a similarity matrix in the manifold space to divide a feature subspace, and performing geometric difference weighted fusion on the feature subspace to obtain a pipeline defect representation vector includes:

[0014] inputting the initial feature data into a feature mapping module, the feature mapping module calculating the near-neighbor relationship of sample points in the initial feature data to construct an adjacency graph, calculating the similarity relationship between sample points in the adjacency graph to obtain a weight matrix, constructing a diagonal matrix according to the weight matrix and calculating a difference matrix, obtaining a projection matrix based on the difference matrix, and mapping the initial feature data to a manifold space using the projection matrix to obtain manifold feature data;

[0015] constructing a local tangent plane in the manifold space in which the manifold feature data is located, converting the global nonlinear space of the underground pipeline network defect into a local linear space, calculating the shortest path between sample points in the manifold feature data as a geodesic distance in the local linear space, and converting it into a similarity score, and constructing a similarity matrix according to the similarity score;

[0016] The manifold feature data is divided into a plurality of feature subspaces based on the similarity matrix, a feature distribution matrix is calculated based on a local distribution feature of the pipeline defect in each feature subspace, a metric tensor is constructed based on the feature distribution matrix, a geometric difference value between the plurality of feature subspaces is calculated according to the metric tensor, the geometric difference value is converted into a corresponding feature subspace weight coefficient, and the feature distribution matrix of the plurality of feature subspaces is multiplied by the corresponding weight coefficient and superimposed to obtain a defect feature vector.

[0017] In an optional embodiment, dividing the manifold feature data into a plurality of feature subspaces based on the similarity matrix comprises:

[0018] In a local linear space where the manifold feature data is located, a search radius is constructed for each sample point of the underground pipeline network defect state, a local density initial value of the sample point is obtained by counting the number of neighbor sample points in the search radius, and a density ratio of the sample point and the corresponding neighbor sample point is calculated according to the local density initial value to obtain a local density weight.

[0019] A density weighting coefficient is determined by multiplying the similarity value between each sample point in the similarity matrix and the corresponding local density weight, and the similarity matrix is updated by using the density weighting coefficient to obtain a density-aware similarity matrix.

[0020] 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 value 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 a preset similarity threshold, the two sample points are divided into the same feature subspace, and a plurality of feature subspaces are obtained.

[0021] In an optional embodiment, the pipeline defect feature vector is constructed as a connection matrix, a spatial distance correlation value and a mechanical action strength between pipeline defect features are calculated to obtain a connection relationship matrix, an iteration propagation is performed to obtain a pipeline state sequence, a pipeline state evolution graph is constructed and a propagation direction is optimized, and a pipeline defect feature map is generated, comprising:

[0022] The pipeline defect feature vector is constructed as a connection matrix, a spatial distance correlation value between each group of pipeline defect features is calculated based on the connection matrix, and normalization processing is performed to obtain an initial connection matrix;

[0023] A mechanical action strength between pipeline defect features is calculated according to the spatial distance correlation value in the initial connection matrix, the mechanical action strength is converted into a connection weight to obtain a connection relationship matrix;

[0024] construct a pipeline state transition matrix based on the connection relationship matrix, perform iterative propagation on the pipeline defect features, and update the current pipeline state by fusing the state information of adjacent pipeline features in each iteration process to obtain a pipeline state sequence;

[0025] construct a state evolution graph based on the pipeline state sequence, calculate the transition probability between nodes in the state evolution graph to obtain a propagation direction, filter the propagation direction according to a preset pipeline structure constraint condition to obtain an optimized direction;

[0026] based on the optimized direction, layer the pipeline state, construct a hierarchical pipeline correlation structure, calculate an importance index of each feature node, filter the feature nodes based on the importance index, and generate a pipeline defect feature map.

[0027] In an alternative embodiment, based on the connection relationship matrix, a pipeline state transition matrix is constructed, and iterative propagation is performed on the pipeline defect features. In each iteration process, the current pipeline state is updated by fusing the state information of adjacent pipeline features to obtain a pipeline state sequence, which includes:

[0028] calculate the eigenvalue decomposition of the connection relationship matrix to obtain a feature vector matrix, construct different scale frequency response functions using the feature vector matrix, and perform convolution operation on the connection relationship matrix and the corresponding frequency response function to obtain pipeline state transition matrices of multiple scales;

[0029] perform pipeline defect feature propagation on each scale of the pipeline state transition matrix to obtain multiple groups of pipeline state sequences; calculate the difference degree of adjacent states in each group of pipeline state sequences to obtain a propagation response value, determine an attention weight based on the propagation response value; and fuse multiple groups of pipeline state sequences using the attention weight to obtain a fused pipeline state;

[0030] calculate the deviation between the fused pipeline state and the initial pipeline state to obtain a state residual error, feed the state residual error back to the pipeline defect feature propagation process; dynamically adjust the attention weight based on the state residual error to adaptively adjust the intensity of pipeline defect feature propagation; repeat the feature propagation update until the state residual error is less than a preset residual threshold to obtain an optimized pipeline state sequence.

[0031] In an alternative embodiment, the pipeline defect feature map is dynamically sampled using an adaptive sliding window to obtain a pipeline defect evolution sequence, an influence factor set is determined based on a counterfactual intervention strategy analysis, a pipeline state transition probability is calculated according to the influence factor set, and an optimal pipeline maintenance scheme is determined based on a multi-objective decision function, which includes:

[0032] The pipeline defect feature atlas is dynamically sampled by using an adaptive sliding window, the size of the sliding window is dynamically adjusted according to the change frequency of the pipeline defect feature, the statistical quantity and the change trend feature of the pipeline feature in the sliding window are calculated to obtain a local feature sequence;

[0033] The local feature sequence is constructed as a time sequence pyramid structure, a pipeline feature representation is extracted at each level of the time sequence pyramid structure, a feature propagation network between levels is constructed to calculate the cooperative evolution relationship between pipeline features, and a cross-level feature recursive update is performed based on the cooperative evolution relationship to obtain a pipeline defect evolution sequence;

[0034] Based on the pipeline defect evolution sequence, a counterfactual intervention strategy is generated to generate a control experiment scene, and a deviation analysis is performed on the control experiment scene and the actual scene to obtain an influence factor set;

[0035] A pipeline defect state descriptor is constructed using the influence factor set, a state transition matrix under different pipeline maintenance schemes is calculated based on the pipeline defect state descriptor, and a pipeline state transition probability is obtained by performing probability distribution estimation on the state transition matrix;

[0036] The pipeline state transition probability, pipeline maintenance cost, maintenance result, and maintenance risk are constructed as a multi-objective decision function, an optimal pipeline maintenance scheme is determined based on the multi-objective decision function, a reliability score of each pipeline maintenance scheme is calculated, and the pipeline maintenance scheme with the highest reliability score is selected.

[0037] In an optional embodiment, based on the pipeline defect evolution sequence, a counterfactual intervention strategy is generated to generate a control experiment scene, and a deviation analysis is performed on the control experiment scene and the actual scene to obtain an influence factor set, which includes:

[0038] A time sequence node with a feature change rate exceeding a preset change rate threshold in the pipeline defect evolution sequence is calculated, the intensity of the pipeline defect state feature at the time sequence node is adjusted, and a counterfactual intervention strategy is generated;

[0039] According to the counterfactual intervention strategy and the pipeline constraint condition, a pipeline defect state simulation is performed, the pipeline defect state under each intervention intensity is evolved and calculated, and a plurality of control experiment scenes are generated;

[0040] The feature deviation between the control experiment scene and the actual pipeline scene is calculated, a cluster analysis is performed on the feature deviation to obtain a feature deviation cluster, the variance contribution rate of the pipeline feature in each feature deviation cluster is calculated, and a dominant factor is selected as the feature with a variance contribution rate exceeding a preset variance threshold;

[0041] calculate a time sequence correlation coefficient based on the dominant factors, determine a contribution weight of each dominant factor according to the time sequence correlation coefficient, and select a dominant factor whose contribution weight is greater than a preset contribution weight threshold value;

[0042] calculate a correlation degree between the dominant factors, rank the dominant factors, and generate an influence factor set.

[0043] In a second aspect of the embodiment, a deep learning-based underground pipe network defect intelligent identification and early warning system is provided, which comprises:

[0044] A first unit is configured to acquire multi-source detection data of an underground pipe network by using a pipe detection robot, and perform preprocessing to obtain initial feature data;

[0045] A second unit is configured to map the initial feature data to a manifold space by a feature mapping module to obtain manifold feature data, construct a similarity matrix in the manifold space to divide a feature subspace, and perform geometric difference weighted fusion on the feature subspace to obtain a pipe defect representation vector;

[0046] A third unit is configured to construct the pipe defect representation vector into a correlation matrix, calculate a spatial distance correlation value and a mechanical action strength between pipe defect features to obtain a connection relationship matrix, perform iterative propagation to obtain a pipe state sequence, construct a pipe state evolution graph and optimize a propagation direction, and generate a pipe defect feature map;

[0047] A fourth unit is configured to perform dynamic sampling on the pipe defect feature map by using an adaptive sliding window to obtain a pipe defect evolution sequence, determine an influence factor set based on a counterfactual intervention strategy analysis, calculate a pipe state transition probability according to the influence factor set, and determine an optimal pipe maintenance scheme based on a multi-objective decision function;

[0048] A fifth unit is configured to generate pipe network defect early warning information when the pipe state transition probability exceeds a preset probability threshold value.

[0049] A sixth unit is configured to send the pipe network defect early warning information and the optimal pipe maintenance scheme to a monitoring terminal.

[0050] In a third aspect of the embodiment, an electronic device is provided, which comprises:

[0051] a processor;

[0052] a memory for storing processor-executable instructions;

[0053] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0054] In a fourth aspect of the embodiment of the present application, a computer readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0055] In the embodiment of the present application, by 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; by using iterative propagation and pipeline state evolution graph construction technology, the correlation between pipeline defect features is deeply mined, and a complete pipeline defect feature map is established, so that the system can fully grasp the pipeline state evolution law and provide reliable data support for early warning decision; combined with counterfactual intervention strategy and multi-objective decision function, intelligent generation of pipeline defect early warning and maintenance scheme is realized, which optimizes the allocation of maintenance resources while ensuring the safe operation of the pipeline network, greatly improves the efficiency and economic benefit of underground pipeline network maintenance, and has significant engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 FIG. 1 is a flowchart of a pipeline defect intelligent recognition and early warning method based on deep learning of an embodiment of the present application.

[0057] Figure 2 FIG. 4 is a pipeline defect feature propagation and adaptive adjustment flowchart. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0059] The technical scheme of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0060] Figure 1 FIG. 1 is a flowchart of a pipeline defect intelligent recognition and early warning method based on deep learning of an embodiment of the present application, as shown in FIG. 1, the method comprises: Figure 1

[0061] The pipeline detection robot is used to obtain multi-source detection data of the underground pipeline network and perform preprocessing to obtain initial feature data;

[0062] ​The initial feature data is mapped to a manifold space by a feature mapping module to obtain manifold feature data, a similarity matrix is constructed in the manifold space to divide a feature subspace, and a pipe defect representation vector is obtained by geometric difference weighted fusion of the feature subspace;

[0063] The pipe defect representation vector is constructed into a correlation matrix, a spatial distance correlation value and a mechanical action strength between pipe defect features are calculated to obtain a connection relationship matrix, an iteration propagation is performed to obtain a pipe state sequence, a pipe state evolution graph is constructed and a propagation direction is optimized, and a pipe defect feature map is generated;

[0064] The pipe defect feature map is dynamically sampled by an adaptive sliding window to obtain a pipe defect evolution sequence, an influence factor set is determined based on a counterfactual intervention strategy analysis, a pipe state transition probability is calculated according to the influence factor set, and an optimal pipe maintenance scheme is determined based on a multi-objective decision function;

[0065] When the pipe state transition probability exceeds a preset probability threshold, a pipe network defect early warning information is generated;

[0066] The pipe network defect early warning information and the optimal pipe maintenance scheme are sent to a monitoring terminal.

[0067] In a specific embodiment, the multi-source detection data acquired by the pipe detection robot mainly comes from various sensing devices. The pipe detection robot is equipped with high-definition cameras, infrared thermal imagers and other visual sensors, which can capture visible light images and thermal imaging images of the inner wall of the pipe. These image data are used to identify visual defects such as surface cracks, corrosion spots and interface misalignment. The robot is also equipped with ultrasonic sensors to measure the thickness of the pipe wall and the integrity of the internal structure, which are used to detect non-surface defects such as pipe wall thinning, internal cavities and material degradation.

[0068] The detection robot also has a laser scanning system that acquires precise geometric dimensions and deformation data inside the pipe through laser ranging and three-dimensional imaging technology, which is used to detect pipe deformation, deviation and geometric anomalies. Acoustic sensors include acoustic emission and sound wave propagation detection devices that collect acoustic signals in the pipe structure, which are used to identify micro-crack 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 electrical conductivity and magnetic anomalies, thereby identifying corrosion weak areas and material fatigue.

[0069] In order to comprehensively evaluate the environmental conditions of the pipe, the detection robot is equipped with environmental parameter sensors to measure parameters such as temperature, humidity, pH value and gas concentration inside the pipe, providing environmental data for defect formation mechanism analysis. The inertial measurement unit of the robot records accurate position and attitude information for defect positioning and spatial mapping.

[0070] The specific acquired data includes a high-definition image sequence of the inner wall of the pipeline and thermal imaging data; pipeline wall thickness measurement data with an accuracy of ±0.1 mm and material sound velocity characteristics; three-dimensional point cloud data and geometric profile information of the inside of the pipeline with an accuracy of ±0.5 mm; acoustic waveform data, spectral characteristics, and acoustic emission event records; electromagnetic characteristic parameters of the metal pipeline and abnormal area distribution map; environmental parameter time series data covering a temperature range of -10°C to 100°C, a humidity range of 0 to 100%, and a pH value range of 0 to 14; and three-dimensional coordinate and motion trajectory data of the detection robot in the pipeline.

[0071] The multi-source data is acquired in real time through data acquisition of the pipeline detection robot and is subjected to spatio-temporal synchronization processing to form a complete pipeline state description data set, thereby providing a comprehensive data basis for subsequent defect identification and early warning. Through fusion analysis of such multi-source heterogeneous data, the system can more comprehensively and accurately identify pipeline defects and improve the reliability of safe operation of underground pipe networks.

[0072] In an optional implementation, the initial feature data is mapped to a manifold space through a feature mapping module to obtain manifold feature data, a similarity matrix is constructed in the manifold space to divide feature subspaces, and a pipeline defect representation vector is obtained by geometric difference weighted fusion of the feature subspaces.

[0073] The initial feature data is input into a feature mapping module, the feature mapping module calculates the near-neighbor relationship of sample points in the initial feature data to construct an adjacency graph, calculates the similarity relationship between the sample points in the adjacency graph to obtain a weight matrix, constructs a diagonal matrix according to the weight matrix and calculates a difference matrix, obtains a projection matrix based on the difference matrix, and maps the initial feature data to a manifold space by using the projection matrix to obtain manifold feature data;

[0074] A local tangent plane is constructed in the manifold space where the manifold feature data is located, a global nonlinear space of underground pipe network defects is converted into a local linear space, a shortest path between sample points in the manifold feature data is calculated as a geodesic distance in the local linear space and is converted into a similarity score, and a similarity matrix is constructed according to the similarity score;

[0075] The manifold feature data is divided into multiple feature subspaces based on the similarity matrix, 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, geometric difference values between the multiple feature subspaces are calculated according to the metric tensor, the geometric difference values are converted into corresponding feature subspace weight coefficients, and the feature distribution matrix of the multiple feature subspaces is multiplied by the corresponding weight coefficients respectively and is superimposed to obtain a defect representation vector.

[0076] In one specific implementation, in the underground pipe network defect identification scenario, the initial feature data usually contains multi-source heterogeneous data such as pipe inner wall image features, acoustic detection features, and size measurement features. For example, for a section of urban underground pipe network, the image features collected by the detection robot can include surface texture, color change, and morphological features, etc., for identifying corrosion and cracks; acoustic features include echo intensity and spectral distribution, for detecting pipe wall thickness changes; and size measurement features reflect pipe deformation and interface misalignment conditions. Mapping these multi-source features to a unified manifold space is a key step for accurately identifying complex defects of underground pipe networks.

[0077] The obtained pipe defect initial feature data is input into the feature mapping module. The feature mapping module constructs an adjacency graph by calculating the near-neighbor relationship between sample points in the initial feature data. In the underground pipe network scenario, the sample points represent the defect features at different positions of the pipe. For example, the data collected by the detection robot along the pipe includes pipe wall thickness changes, crack features, interface states, and other information, which together constitute the initial feature data of the underground pipe network defects. When constructing the adjacency graph, the pipe network defect points with similar features and close spatial positions are connected to form a topological structure representing the distribution characteristics of the underground pipe network defects. Specifically, for each sample point, K nearest neighbors are selected, for example, K = 10, and these nearest neighbors are determined by calculating the Euclidean distance. When sample point i and sample point j are nearest neighbors of each other, a connection is established in the adjacency graph. Next, the similarity relationship between the sample points in the adjacency graph is calculated to obtain a weight matrix. The similarity relationship is calculated by 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, then the similarity weight between them is about 0.04. If the sample points are not adjacent, the corresponding weight is 0.

[0078] Based on the weight matrix, a diagonal matrix is constructed, where the diagonal elements are the sum of all elements of the corresponding row of the weight matrix. For example, if the weight sum of a sample point and all its neighbors is 5.3, then the value at that position in the diagonal matrix is 5.3. Then the difference matrix is calculated, which is the difference between the diagonal matrix and the weight matrix. The difference matrix is used to solve the generalized eigenvalue problem to obtain the projection matrix. The column vectors of the projection matrix are the eigenvectors of the difference matrix corresponding to the smallest non-zero eigenvalue. The first d eigenvectors are selected to form the projection matrix, for example, d = 50, for dimension reduction. The projection matrix is used to map the initial feature data to the manifold space to obtain the manifold feature data.

[0079] In the manifold space where the manifold feature data resides, a local tangent plane needs to be constructed to convert the global nonlinear space of underground pipe network defects into a local linear space. For each sample point on the manifold, m of its nearest neighbors (e.g., m = 15) are selected to form a local region, and the tangent plane approximation is performed within the local region. The tangent plane is determined through local principal component analysis, that is, the covariance matrix of the sample points within the local region is calculated, and the principal direction thereof is extracted as the basis vector of the tangent plane.

[0080] 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 the sample points, and the weight of the edge is the Euclidean distance between the corresponding sample points in the local linear space. If two points are not in the same local region, there is no direct connection between them. The Dijkstra algorithm is used to calculate the shortest path distance between any two points. For example, the geodesic distance of sample A and sample B is calculated as 7.2. The geodesic distance is then converted into a similarity score using a negative exponential transformation: similarity score = negative exponential function (negative geodesic distance divided by an adjustment parameter). If the geodesic distance is 7.2 and the adjustment parameter is 3.5, the similarity score is about 0.13. Based on the similarity scores between all pairs of sample points, a complete similarity matrix is constructed.

[0081] 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, and its eigenvalues and eigenvectors are calculated. The first k (e.g., k = 8) eigenvectors are selected to form a new feature representation, and the K-means algorithm is used to cluster these new features to divide the samples into different feature subspaces.

[0082] In each feature subspace, the local distribution characteristics of the pipe defects are calculated to obtain a feature distribution matrix. The local distribution characteristics are obtained by calculating the statistics of the samples in the subspace, including the mean vector, the covariance matrix, the skewness, and the kurtosis, etc. 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 feature distribution matrix of the subspace.

[0083] Based on the feature distribution matrix, a metric tensor is constructed to represent the local geometric structure of the subspace. The metric tensor is obtained by calculating the local scatter of the sample points in the subspace, which reflects the direction sensitivity of the region. Using the covariance information in the feature distribution matrix, the Riemann metric tensor is constructed to represent the local deformation in different directions.

[0084] The geometric difference values between the plurality of feature subspaces are calculated using the metric tensors. The geometric difference between two subspaces is measured by the log-Euclidean distance between their metric tensors. If the log-Euclidean distance between the metric tensors of subspace 1 and subspace 2 is calculated as 2.8, it means that they have a large geometric difference.

[0085] The geometric difference values are converted into corresponding feature subspace weight coefficients. The softmax normalization method is adopted to make the sum of the weight coefficients equal to 1, and the subspace with a larger geometric difference obtains a higher weight. For example, if there are 3 subspaces with geometric difference values of 2.8, 1.5, and 0.9 respectively, the converted weight coefficients can be 0.5, 0.3, and 0.2.

[0086] The feature distribution matrices of the plurality of feature subspaces are multiplied by the corresponding weight coefficients respectively, and the results are superimposed to obtain the final defect representation vector. This representation vector fuses the feature distribution information of different subspaces and takes into account the geometric difference of each subspace, which can more comprehensively represent the characteristics of the pipeline defects and is beneficial to the subsequent defect classification and identification tasks.

[0087] In an optional implementation, the step of dividing the manifold feature data into a plurality of feature subspaces based on the similarity matrix comprises:

[0088] In a 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 neighbor sample points within the search radius is counted to obtain a local density initial value of the sample point, and a density ratio value of the sample point and the corresponding neighbor sample point is calculated according to the local density initial value to obtain a local density weight;

[0089] The similarity value between each sample point in the similarity matrix is multiplied by the corresponding local density weight to determine a density weighting coefficient, and the density weighting coefficient is used to update the similarity matrix to obtain a density-aware similarity matrix;

[0090] The manifold feature data is grouped and divided according to the similarity relationship between the sample points in the density-aware similarity matrix, and 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 a preset similarity threshold, the two sample points are divided into the same feature subspace to obtain a plurality of feature subspaces.

[0091] In one specific implementation, a set of manifold feature data of underground pipe network defect states is obtained. These manifold feature data come from multi-source data collected in the underground pipe network detection process, including visual images, acoustic signals, and laser scanning, etc. In the underground pipe network environment, different types of defects such as inner wall scaling, pipe wall corrosion, interface loosening, etc. form complex distribution patterns in the feature space. In particular, for the old pipe network that has been in operation for a long time, the defect density and type in different areas are obviously different, and these differences are reflected in the manifold feature data, which is helpful for subsequent defect identification and classification. The data set contains multiple sample points, each of which represents an underground pipe network defect state, such as pipe rupture, corrosion degree, interface loosening, etc. These sample points are distributed in a high-dimensional feature space and exhibit certain manifold structure characteristics. In the analysis of underground pipe network defects, the density distribution characteristics of manifold feature data are particularly important. Due to the complexity of the underground environment, pipe network defects often exhibit an aggregated distribution characteristic, such as corrosion areas usually concentrated in pipe sections under specific environmental conditions, and cracks often occur in stress concentration areas. By constructing a density-aware similarity matrix, the spatial distribution pattern of underground pipe network defects can be more accurately captured, thereby improving the accuracy of subspace partitioning and providing a more reliable data foundation for subsequent defect assessment and early warning.

[0092] For the obtained manifold feature data, a search radius is constructed in the local linear space to determine the neighborhood 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 neighbor sample points within the search radius is counted for each sample point to obtain the initial local density of the sample point. For example, for sample point P1, there are 8 neighbor sample points within the search radius, and the initial local density of P1 is 8; for sample point P2, there are 5 neighbor sample points within the search radius, and the initial local density of P2 is 5.

[0093] According to the calculated initial local density, the density ratio between the sample point and its neighbor sample points is calculated to obtain the local density weight. For each sample point, divide the initial local density of the sample point by the initial local density of all neighbor sample points within the search radius to calculate the density ratio. For example, the initial local density of sample point P1 is 8, and the initial local density of one of its neighbor sample points P3 is 4, so the density ratio between P1 and P3 is 8 / 4=2. The average of the density ratios between the sample point and all its neighbor sample points is taken to obtain the local density weight of the sample point. If the average of the density ratios between sample point P1 and its 8 neighbor sample points is 1.5, then the local density weight of P1 is 1.5.

[0094] The similarity matrix is updated to generate a density-aware similarity matrix. Assuming 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 x 1.2 = 1.8. Multiplying the original similarity value 0.8 by the density weighting coefficient 1.8, the new similarity value between P1 and P2 in the density-aware similarity matrix is 0.8 x 1.8 = 1.44. In this way, the similarity value between each pair of sample points in the similarity matrix is updated, and the density-aware similarity matrix DS is obtained.

[0095] After obtaining the density-aware similarity matrix DS, the manifold feature data is grouped and divided according to the similarity relationship 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 difference between the local density initial values 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.

[0096] For any two sample points, if the absolute value of the difference between their local density initial 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 divided into the same feature subspace. For example, the local density initial values of sample points P1 and P2 are 8 and 6 respectively, the difference between their local density initial 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, so P1 and P2 are divided into the same feature subspace.

[0097] By iteratively processing all pairs of sample points, the entire manifold feature data set is finally divided into multiple feature subspaces. For example, in the underground pipe network defect state data containing 100 sample points, it may be divided into 5 feature subspaces, representing different types or different severity of defect states, such as a slight corrosion subspace (containing 23 sample points), a moderate corrosion subspace (containing 18 sample points), a severe corrosion subspace (containing 15 sample points), a slight rupture subspace (containing 25 sample points), and a severe rupture subspace (containing 19 sample points).

[0098] This manifold feature data division method based on the density-aware similarity matrix fully considers the density distribution characteristics of sample points in the local region, can more accurately identify the internal relevance of underground pipe network defect states, and provides more accurate data support for subsequent defect analysis and prediction, thereby improving the efficiency of underground pipe network maintenance and management.

[0099] In an optional implementation, the pipeline defect feature vector is constructed as a correlation matrix, a connection relationship matrix is obtained by calculating the spatial distance correlation value and the mechanical action strength between pipeline defect features, an iteration propagation is performed to obtain a pipeline state sequence, a pipeline state evolution graph is constructed and a propagation direction is optimized, and a pipeline defect feature map is generated, including:

[0100] The pipeline defect feature vector is constructed as a correlation matrix, the spatial distance correlation value between each group of pipeline defect features is calculated based on the correlation matrix, and normalization processing is performed to obtain an initial connection matrix;

[0101] The mechanical action strength between pipeline defect features is calculated according to the spatial distance correlation value in the initial connection matrix, and the mechanical action strength is converted into a connection weight to obtain a connection relationship matrix;

[0102] A pipeline state transition matrix is constructed based on the connection relationship matrix, and an iteration propagation is performed on the pipeline defect features, in which the state information of adjacent pipeline features is fused to update the current pipeline state in each iteration process, and a pipeline state sequence is obtained;

[0103] A state evolution graph is constructed based on the pipeline state sequence, a transition probability between nodes in the state evolution graph is calculated to obtain a propagation direction, the propagation direction is filtered according to a preset pipeline structure constraint condition, and an optimized direction is obtained;

[0104] The pipeline state is layered based on the optimized direction, a hierarchical pipeline correlation structure is constructed, an importance index of each layer of feature nodes is calculated, the feature nodes are screened based on the importance index, and a pipeline defect feature map is generated.

[0105] In one specific implementation, the defect state of the underground pipe network has obvious spatiotemporal correlation and evolution law. For example, the corrosion defect of the pipeline will expand over time and may induce structural damage in the surrounding area, forming a defect chain. By constructing a correlation matrix based on the defect characterization vector, the mutual influence relationship between defects at different positions in the underground pipe network can be quantified, and the propagation path and mechanism of the defects can be revealed. This feature map construction method based on deep learning can upgrade the traditional isolated defect detection to systematic pipe network health state evaluation. The collected pipeline defect data includes defect type, location, size, shape, and other feature information. 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 position coordinates, the fourth value represents the defect type, the next three values represent the defect size, and the last value represents the material parameter. In the underground pipe network detection scene, these defect data are usually collected from detection robots equipped with multiple sensors, which can identify multiple defect types such as pipe wall deformation, cracks, corrosion, etc. The special environment of the underground pipe network makes the defects present diversified features, such as pipe wall corrosion depth, crack length and width, interface misalignment degree, etc. These data collectively constitute the pipeline defect characterization vector, which is used for subsequent feature map construction and defect evaluation.

[0106] All pipeline defect characterization vectors are combined to construct a correlation matrix. Assuming that there are f pipeline defect points, each defect point is represented by a g-dimensional feature vector, and the dimension of the correlation matrix is f x g. Based on the correlation matrix, the spatial distance correlation value between each group of pipeline defect features is calculated. The spatial distance correlation value is calculated using the Euclidean distance. For defect points 1 and 2, the spatial distance is the Euclidean distance of the spatial position coordinates in the characterization vectors of the two points. For example, 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 spatial distance calculation results of all defect point pairs form a spatial distance matrix, and the dimension of the spatial distance matrix is f x f.

[0107] The spatial distance matrix is normalized to obtain an initial connection matrix. The normalization method uses a Gaussian kernel function to convert the spatial distance to a correlation degree value between 0 and 1. The smaller the spatial distance, the greater the correlation degree. For example, the spatial distance between defect points 1 and 2 is 3.01, and the correlation degree value after normalization is 0.82. Each element in the initial connection matrix represents the spatial correlation degree between two defect points.

[0108] The mechanical action strength between the pipeline defect features is calculated according to the spatial distance correlation value in the initial connection matrix. The mechanical action strength considers factors such as defect type, size, and material parameters. For defect points 1 and 2, if the spatial correlation degree exceeds the threshold value 0.7 and the defect type similarity is greater than 0.6, the mechanical action strength is calculated, otherwise the mechanical action strength is 0. For example, the spatial correlation degree of defect points 1 and 2 is 0.82, the defect type similarity is 0.85, and the calculated mechanical action strength is 0.76.

[0109] The mechanical action strength is converted into a connection weight to obtain a connection relationship matrix. The conversion method uses a sigmoid function to map the mechanical action strength to the [0, 1] interval as the connection weight. For example, the mechanical action strength of defect points 1 and 2 is 0.76, and the converted connection weight is 0.82. The dimension of the connection relationship matrix is f x f, and each element in the connection relationship matrix represents the connection weight between two defect points.

[0110] A pipeline state transition matrix is constructed based on the connection relationship matrix. The elements in the state transition matrix represent the probability of transitioning from one defect point state to another defect point state. Each row of the connection relationship matrix is normalized so that the sum of the elements in each row is 1 to obtain the state transition matrix. For example, the connection weight of defect point 1 to other defect points is [0, 0.82, 0.56, 0.43, 0], and after normalization, it becomes [0, 0.45, 0.31, 0.24, 0].

[0111] 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 process, the state transition matrix is multiplied by the current state vector S t to obtain 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} is obtained.

[0112] A state evolution graph is constructed based on the pipeline state sequence. The nodes in the state evolution graph represent defect points, and the edges represent the connection relationship between defect points, with the edge weight being the corresponding value in the connection relationship matrix. The transition probability between nodes in the state evolution graph is calculated to obtain the propagation direction. For example, the transition probability from defect point 1 to defect point 2 is 0.45, indicating the probability of defect propagation from point 1 to point 2.

[0113] The propagation direction is filtered according to a preset pipeline structure constraint condition to obtain an optimized direction. The pipeline structure constraint condition includes: the propagation direction should comply with the pipeline flow direction; the propagation distance should not exceed a preset threshold; the propagation path should avoid special structure areas, etc. For example, if the propagation from the defect point 3 to the point 4 is opposite to the pipeline flow direction, the propagation direction is filtered out.

[0114] The pipeline state is layered based on the optimized direction to construct a hierarchical pipeline correlation structure. From the source point, the defect points are divided into different levels according to the propagation direction. For example, the defect point 1 is the source point (0th level), the defect points 2, 3 and 4 directly connected thereto are the 1st level, other defect points directly connected to the 1st level are the 2nd level, and so on.

[0115] The importance index of each level feature node is calculated. The importance index comprehensively considers the degree centrality, closeness centrality and betweenness centrality of the node. For example, the degree centrality of the defect point 2 is 3, the closeness centrality is 0.68, the betweenness centrality is 0.45, and the calculated importance index is 0.72. According to the importance index, the feature nodes are screened, and the nodes with an importance index greater than a threshold value of 0.5 are retained to generate a pipeline defect feature map. The finally generated pipeline defect feature map contains the hierarchical relationship, connection relationship and importance information between the defect points, and can directly display the distribution characteristics and evolution trend of the pipeline defects.

[0116] In an optional implementation, a pipeline state transition matrix is constructed based on the connection relationship matrix, and an iterative propagation of pipeline defect features is performed, in each iteration process, the state information of adjacent pipeline features is fused to update the current pipeline state, and a pipeline state sequence is obtained, including:

[0117] Eigenvalue decomposition of the connection relationship matrix is calculated to obtain an eigenvector matrix, different scale frequency response functions are constructed by using the eigenvector matrix, and convolution operation is performed on the connection relationship matrix and the corresponding frequency response function to obtain pipeline state transition matrices of multiple scales;

[0118] Pipeline defect feature propagation is performed on each scale pipeline state transition matrix to obtain multiple groups of pipeline state sequences; the difference degree of adjacent states in each group of pipeline state sequences is calculated to obtain a propagation response value, and an attention weight is determined based on the propagation response value; and the multiple groups of pipeline state sequences are fused by using the attention weight to obtain a fused pipeline state.

[0119] The deviation between the fusion pipeline state and the initial pipeline state is calculated to obtain a state residual, and the state residual is fed back to a pipeline defect feature propagation process; the attention weight is dynamically adjusted based on the state residual, and the intensity of pipeline defect feature propagation is adaptively adjusted; the feature propagation update is repeatedly performed until the state residual is less than a preset residual threshold, and an optimized pipeline state sequence is obtained.

[0120] In a specific embodiment, the propagation process of underground pipe network defects generally follows specific physical laws and topological structure constraints. Unlike traditional single-point defect analysis, this method can comprehensively capture the propagation characteristics of defect signals in underground pipe networks at different spatial scales through eigenvalue decomposition and multi-scale frequency response functions. For example, the local scale (parameter 0.5) mainly reflects the correlation between adjacent pipe sections, which is suitable for identifying microscopic crack propagation; the medium scale (parameter 1.0) captures the interaction between pipe section groups, which is suitable for analyzing the diffusion of corrosion regions; and the global scale (parameter 2.0) focuses on the structural integrity changes of the entire pipe network system, which helps to evaluate the impact of large-scale settlement or geological changes on the pipe network. For the eigenvalue decomposition process of the connection relationship matrix, the eigenvalues and corresponding eigenvectors of the connection relationship matrix are obtained through iterative calculation. Taking an example of an underground pipe network, which contains 200 pipe nodes, the connection relationship matrix is a 200x200 matrix. The eigenvalue decomposition of this matrix is performed to obtain 200 eigenvalues and their corresponding eigenvectors, which form the eigenvector matrix. Each column of the eigenvector matrix represents an eigenvector, which represents the topological information in the pipe network structure.

[0121] When constructing frequency response functions of different scales based on the eigenvector matrix, three different scale parameters are set, which are 0.5, 1.0 and 2.0, corresponding to the frequency responses of local, medium and global scales. By combining the eigenvector matrix with the corresponding scale parameters, three groups of frequency response functions are constructed. Each group of frequency response functions can capture the signal propagation characteristics of the pipe network structure at different scales. For the frequency response function with scale parameter 0.5, it mainly focuses on the local connection characteristics between adjacent nodes in the pipe network; for the frequency response function with scale parameter 1.0, it can capture the connection mode within a medium range; and the frequency response function with scale parameter 2.0 focuses on the expression of global connection structure.

[0122] When performing convolution operation between connection relationship matrix and frequency response function, convolution operation is performed between connection relationship matrix and frequency response function of each scale. Specifically, for each scale parameter, the connection relationship matrix is multiplied by the corresponding frequency response function, and the result matrix is normalized to finally obtain three pipe state transition matrices of different scales. These transition matrices represent the state transition probabilities between pipe network nodes at different scales.

[0123] In the pipeline defect feature propagation stage, a feature propagation process is performed for each scale of the pipeline state transition matrix. With the initial pipeline state vector as input, the pipeline defect features are propagated along the path defined by the transition matrix through iterative calculation. In specific implementation, the propagation step number is set to 10, and for each propagation step, the current state vector is multiplied by the state transition matrix to obtain the state vector at the next time. After 10 steps of propagation, 3 groups of pipeline state sequences each containing 11 state vectors are obtained (including the initial state).

[0124] In the calculation of the propagation response value, the difference between adjacent state vectors in each group of pipeline state sequences is analyzed. Specifically, the Euclidean distance between adjacent state vectors is calculated, and these distance values are normalized to obtain the propagation response value. For the 3 groups of state sequences in the example, 10 propagation response values are obtained respectively. The larger the propagation response value, the more significant the change in the pipeline state in that propagation step, and there may be more important defect information.

[0125] In the process of determining the attention weight based on the propagation response value, the softmax function is applied to convert the propagation response value into the attention weight. For each scale of the state sequence, the corresponding 10 propagation response values are converted into weight coefficients with a sum of 1 by the softmax function. In this way, for the propagation step with significant changes, a higher attention weight is given; for the step with smaller changes, a lower weight is given.

[0126] In the fusion stage of multiple groups of pipeline state sequences, the calculated attention weights are used to weight and fuse the 3 groups of state sequences. Specifically, for each pipeline node, the weighted state values at 3 scales are calculated respectively, and the 3 weighted values are added to obtain the fused pipeline state. The fusion process takes into account the importance of defect information at different scales, and can more comprehensively represent the actual state of the pipeline.

[0127] In the state residual calculation process, the fused pipeline state is compared with the initial pipeline state to calculate the difference between them, and the state residual vector is obtained. The state residual reflects the degree of correction of the initial state through the feature propagation process. In the example, the initial residual threshold is set to 0.01, and if the norm of the state residual is less than the threshold, it is considered that the state has converged, otherwise further optimization is needed.

[0128] In the state residual feedback and adaptive adjustment stage, the state residual information is fed back to the feature propagation process to dynamically adjust the attention weight and the propagation intensity. According to the size of the state residual, the propagation step length at 3 scales is adjusted in proportion, and for the areas with larger residuals, the propagation step length is increased; for the areas with smaller residuals, the propagation step length is decreased. This adaptive adjustment mechanism can more accurately optimize the feature propagation for problem areas.

[0129] The process of feature propagation, state fusion and residual calculation is repeatedly performed until the state residual is less than a preset threshold 0.01 or the maximum number of iterations 50 is reached. In the example, after 12 iterations, the state residual is reduced to 0.0092, which is less than the threshold, so the iteration is stopped, and the final optimized pipeline state sequence is obtained.

[0130] Generally, the underground pipe network defect identification method mainly relies on manual detection or simple signal processing technology, which cannot effectively process the defect propagation characteristics in complex pipe network structure. Although the existing deep learning method introduces a neural network model, it mainly uses fixed network structure and parameters, which is difficult to adapt to pipe networks of different scales and complexities. The method of the embodiment proposes an adaptive method based on eigenvalue decomposition and multi-scale feature propagation. The multi-scale frequency response function is introduced to capture defect features of different ranges, and the adaptive adjustment is realized through attention mechanism and residual feedback.

[0131] As shown in Figure 2 , a pipeline defect feature propagation and adaptive adjustment flowchart is shown.

[0132] In an optional implementation, an adaptive sliding window is used to dynamically sample the pipeline defect feature map to obtain a pipeline defect evolution sequence, an influence factor set is determined based on a counterfactual intervention strategy analysis, a pipeline state transition probability is calculated according to the influence factor set, and an optimal pipeline maintenance scheme is determined based on a multi-objective decision function, including:

[0133] 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 change frequency of the pipeline defect feature, the statistics and change trend features of the pipeline features in the sliding window are calculated to obtain a local feature sequence;

[0134] The local feature sequence is constructed into a time series pyramid structure, the pipeline feature representation is extracted at each level of the time series pyramid structure, a cross-level feature propagation network is constructed to calculate the cooperative evolution relationship between pipeline features, and a pipeline defect evolution sequence is obtained based on the cooperative evolution relationship.

[0135] Based on the pipeline defect evolution sequence, a counterfactual intervention strategy is generated to construct a control experiment scene, and a deviation analysis is performed on the control experiment scene and the actual scene to obtain an influence factor set;

[0136] A pipeline defect state descriptor is constructed using the influence factor set, a state transition matrix under different pipeline maintenance schemes is calculated based on the pipeline defect state descriptor, and a pipeline state transition probability is obtained by probability distribution estimation of the state transition matrix;

[0137] The pipeline state transition probability is constructed into a multi-objective decision function with pipeline maintenance cost, maintenance result and maintenance risk, and an optimization solution is performed based on the multi-objective decision function to obtain a pipeline maintenance scheme set, a reliability score of each pipeline maintenance scheme set is calculated, a pipeline maintenance scheme with the highest reliability score is selected, and an optimal pipeline maintenance scheme is determined.

[0138] In a specific embodiment, in the underground pipe network scene, the defect feature map contains the structural damage distribution information of the pipeline along the pipeline. The underground pipe network is eroded by the external environment and affected by the internal medium for a long time, forming different degrees of damage. The adaptive sliding window sampling can dynamically adjust the sampling interval according to the defect distribution density, provide more detailed data collection in the area where the defect changes significantly, and thus accurately depict the spatial distribution characteristics and time evolution trend of the underground pipe network defects. After obtaining the pipeline defect feature map, adaptive sliding window sampling is performed. The initial size of the sliding window is set to 30 days of data, and is dynamically adjusted according to the frequency of change of the pipeline defect features. Specifically, when the feature change rate of the continuous 5 sampling points exceeds the preset threshold of 10%, the window size is reduced to 80% of the current window; when the feature change rate of the continuous 10 sampling points is lower than the preset threshold of 5%, the window size is increased to 120% of the current window. For each sliding window, the statistics of the pipeline features in the window are calculated, including the mean, standard deviation, kurtosis and skewness, and the change trend features are calculated, including the first-order difference mean, second-order difference mean and trend slope, so as to obtain the local feature sequence. In actual application, for example, for a 500-meter long oil pipeline, the average corrosion depth in a sliding window is 2.5 mm, the standard deviation is 0.8 mm, and the first-order difference mean is 0.15 mm / day, indicating that the corrosion degree of the pipeline is accelerating. The defect evolution of the underground pipe network often presents nonlinear and multi-stage characteristics. The adaptive sliding window technology can dynamically adjust the sampling frequency according to the rate of change of the defects of different pipe sections, which is of great significance for accurately capturing sudden defects (such as sudden cracks) and gradual defects (such as slow corrosion) in the underground pipe network. By constructing a time pyramid structure, the defect evolution patterns at different time scales can be analyzed simultaneously, such as short-term pipe wall stress changes, medium-term material degradation processes and long-term structural stability trends, providing data support for the whole life cycle management of the underground pipe network.

[0139] The local feature sequence is constructed into a time sequence pyramid structure, which contains three levels. The first level retains all local features at the original sampling frequency; the second level is formed by aggregating every 2 adjacent window features, with the sampling frequency reduced by half; the third level is formed by aggregating every 4 adjacent window features, with the sampling frequency reduced to 1 / 4. On each level, a feature extraction module is used to extract a pipeline feature representation, which adopts a three-layer convolutional network structure with convolution kernel sizes of 3x3, 5x5 and 7x7, and channel numbers of 16, 32 and 64, respectively. The inter-level feature propagation network calculates the correlation weights of different level features through a self-attention mechanism, and the weight threshold is set to 0.3. The correlation below the threshold will be pruned. Based on the co-evolution relationship, cross-level feature recursive updating is performed, with an update step of 0.05 and an iteration number of 50. Finally, the pipeline defect evolution sequence is obtained. For example, the analysis of the pipeline wall thickness thinning phenomenon shows that the initial micro crack is identified at the first level, the crack propagation trend is captured at the second level, and the long-term evolution result is predicted at the third level, which completely describes the evolution process from a small defect to a potential leak.

[0140] Based on the pipeline defect evolution sequence, counterfactual intervention strategies are constructed. First, five control intervention scenarios are set, including: environmental temperature increase of 10℃, environmental humidity increase of 15%, pipeline pressure increase of 20%, fluid pH change of ±1.5, and pipeline service life increase of 2 years. For each intervention scenario, a homogeneous simulator is used to generate pipeline defect evolution data, and deviation analysis is performed with actual scene data. The deviation calculation uses the root mean square error (RMSE), and the deviation threshold is set to 15%. When the deviation of a certain intervention scenario exceeds the threshold, the corresponding factor is included in the influence factor set. In a case analysis, the deviation caused by environmental temperature change is 17.3%, the deviation caused by pipeline pressure change is 19.8%, and the deviation caused by environmental humidity change is 12.5%, so the environmental temperature and pipeline pressure are determined as the key influence factors.

[0141] A set of influencing factors is used to construct a pipeline defect state descriptor, which includes 8 dimensions: defect depth, defect area, defect propagation rate, environmental temperature, pipeline pressure, fluid corrosiveness, pipeline material type, and service life. For each dimension, a discretization process is adopted to divide the continuous value into 5 levels. Based on the pipeline defect state descriptor, a state transition matrix is calculated under different pipeline repair schemes. The repair schemes include: no repair, local repair, partial replacement, and full replacement. For each repair scheme, an initial state transition matrix is constructed based on historical data and expert knowledge, and the matrix size is 125x125 (8 times of 5 state levels). Monte Carlo simulation is used to sample 10000 times to estimate the probability distribution of the state transition matrix, and the pipeline state transition probability is obtained. For example, for the local repair scheme, the probability of transitioning from state [3, 2, 4, 3, 4, 2, 1, 3] to state [2, 1, 3, 3, 4, 2, 1, 3] is 0.73, indicating that this repair scheme can effectively reduce defect depth and area.

[0142] The pipeline state transition probability, pipeline repair cost, repair result, and repair risk are constructed into a multi-objective decision function. The repair cost includes material cost, labor cost, and production loss; the repair result evaluation includes expected service life extension and performance recovery degree; the repair risk evaluation includes construction risk and secondary damage risk. The weights of these three dimensions are set to 0.4, 0.35, and 0.25, respectively. The multi-objective decision function adopts a weighted summation form, and a simulated annealing algorithm is executed based on this function to obtain the optimal solution. The initial temperature is set to 100, the cooling coefficient is 0.95, and the iteration number is 1000 times, and the pipeline repair scheme set is obtained. The reliability score is calculated for each repair scheme, considering three aspects: historical success rate (weight 0.5), technical maturity (weight 0.3), and adaptability (weight 0.2). Finally, the pipeline repair scheme with the highest reliability score is selected as the optimal scheme. In a practical case, the score of local repair is 0.86, the score of partial replacement is 0.92, and the score of full replacement is 0.78, so partial replacement is determined as the optimal pipeline repair scheme.

[0143] In an alternative embodiment, a counterfactual intervention strategy is generated based on the pipeline defect evolution sequence to construct 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, which includes:

[0144] The time sequence nodes with feature change rates exceeding a preset change rate threshold in the pipeline defect evolution sequence are calculated, and the intensity of the pipeline defect state feature at the time sequence nodes is adjusted to generate a counterfactual intervention strategy;

[0145] Performing pipeline defect state simulation according to the counterfactual intervention strategy and pipeline constraint conditions, performing evolutionary calculation on the pipeline defect state under each intervention intensity to generate multiple sets of control experiment scenarios;

[0146] Calculating the feature deviation between the control experiment scenarios and the actual pipeline scenarios, performing cluster analysis on the feature deviation to obtain feature deviation clusters, calculating the variance contribution rate of the pipeline features in each feature deviation cluster, and selecting features with a variance contribution rate exceeding a preset variance threshold as dominant factors;

[0147] Calculating the time series correlation coefficient based on the dominant factors, determining the contribution weight of each dominant factor according to the time series correlation coefficient, and selecting dominant factors with a contribution weight greater than a preset contribution weight threshold;

[0148] Calculating the degree of association between the dominant factors, hierarchically sorting the dominant factors, and generating an impact factor set.

[0149] In a specific implementation, the pipeline defect evolution sequence generally includes defect state feature data at multiple time points. In underground pipeline network monitoring applications, the defect evolution sequence reflects the process of pipeline damage accumulation over time. Underground pipeline networks are affected by various internal and external factors during long-term operation, and the defect state will exhibit certain evolution rules. By recording the changes of key parameters (such as wall thickness reduction rate, crack propagation speed, etc.) over time, a complete defect evolution sequence is constructed. Counterfactual intervention analysis can identify key factors affecting the development of underground pipeline defects, such as environmental humidity and internal pressure changes, providing a scientific basis for pipeline maintenance decisions. Read the pipeline defect evolution sequence dataset, which includes pipeline state parameters at different time points, such as wall thickness, corrosion depth, crack length, etc. For the evolution sequence data of a certain pipeline, the system calculates the change rate of each feature between adjacent time points. For example, for the corrosion depth feature, if the time value 1 is 2.5 mm and the time value 2 is 3.2 mm, the change rate is 28%. The system compares the change rate of each feature with the preset change rate threshold (usually set to 15%) and marks the time nodes with a change rate exceeding the threshold. In practical applications, the wall thickness reduction rates of a certain pipeline segment on the 45th day and the 78th day are 17% and 22%, respectively, exceeding the preset threshold, so they are marked as key time nodes.

[0150] For the marked time nodes, the system designs a counterfactual intervention strategy to adjust the intensity of the pipeline defect state features. Intensity adjustment usually uses three adjustment coefficients of 0.8, 1.0, and 1.2, representing weakened, maintained, and enhanced intervention, respectively. For example, for a node with a wall thickness reduction rate of 22%, three intervention intensity counterfactual scenarios of 17.6% (0.8 x 22%), 22%, and 26.4% (1.2 x 22%) are generated.

[0151] Defect state simulation is performed according to counterfactual intervention strategies and pipeline constraint conditions. Pipeline constraint conditions include material properties (such as elastic modulus 210 GPa, Poisson's ratio 0.3), operating parameters (such as internal pressure 8 MPa, temperature 65°C) and geometric constraints. Finite element analysis method is used to calculate the evolution of pipeline defect state under each intervention intensity. For example, for the cases of wall thickness reduction rates of 17.6%, 22% and 26.4%, the defect development trend in the next 30 days is calculated respectively to obtain three sets of control experiment scene data. In actual calculation, a certain pipeline section under the intervention of 26.4% reduction rate has a maximum deformation of 2.1 mm after 30 days, while the standard scene is only 1.7 mm.

[0152] The feature deviation of the control experiment scene and the actual pipeline scene is calculated. Feature deviation calculation involves multiple feature dimensions, such as stress distribution, deformation, energy release rate, etc. For each feature, the deviation value is calculated and stored as a deviation matrix. For example, for the deformation feature, the intervention scene value is 2.1 mm, the actual scene value is 1.7 mm, the deviation is 0.4 mm, and the deviation rate is about 23.5%.

[0153] Cluster analysis is performed on the feature deviation, and K-means clustering algorithm is used to classify similar deviation patterns into a group. In practical application, the number of clusters is usually set to 3-5. After clustering, the variance contribution rate of each pipeline feature is calculated for each feature deviation cluster. Variance contribution rate calculation is based on principal component analysis method to determine the contribution of each feature to the total variance. For example, in a certain analysis, the variance contribution rates of wall thickness reduction, crack length and material hardness are 42%, 31% and 15% respectively. Select the features with variance contribution rate exceeding the preset variance threshold (usually 20%) as the dominant factors, and in this example, wall thickness reduction and crack length are identified as the dominant factors.

[0154] For the identified dominant factors, the time series correlation coefficient is calculated, and the lag correlation analysis method is used to evaluate the correlation degree of each factor with time. Time series correlation calculation is based on the change trend of each dominant factor in historical data. For example, the time series correlation coefficient of wall thickness reduction and crack length is 0.76, indicating that the two have strong time correlation. According to the time series correlation coefficient, the contribution weight of each dominant factor is determined, usually through normalization processing. In the example, the contribution weights of wall thickness reduction and crack length are 0.58 and 0.42 respectively. Select the dominant factors with contribution weight greater than the preset contribution weight threshold (usually 0.3), and in this example, both factors are retained.

[0155] The correlation degree between the dominant factors is calculated, and a factor correlation network is constructed. The correlation degree calculation is based on conditional mutual information and partial correlation analysis. For example, the analysis shows that the correlation coefficient between wall thickness thinning and crack length is 0.65, indicating a moderate mutual influence. The system ranks the dominant factors according to the correlation degree and generates the final set of influencing factors. In this example, the final set of influencing factors is: wall thickness thinning (first-level factor, weight 0.58) and crack length (second-level factor, weight 0.42).

[0156] Through the above method, the key influencing factors in the evolution process of pipeline defects are successfully identified, providing data support for subsequent preventive maintenance and risk management.

[0157] The underground pipe network defect intelligent identification and early warning system based on deep learning of the embodiment of the application comprises:

[0158] The first unit is configured to obtain multi-source detection data of the underground pipe network by using a pipeline detection robot, and to pre-process the data to obtain initial feature data;

[0159] The second unit is configured to map the initial feature data to a manifold space through a feature mapping module to obtain manifold feature data, construct a similarity matrix in the manifold space to divide a feature subspace, and perform geometric difference weighted fusion on the feature subspace to obtain a pipeline defect representation vector;

[0160] The third unit is configured to construct the pipeline defect representation vector into a correlation matrix, calculate spatial distance correlation values and mechanical action intensities between pipeline defect features to obtain a connection relationship matrix, perform iterative propagation to obtain a pipeline state sequence, construct a pipeline state evolution graph and optimize the propagation direction, and generate a pipeline defect feature map;

[0161] The fourth unit is configured to dynamically sample the pipeline defect feature map by using an adaptive sliding window to obtain a pipeline defect evolution sequence, determine an influencing factor set based on a counterfactual intervention strategy analysis, calculate a pipeline state transition probability according to the influencing factor set, and determine an optimal pipeline maintenance scheme based on a multi-objective decision function;

[0162] The fifth unit is configured to generate pipe network defect early warning information when the pipeline state transition probability exceeds a preset probability threshold;

[0163] The sixth unit is configured to send the pipe network defect early warning information and the optimal pipeline maintenance scheme to a monitoring terminal.

[0164] In a third aspect, the embodiment of the application provides an electronic device, comprising:

[0165] a processor;

[0166] a memory for storing processor-executable instructions;

[0167] The processor is configured to invoke instructions stored in the memory to perform the method described above.

[0168] In a fourth aspect, the present application provides a computer readable storage medium having stored thereon computer program instructions, which when executed by a processor implement the method described above.

[0169] The present application can be a method, apparatus, system, and / or computer program product. Computer program products can include computer readable storage media having computer readable program instructions thereon for performing various aspects of the present application.

[0170] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A deep learning-based underground pipe network defect intelligent identification and early warning method, characterized in that, The application relates to a pipeline defect feature mapping method and device. The underground pipeline network is detected by a pipeline detection robot to obtain multi-source detection data, and initial feature data is obtained after preprocessing; The initial feature data is mapped to a manifold space by a feature mapping module to obtain manifold feature data, a similarity matrix is constructed in the manifold space to divide a feature subspace, and a pipeline defect representation vector is obtained by geometric difference weighted fusion of the feature subspace, including: The initial feature data is input into the feature mapping module, the feature mapping module calculates the near-neighbor relationship of sample points in the initial feature data to construct an adjacency graph, calculates the similarity relationship between the sample points in the adjacency graph to obtain a weight matrix, constructs a diagonal matrix according to the weight matrix and calculates a difference matrix, obtains a projection matrix based on the difference matrix, and maps the initial feature data to a manifold space by using the projection matrix to obtain manifold feature data; A local tangent plane is constructed in the manifold space where the manifold feature data is located, the global nonlinear space of the underground pipeline network defect is converted into a local linear space, the shortest path between the sample points in the manifold feature data is calculated as a geodesic distance in the local linear space, and the geodesic distance is converted into a similarity score to construct a similarity matrix; The manifold feature data is divided into multiple feature subspaces based on the similarity matrix, the local distribution characteristics of the 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, the geometric difference values are converted into corresponding feature subspace weight coefficients, the feature distribution matrix of the multiple feature subspaces is multiplied by the corresponding weight coefficients respectively and is superimposed to obtain a defect representation vector; The pipeline defect representation vector is constructed into a correlation matrix, the spatial distance correlation value and the mechanical action strength between the pipeline defect characteristics are calculated to obtain a connection relationship matrix, an iteration propagation is performed to obtain a pipeline state sequence, a pipeline state evolution graph is constructed and the propagation direction is optimized, and a pipeline defect feature map is generated; The pipeline defect feature map is dynamically sampled by using an adaptive sliding window to obtain a pipeline defect evolution sequence, an influence factor set is determined based on a counterfactual intervention strategy analysis, a pipeline state transition probability is calculated according to the influence factor set, and an optimal pipeline maintenance scheme is determined based on a multi-objective decision function; When the pipeline state transition probability exceeds a preset probability threshold, pipeline network defect early warning information is generated; The pipeline network defect early warning information and the optimal pipeline maintenance scheme are sent to a monitoring terminal.

2. The method of claim 1, wherein, The manifold feature data is divided into multiple feature subspaces based on the similarity matrix, 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 pipeline network defect state, the number of near-neighbor sample points in the search radius is counted to obtain a local density initial value of the sample point, and a density ratio of the sample point and the corresponding near-neighbor sample point is calculated according to the local density initial value to obtain a local density weight; The similarity value between each sample point in the similarity matrix is multiplied by the corresponding local density weight to determine a density weighting coefficient, and the density weighting coefficient is used to update the similarity matrix to obtain a density-aware similarity matrix. Grouping the manifold feature data according to the similarity relationship between 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 a preset similarity threshold, the two sample points are divided into the same feature subspace, to obtain a plurality of feature subspaces.

3. The method of claim 1, wherein, The pipeline defect feature graph is generated by constructing a correlation matrix from the pipeline defect feature vectors, calculating the spatial distance correlation values between the pipeline defect features to obtain an initial connection matrix, calculating the mechanical action strength between the pipeline defect features according to the spatial distance correlation values in the initial connection matrix, converting the mechanical action strength into connection weights to obtain a connection relationship matrix, constructing a pipeline state transition matrix based on the connection relationship matrix, performing iterative propagation of the pipeline defect features, updating the current pipeline state by fusing the state information of adjacent pipeline features in each iteration process to obtain 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, filtering the propagation direction according to a preset pipeline structure constraint condition to obtain an optimized direction, and hierarchically layering the pipeline state based on the optimized direction to construct a hierarchical pipeline correlation structure, calculating an importance degree index of each layer of feature nodes, screening the feature nodes based on the importance degree index, and generating the pipeline defect feature graph. The pipeline defect feature graph is generated by constructing a correlation matrix from the pipeline defect feature vectors, calculating the spatial distance correlation values between the pipeline defect features to obtain an initial connection matrix, calculating the mechanical action strength between the pipeline defect features according to the spatial distance correlation values in the initial connection matrix, converting the mechanical action strength into connection weights to obtain a connection relationship matrix, constructing a pipeline state transition matrix based on the connection relationship matrix, performing iterative propagation of the pipeline defect features, updating the current pipeline state by fusing the state information of adjacent pipeline features in each iteration process to obtain 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, filtering the propagation direction according to a preset pipeline structure constraint condition to obtain an optimized direction, and hierarchically layering the pipeline state based on the optimized direction to construct a hierarchical pipeline correlation structure, calculating an importance degree index of each layer of feature nodes, screening the feature nodes based on the importance degree index, and generating the pipeline defect feature graph. The pipeline defect feature graph is generated by constructing a correlation matrix from the pipeline defect feature vectors, calculating the spatial distance correlation values between the pipeline defect features to obtain an initial connection matrix, calculating the mechanical action strength between the pipeline defect features according to the spatial distance correlation values in the initial connection matrix, converting the mechanical action strength into connection weights to obtain a connection relationship matrix, constructing a pipeline state transition matrix based on the connection relationship matrix, performing iterative propagation of the pipeline defect features, updating the current pipeline state by fusing the state information of adjacent pipeline features in each iteration process to obtain 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, filtering the propagation direction according to a preset pipeline structure constraint condition to obtain an optimized direction, and hierarchically layering the pipeline state based on the optimized direction to construct a hierarchical pipeline correlation structure, calculating an importance degree index of each layer of feature nodes, screening the feature nodes based on the importance degree index, and generating the pipeline defect feature graph. The pipeline defect feature graph is generated by constructing a correlation matrix from the pipeline defect feature vectors, calculating the spatial distance correlation values between the pipeline defect features to obtain an initial connection matrix, calculating the mechanical action strength between the pipeline defect features according to the spatial distance correlation values in the initial connection matrix, converting the mechanical action strength into connection weights to obtain a connection relationship matrix, constructing a pipeline state transition matrix based on the connection relationship matrix, performing iterative propagation of the pipeline defect features, updating the current pipeline state by fusing the state information of adjacent pipeline features in each iteration process to obtain 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, filtering the propagation direction according to a preset pipeline structure constraint condition to obtain an optimized direction, and hierarchically layering the pipeline state based on the optimized direction to construct a hierarchical pipeline correlation structure, calculating an importance degree index of each layer of feature nodes, screening the feature nodes based on the importance degree index, and generating the pipeline defect feature graph. ​ ​ 4. The method of claim 3, wherein, ​ ​ ​ ​ 5. The method of claim 1, wherein, The pipeline defect feature atlas is dynamically sampled by using an adaptive sliding window to obtain a pipeline defect evolution sequence, an influence factor set is determined based on a counterfactual intervention strategy analysis, a pipeline state transition probability is calculated according to the influence factor set, and an optimal pipeline maintenance scheme is determined based on a multi-objective decision function, including: The pipeline defect feature atlas is dynamically sampled by using an adaptive sliding window, the size of the sliding window is dynamically adjusted according to the change frequency of the pipeline defect feature, the statistical quantity and change trend feature of the pipeline feature in the sliding window are calculated to obtain a local feature sequence; The local feature sequence is constructed into a time series pyramid structure, pipeline feature representations are extracted at each level of the time series pyramid structure, a hierarchical feature propagation network is constructed to calculate the cooperative evolution relationship between pipeline features, and a pipeline defect evolution sequence is obtained based on the cooperative evolution relationship; A counterfactual intervention strategy is generated based on the pipeline defect evolution sequence, a control experiment scene is constructed, and a deviation analysis is performed on the control experiment scene and the actual scene to obtain an influence factor set; A pipeline defect state descriptor is constructed using the influence factor set, a state transition matrix under different pipeline maintenance schemes is calculated based on the pipeline defect state descriptor, and a pipeline state transition probability is obtained by probability distribution estimation on the state transition matrix; The pipeline state transition probability, pipeline maintenance cost, maintenance result and maintenance risk are constructed into a multi-objective decision function, and an optimal pipeline maintenance scheme is obtained by optimization solving based on the multi-objective decision function.

6. The method of claim 5, wherein, The pipeline defect evolution sequence is constructed based on the counterfactual intervention strategy to generate a control experiment scene, and a deviation analysis is performed on the control experiment scene and the actual scene to obtain an influence factor set, including: Calculate the time series node in the pipeline defect evolution sequence whose feature change rate exceeds the preset change rate threshold, adjust the intensity of the pipeline defect state feature at the time series node, and generate a counterfactual intervention strategy; According to the counterfactual intervention strategy and the pipeline constraint condition, the pipeline defect state simulation is performed, and the evolution calculation of the pipeline defect state under each intervention intensity is performed to generate multiple control experiment scenes; Calculate the feature deviation between the control experiment scene and the actual pipeline scene, perform clustering analysis on the feature deviation to obtain a feature deviation cluster, calculate the variance contribution rate of the pipeline feature in each feature deviation cluster, and select the feature whose variance contribution rate exceeds the preset variance threshold as the dominant factor; Based on the dominant factor, calculate the time series correlation coefficient, determine the contribution weight of each dominant factor according to the time series correlation coefficient, and select the dominant factor whose contribution weight is greater than the preset contribution weight threshold; Calculate the correlation degree between the dominant factors, sort the dominant factors by level, and generate an influence factor set.

7. A deep learning based underground pipe network defect intelligent identification and early warning system for implementing the method of any one of the preceding claims 1-6, characterized in that, The first unit is configured to acquire multi-source detection data of an underground pipeline network by using a pipeline detection robot, and to preprocess the data to obtain initial feature data. ​ The second unit is configured to map the initial feature data to a manifold space through a feature mapping module to obtain manifold feature data, construct a similarity matrix in the manifold space, divide a feature subspace, and perform geometric difference weighted fusion on the feature subspace to obtain a pipeline defect feature vector. The third unit is configured to construct the pipeline defect feature vector into a correlation matrix, calculate a spatial distance correlation value and a 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 graph, optimize a propagation direction, and generate a pipeline defect feature map. The fourth unit is configured to dynamically sample the pipeline defect feature map using an adaptive sliding window to obtain a pipeline defect evolution sequence, determine an influence factor set based on a counterfactual intervention strategy analysis, calculate a pipeline state transition probability according to the influence factor set, and determine an optimal pipeline maintenance scheme based on a multi-objective decision function. The fifth unit is configured to generate pipeline network defect early warning information when the pipeline state transition probability exceeds a preset probability threshold. The sixth unit is configured to send the pipeline network defect early warning information and the optimal pipeline maintenance scheme to a monitoring terminal.

8. An electronic device, comprising: The computer program instructions are executed by the processor to implement the method of any one of claims 1-6. The computer program instructions are executed by the processor to implement the method of any one of claims 1-6. ​ ​ 9. A computer-readable storage medium having stored thereon computer program instructions, wherein, ​

Images