Intelligent sealing method and system for underground pipe network leakage based on robot learning

Abstract

The present application relates to a robot learning-based underground pipe network leakage intelligent plugging method and system, comprising the following steps: S1: obtaining abnormal alarm information, historical operation and maintenance data and pipe network basic archives of a pipe network monitoring system, and constructing a work plan and resource allocation; S2: a pipe network robot inspects along a predetermined route, simultaneously collects multi-sensor data, and pre-processes; S3: based on the pre-processed multi-sensor data, based on a multi-modal leakage discovery model, a leakage candidate point list is obtained; S4: according to the leakage candidate list and confidence information, combined with the accurate pose data provided by the robot SLAM system, the accurate positioning and geometric reconstruction of each candidate point are carried out, and the accurate positioning result and three-dimensional geometric reconstruction model are obtained; S5: according to the accurate positioning result and three-dimensional geometric reconstruction model, combined with the plugging process database and the field constraint condition, the optimal plugging strategy is generated through a multi-objective optimization algorithm. The present application effectively improves the efficiency, accuracy and safety of underground pipe network leakage disposal.

Description

Technical Field

[0001] This invention relates to the field of underground pipeline repair, and in particular to an intelligent method and system for sealing leaks in underground pipelines based on robot learning. Background Technology

[0002] Underground pipeline networks (such as water supply, gas supply, heating, drainage, and industrial media pipelines) are typically located in confined spaces, humid environments, poor visibility, and complex media and noise interference conditions. They are susceptible to long-term corrosion, settlement, external disturbances, joint aging, and construction defects, making them prone to leaks, cracks, pitting corrosion, and loose joints. Leaks not only lead to resource loss and reduced operational efficiency but can also trigger major safety accidents such as ground subsidence, secondary pollution, and explosion risks. Due to the deep burial depth, wide coverage, and complex topology of underground pipeline networks, and the fact that they are often operated without interruption of supply or maintenance, the rapid detection, accurate location, and low-disturbance repair of leak points have always been core challenges in the operation and maintenance field. Summary of the Invention

[0003] To address the aforementioned problems, the present invention aims to provide an intelligent method and system for sealing leaks in underground pipelines based on robot learning, which effectively improves the efficiency and reliability of sealing leaks in underground pipelines.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A machine learning-based intelligent leak sealing method for underground pipelines includes the following steps: S1: Obtain abnormal alarm information, historical operation and maintenance data, and basic pipeline archives from the pipeline monitoring system, and construct work plans and resource configurations; S2: Based on the work plan and resource allocation, the pipeline robot performs inspections along a predetermined route, while collecting data from multiple sensors and preprocessing it. S3: Based on the preprocessed multi-sensor data and the multi-modal leak detection model, obtain a list of leak candidate points; S4: Based on the leak candidate list and confidence information, combined with the precise pose data provided by the robot SLAM system, each candidate point is precisely located and geometrically reconstructed to obtain the precise location results and three-dimensional geometric reconstruction model. S5: Based on the precise positioning results and the three-dimensional geometric reconstruction model, combined with the sealing process database and on-site constraints, the optimal sealing strategy is generated through a multi-objective optimization algorithm.

[0005] Furthermore, abnormal alarm information, historical operation and maintenance data, and basic pipeline archives from the pipeline monitoring system are obtained, and work plans and resource configurations are constructed, as follows: Obtain real-time abnormal alarm information from the pipeline network SCADA monitoring system, and extract historical operation and maintenance data of relevant pipeline sections as well as basic pipeline network information; Through data cleaning and standardization, false alarms and redundant information are eliminated, a complete data profile of the target area is established, and based on the frequency, intensity and duration characteristics of abnormal alarms, combined with the statistical patterns of similar historical failures and pipeline aging models, the preliminary probability and risk level of leakage are calculated to obtain preliminary risk assessment results. Based on the preliminary risk assessment results, and taking into account the geometric characteristics, environmental constraints, and safety requirements of the target pipe section, a work plan is generated. Construct a hierarchical risk control matrix, and based on the hazard of the medium, pressure level, and environmental sensitivity, clarify the operational constraints, safety protection requirements, emergency evacuation triggering conditions, and response measures under different risk levels.

[0006] Furthermore, by combining the statistical patterns of similar historical failures and pipeline aging models, the preliminary probability and risk level of leakage are calculated, yielding preliminary risk assessment results as follows: Extract multi-dimensional feature vectors from the cleaned alarm data, including the time features, intensity features, spatial features, and correlation features of abnormal alarms; By analyzing historical data on similar failures using big data mining techniques, a failure feature pattern library is constructed to identify typical combinations of leakage symptoms and establish feature fingerprint spectra for different defect types. Combined with pipeline aging models, the degree of deterioration and baseline failure probability of the target pipe section are analyzed. A leakage probability calculation model is established using a Bayesian inference framework. Historical fault statistics are used as prior probabilities, and current abnormal features are used as observational evidence. The posterior probability of leakage is obtained by updating the likelihood function. The probability calculation comprehensively considers the weighted contribution of multiple factors, including the matching degree of abnormal alarms, the degree of pipeline aging, the impact of operating conditions, and external environmental factors. Based on the probability calculation, combined with the severity assessment of leakage consequences, a risk level assessment matrix is ​​established, and the initial risk is divided into four levels.

[0007] Furthermore, the multi-sensor data includes data from visual sensors, point cloud sensors, and acoustic sensors. The visual sensors continuously acquire images to obtain visual clues such as cracks, corrosion, deformation, and water seepage traces. The point cloud sensors acquire three-dimensional point clouds of the inner wall of the pipe to obtain the geometry of pipe diameter changes, ellipticization, misalignment, settlement, obstacles, and attachments. The acoustic sensors acquire raw waveforms at a high sampling rate to capture broadband noise, turbulent sound, and their energy distribution as a function of location caused by the leak.

[0008] Furthermore, based on the preprocessed multi-sensor data and a multi-modal leak detection model, a list of leak candidate points is obtained, as detailed below: Based on the preprocessed three-modal data stream, visual frame I k Point cloud block P kAcoustic waveform segment a k (t), map it to uniform mileage coordinates s using SLAM, and construct samples according to mileage windows: ; Among them, w i For the i-th mileage window; s i The mileage corresponding to the center of the window; W is the window length; x v (i) represents the visual modal input features of the i-th window; x p (i) represents the modal input features of the point cloud in the i-th window; x a (i) represents the acoustic modal input features of the i-th window; Φv, Φp, and Φa are the encoding mapping functions for the three modalities, which transform the original data set within the window into features usable by the model; s k Let {k} be the mileage position corresponding to the k-th sample; {·} represents the set of all samples within the window; Calculate the leakage probability p for each window i mode m. m (i)∈(0,1), m∈{v,p,a}; v,p,a represent the visual modality, point cloud modality, and acoustic modality, respectively, and a reliability q is calculated for each modality. m (i)∈[0,1], obtain the normalized weights: ; Among them, w m (i) represents the fusion weights of mode m in the i-th window; q m (i) represents the quality score of mode m in window i; It is a constant; We use weighted log-probability fusion to obtain the fusion confidence sequence P along the mileage window. leak (i): ; Where logit(·) is the log-odds function, mapping the probability p∈(0,1) to the real number domain; p is the event probability; L(i) is the fused log-odds score for the i-th window; b is the bias term; p m (i) The single-mode leakage prediction probability of mode m in window i; P leak (i) represents the fusion leakage probability of window i; σ(·) is the Sigmoid function; P leak (i) Consider the leakage probability thermodynamic curve to generate candidate point centers and their locations: ; For each candidate peak i Given the position range: ; ; Where I represents the candidate peak i The corresponding set of candidate leakage intervals For the i-th The central mileage location corresponding to each peak; These represent the minimum and maximum mileage, respectively; λ' is the contour scale factor; and s is the mileage coordinate.

[0009] Furthermore, a confidence map q(u,v) of leakage-related appearance cues is obtained using a defect detection network, and the ROIs are aggregated and mapped to probabilities: ; ; Where ROI is the suspected leakage region obtained within window i; (u,v) are pixel coordinates; rv(i) is the strength of the original visual evidence of window i; p v (i) represents the visual modal leakage probability of window i; α v θ is the slope coefficient; v The threshold for visual evidence; By fitting a cross-sectional circle onto the window point cloud, the ellipticity is obtained, the depression depth is calculated, and these are combined to form a point cloud anomaly score, which is then mapped to a probability. ; ; ; ; Where a(i) is the major axis obtained by fitting the pipe cross-section at window i; b(i) is the minor axis obtained by fitting the cross-section at window i; e(i) is the ellipticity / deformation index; the condition a≥b specifies the order of major and minor axes; P wi For window w i A subset of the point cloud; x is a 3D point in the point cloud; R fit (i) is the fitting radius of the pipe wall in window i; c(i) is the center point of the fitting circle axis to the cross section; Let x be the distance from point x to the center of the circle; d(i) be the maximum concavity depth of window i; step(i) be the step strength index of window i; r p (i) represents the original evidence score of the point cloud at window i; β1, β2, β3 are the weight coefficients of each geometric index; p p (i) represents the point cloud modal leakage probability of window i; α p θ is the slope coefficient of the point cloud fraction to probability mapping.p The threshold for point cloud score mapping; Perform STFT on the acoustic segment to obtain the power spectrum P(f,i), and define the leakage band energy ratio: ; Where P(f,i) is the acoustic power spectral density at frequency f of window i; f1 and f2 are the upper and lower limits of the calibrated leakage characteristic frequency band, respectively; f max ρ is the highest frequency in the spectral analysis; ρ(i) is the energy proportion of the leakage band; And add spectral flatness: ; Where SF(i) is the spectral flatness; P k (i) represents the spectral energy of window i at the k-th frequency point; K is the number of frequency points involved in the calculation; Acoustic evidence score: ; Where, r a (i) represents the acoustic raw evidence score of window ii; γ1 and γ2 are the weighting coefficients of ρρ and SF, respectively. Mapped to probability: ; Where, p a (i) represents the acoustic modal leakage probability of window i; α a θ is the slope coefficient; a The threshold value is used.

[0010] Furthermore, based on the leak candidate list and confidence information, combined with the precise pose data provided by the robot SLAM system, each candidate point is precisely located, as follows: For each leak candidate point c i Known: Mileage range Fusion confidence level P i SLAM provides the robot's pose at time t: ; Where W is the network map coordinate system and B is the robot base coordinate system. (t) represents the rigid body pose transformation from the B coordinate system to the W coordinate system at time t; SE(3) represents the three-dimensional rigid body motion group; R(t) represents the rotation matrix at time t; p(t) represents the translation vector at time t; It is a 1×3 zero vector; Map candidate mileage intervals to time periods: ; In Ti Internally, pixel observation is performed on the ROI output by the visual network. Given the camera intrinsic parameter K, the normalized line-of-sight direction is obtained: ; Transform it to world coordinates to obtain the ray: ; Where k is the index of the kth observed ray; t k To generate the timestamp for the kth observation; o W,k Let r be the starting point of the k-th ray in the world system W; W,k Let p(t) be the direction vector of the k-th ray in the world system W; k (t) represents time t k The location of the machine's origin in the world system; R(t) k (t) represents time t k Rotation from machine body to world system The position of the origin of the camera coordinate system C in the body coordinate system B Let C be the extrinsic rotation matrix from camera system C to machine system B; Let be the normalized imaging ray direction of the k-th ray in the camera system; Corresponding ray equation: ; in, Let λ be the parameterized line of the k-th ray in the world system; λ is the scale parameter along the ray direction. Simultaneously, a set P of local pipe wall points is extracted from the point cloud in the neighborhood at the same time. k Transfer the radar point from the radar system to the world system: ; Where, x L Let the coordinates of the point be in the lidar coordinate system L. The extrinsic pose of the lidar system L to the machine system B; x W Let W be the coordinates of the same point in the world system W. A local plane is fitted as an approximate surface using the local pipe wall point cloud. The plane form is: ; Where, π k The equation for the k-th plane; n k dk is the plane normal vector; x is an arbitrary point in space; dk is the signed distance parameter of the plane offset; Intersection of ray and plane: ; in, For the k-th ray and the plane π k Intersection parameters; It is the dot product of the ray direction and the plane normal; To obtain the three-dimensional coordinates of the intersection point; Using the nearest neighbor method from ray to point cloud: ; in, Let be the 3D matching point corresponding to the k-th ray; be the point cloud set in the world system; For the k-th ray; dist(x, From point x to the line The shortest Euclidean distance; For candidate point i, sum all single-frame estimates and use a weighted robust mean: ; Where, q k For data quality; w k P represents the weight of the k-th observation; i The prior weights associated with the i-th candidate leak segment; q k Let k be the quality factor of the kth observation itself; The final location was obtained through optimization: ; Where ρ(·) is the robust loss function; The location covariance can then be approximated by a weighted distribution: ; Where, Σ i Let ∑ be the weighted covariance matrix of the i-th localization result; k w k The weighted sum is used for normalization; the output is the precise localization result. With uncertainty Σ i .

[0011] Further geometric reconstruction is performed as follows: For each candidate point i, within its neighborhood mileage range Internal extraction of point cloud frames By utilizing SLAM poses, they are unified into the world system and integrated: ; Among them, P j Let x be the original point cloud set of the j-th frame in the radar system LL; L Let L be the coordinates of a point in the point cloud in the radar coordinate system L. To achieve the external parameter position of the radar Let time t j The body's position in the world; t j x is the timestamp corresponding to the j-th frame; W Let x be the coordinates of point x after transformation to the world system; P W,j Let j be the point set of the point cloud in the world system; Initial assembly: ; Among them, P W For all frames / global point cloud / map after transformation to world system; To perform a union over all j; To eliminate the small drift of SLAM and the sensor delay residual, registration fine-tuning is performed on adjacent frames, and the incremental value for each frame is calculated. : ; Among them, C j For the set of corresponding point pairs, C p C q For local covariance; ΔR j For ΔT j Rotation increment in; Δt j For ΔT j Translation increment in C; j Let be the set of correspondences for the j-th frame, obtained by nearest neighbor, feature matching, or projection association; (p,q) is a pair of matching points; p is the source point; q is the target point; The information matrix representing the uncertainty of synthesis; This indicates rotating the source point covariance to the target coordinate direction; Updated point cloud fusion: ; in, For P W,j Corrected point cloud after applying the optimal increment; The corrected global map is the result of merging all corrected point clouds; Using TSDF voxel fusion, the truncation distance d for voxel v in each frame j (v) and weight w j (v) Accumulation: ; Where D(v) is the fused depth value of voxel v; The local geometric model M is obtained by extracting the mesh at D(v)=0D using Marching Cubes and reconstructing it. i .

[0012] Furthermore, based on the precise positioning results and the three-dimensional geometric reconstruction model, combined with the sealing process database and on-site constraints, the optimal sealing strategy is generated through a multi-objective optimization algorithm, as follows: With the precise location results of each leak point and local three-dimensional geometric reconstruction model M i Based on this, a structured feature set for process matching is extracted within the defect neighborhood: defect type, size parameters, morphological parameters, and geometric boundaries related to construction; at the same time, on-site constraints and operational constraints are incorporated to form a constraint set for that point. A two-level search of the sealing process database is performed, consisting of hard constraint filtering and soft constraint scoring: First, schemes that do not meet the applicable range of medium and temperature, pipe diameter range, curing time and window mismatch, or insufficient equipment capacity are eliminated; then, they are sorted according to historical success rate, impact on traffic cross section, reversibility of subsequent maintenance, and stringency of surface treatment requirements, and finally, several feasible candidate process packages are generated for each leak point, and adjustable parameters are specified for each candidate process; The process selection and process parameters are modeled as a multi-objective optimization problem: under the premise of satisfying all hard constraints, the remaining risk after repair, cost, construction time, and disturbance to operation are minimized. The cost objective includes materials, equipment and manpower, transportation and re-inspection. The schedule objective consists of site preparation, actual construction, curing and waiting, and acceptance and re-inspection. The operational impact is measured by pressure drop, flow reduction or operational constraints during construction. The NSGA-II multi-objective evolutionary algorithm, which is suitable for the mixed optimization of discrete process ID and continuous parameters, is adopted, and feasible solutions are prioritized for constraints to obtain the optimized output.

[0013] The intelligent sealing system for underground pipeline leaks based on robot learning includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent sealing method for underground pipeline leaks based on robot learning as described above.

[0014] The present invention has the following beneficial effects: 1. This invention uses pipeline monitoring alarms, historical operation and maintenance data, and basic archives as task prerequisites, and incorporates scheduling into the work plan in advance, so that inspection is transformed from passive search to target-oriented; the robot executes according to the plan and collects multi-sensor data, and then the multi-modal leak detection model outputs a list of candidate points, compressing massive continuous inspection data into a small number of high-value suspicious points, significantly reducing the frequency of manual review and blind excavation; 2. This invention suppresses false alarms caused by single-frequency mechanical noise, stain reflection, and sparse point clouds through a confidence and quality weighting mechanism. In the localization and reconstruction stage, candidate points are aligned with the precise pose output by SLAM, suspected locations are upgraded to landable world coordinates and odometer information, and a local 3D geometric model is generated. 3. This invention can achieve robust leak detection based on multi-source monitoring and robot inspection data, achieve high-precision positioning and defect quantification with the help of SLAM and geometric reconstruction, and further combine sealing process knowledge and on-site constraints to generate sealing strategies through an optimizable decision-making method, so as to improve the efficiency, accuracy and safety of underground pipeline leak handling. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1 In this embodiment, a smart sealing method for underground pipeline leaks based on robot learning is provided, including the following steps: S1: Obtain abnormal alarm information, historical operation and maintenance data, and basic pipeline archives from the pipeline monitoring system, and construct work plans and resource configurations; S2: Based on the work plan and resource allocation, the pipeline robot performs inspections along a predetermined route, while collecting data from multiple sensors and preprocessing it. S3: Based on the preprocessed multi-sensor data and the multi-modal leak detection model, obtain a list of leak candidate points; S4: Based on the leak candidate list and confidence information, combined with the precise pose data provided by the robot SLAM system, each candidate point is precisely located and geometrically reconstructed to obtain the precise location results and three-dimensional geometric reconstruction model. S5: Based on the precise positioning results and the three-dimensional geometric reconstruction model, combined with the sealing process database and on-site constraints, the optimal sealing strategy is generated through a multi-objective optimization algorithm.

[0017] In this embodiment, abnormal alarm information, historical operation and maintenance data, and basic pipeline archives of the pipeline monitoring system are acquired, and work plans and resource configurations are constructed, as follows: Real-time abnormal alarm information is obtained from the pipeline SCADA monitoring system, including various alarm data such as pressure drop, abnormal flow, and user complaints. At the same time, historical operation and maintenance data of relevant pipeline sections (including historical fault records, maintenance history, material replacement records, and trends of operating parameters) and basic pipeline archive information (pipeline diameter, material type, burial depth, design pressure rating, service life, and geographical coordinates) are extracted. Through data cleaning and standardization, false alarms and redundant information are eliminated, a complete data profile of the target area is established, and based on the frequency, intensity and duration characteristics of abnormal alarms, combined with the statistical patterns of similar historical failures and pipeline aging models, the preliminary probability and risk level of leakage are calculated to obtain preliminary risk assessment results. Based on the preliminary risk assessment results, and taking into account the geometric characteristics (pipe diameter, elbow density, accessibility), environmental constraints (groundwater level, traffic conditions, surrounding buildings), and safety requirements (medium type, pressure level, surrounding population density) of the target pipeline section, an operation plan is generated. In this embodiment, geometric feature analysis extracts the precise geometric parameters of the target pipe section through the pipeline network GIS system: pipe diameter classification determines the robot platform selection (DN150-300mm uses an ultra-small tracked platform, DN300-800mm uses a standard wheeled platform, and DN800 and above can use a large tracked or walking platform); elbow density statistics assess the robot's passage difficulty and cumulative positioning error by the number of elbows per 100 meters (<2 is low density, 2-5 is medium density, >5 is high density); accessibility assessment calculates the straight-line distance from the nearest manhole to the target location, the actual pipeline path length, and the distribution of intermediate blockages (silt, foreign objects, gate valves, etc.). Environmental constraint analysis obtains key environmental parameters through on-site surveys and historical data: groundwater level is determined through drilling data or water level measurements in adjacent well chambers (0.5m above the bottom of the pipe indicates high risk, 0.5-2m indicates medium risk, and >2m indicates low risk), directly affecting the safety of downhole operations and the waterproofing requirements of equipment; traffic condition analysis includes surface road grades, traffic density, and control feasibility, determining equipment transportation routes and time windows for road occupancy operations; surrounding building density statistics include the distribution and distance of sensitive buildings such as residences, schools, and hospitals, affecting noise control standards and safety protection range; The work plan includes key elements such as robot platform selection (in-pipe inspection type for large-diameter accessible pipe sections, well chamber operation type for space-constrained scenarios, and ground support type for communication relay and material supply), estimated work duration, entry path planning, and personnel allocation plan; at the same time, based on information such as pipeline material, pressure level, and defect type prediction, the system pre-selects suitable material types from the sealing process database (injection-type two-component materials for rapid sealing of small defects, adhesive composite materials for large-area defects, and mechanical devices for temporary sealing), estimates material usage, configures special tools, and forms a complete resource allocation list. Construct a hierarchical risk control matrix, and based on the hazard of the medium (gas requires explosion and poison prevention, water supply requires pollution prevention, and heat requires scalding prevention), pressure level (high-pressure pipelines require pressure reduction operations, and medium and low pressure pipelines can be handled under pressure), and environmental sensitivity (sensitive areas such as schools and hospitals require special protection), clarify the operational constraints, safety protection requirements, emergency evacuation triggering conditions, and response measures under different risk levels.

[0018] In this embodiment, by combining the statistical patterns of similar historical failures and the pipeline aging model, the preliminary probability and risk level of leakage are calculated, and the preliminary risk assessment results are obtained as follows: Multi-dimensional feature vectors are extracted from the cleaned alarm data, including the time features of abnormal alarms (occurrence time, duration, interval period), intensity features (pressure drop magnitude, flow change rate, concentration detection value), spatial features (alarm point distribution, impact range, propagation direction), and correlation features (changes in other parameters during the same period, responses of adjacent pipe sections, and feedback from downstream users). By analyzing historical data on similar failures using big data mining techniques, a failure feature pattern library is constructed to identify typical combinations of leakage symptoms (such as leakage patterns of "continuous pressure drop + abnormal flow increase + insufficient downstream pressure"), and feature fingerprint spectra of different defect types are established. Combined with pipeline aging models (considering material corrosion rate, fatigue cumulative damage, environmental impact factors, etc.), the degree of deterioration and baseline failure probability of the target pipe section are analyzed. A leakage probability calculation model was established using a Bayesian inference framework. Historical fault statistics were used as prior probabilities, and current abnormal features were used as observational evidence. The posterior probability of leakage was obtained through likelihood function updates. The probability calculation comprehensively considered the weighted contributions of multiple factors, including the matching degree of abnormal alarms (similarity between current features and historical leakage patterns), pipeline aging degree (deterioration assessment based on service life, material properties, and environmental corrosion), operating condition impact (impact of pressure fluctuations, temperature cycles, and load changes on pipeline stress), and external environmental factors (construction disturbances, geological changes, seasonal impacts, etc.). Based on the probability calculation, combined with the severity assessment of leakage consequences (considering the hazard of the medium, sensitivity of the surrounding environment, and scale of economic losses), a risk level assessment matrix was established, classifying the initial risk into four levels: Level I (probability < 5% and minor consequences, enhanced monitoring recommended), Level II (5% ≤ probability < 20% or moderate consequences, inspection recommended within 1 month), Level III (20% ≤ probability < 50% or severe consequences, handling recommended within 1 week), and Level IV (probability ≥ 50% or extremely severe consequences, immediate response recommended).

[0019] In this embodiment, the multi-sensor data includes visual sensor, point cloud sensor, and acoustic sensor data. The visual sensor (in-pipe camera / surround view camera) continuously acquires images to obtain visual clues of cracks, corrosion, deformation, and water seepage traces. The point cloud sensor (LiDAR or structured light) acquires three-dimensional point clouds of the inner wall of the pipe to obtain the geometry of pipe diameter changes, ellipticization, misalignment, settlement, obstacles, and attachments. The acoustic sensor (array microphone / hydrophone) acquires raw waveforms at a high sampling rate to capture broadband noise, turbulent sound, and their energy distribution as a function of position caused by leakage.

[0020] In this embodiment, based on preprocessed multi-sensor data and a multi-modal leakage detection model, a list of candidate leakage points is obtained, as detailed below: Based on the preprocessed three-modal data stream, visual frame I k Point cloud block P k Acoustic waveform segment a k (t), map it to uniform mileage coordinates s using SLAM, and construct samples according to mileage windows: ; Among them, w i For the i-th mileage window; s i The mileage corresponding to the center of the window; W is the window length; x v (i) represents the visual modal input features of the i-th window (e.g., keyframes, video clip features, or their embeddings); x p (i) represents the point cloud modal input features for the i-th window (e.g., local point cloud sub-images, voxel / range image features, etc.); x a (i) represents the acoustic modal input features (e.g., Mel spectrum / time-frequency plot) of the i-th window; Φv, Φp, and Φa are the encoding mapping functions for the three modalities, transforming the original data set within the window into features usable by the model; s k This represents the mileage position corresponding to the k-th sample (obtained through SLAM / odometer / time alignment). {·} denotes the set of all samples within the window; Calculate the leakage probability p for each window i mode m. m (i)∈(0,1), m∈{v,p,a}; v,p,a represent the visual modality, point cloud modality, and acoustic modality, respectively, and a reliability q is calculated for each modality. m (i)∈[0,1], obtain the normalized weights: ; Among them, w m (i) represents the fusion weights of mode m in the i-th window; q m(i) is the quality score of mode m in window i; using visual sharpness / exposure, point cloud coverage / echo intensity, and acoustic SNR / noise level. It is a constant; The reliability calculation is as follows: visual modality is calculated using sharpness, exposure saturation, and occlusion ratio; point cloud modality is calculated using coverage, point density, and echo anomaly rate; acoustic modality is calculated using SNR, saturation, and background noise level. We use weighted log-probability fusion to obtain the fusion confidence sequence P along the mileage window. leak (i): ; Where logit(·) is the log-odds function, mapping the probability p∈(0,1) to the real number domain; p is the event probability (the predicted probability of a certain modality for "leakage"); L(i) is the fused log-odds score for the i-th window; b is the bias term; p m (i) The single-mode leakage prediction probability of mode m in window i; P leak (i) represents the fusion leakage probability of window i; σ(·) is the Sigmoid function; P leak (i) Consider the leakage probability thermodynamic curve to generate candidate point centers, and the candidate point locations (center mileage): ; For each candidate peak i Given the position range: ; ; Where I represents the candidate peak i The corresponding set of candidate leakage intervals For the i-th The central mileage location corresponding to each peak; These represent the minimum and maximum mileage, respectively; λ' is the contour scale factor; and s is the mileage coordinate.

[0021] In this embodiment, a defect detection network is used to obtain confidence maps q(u,v) of leakage-related visual clues (water marks, cracks, corrosion pits, spray atomization, etc.), and the ROIs are aggregated and mapped to probabilities: ; ; Where ROI is the suspected leakage region obtained within window i; (u,v) are pixel coordinates; rv(i) is the strength of the original visual evidence of window i; p v(i) represents the visual modal leakage probability of window i; α v θ is the slope coefficient; v The threshold for visual evidence; By fitting a cross-sectional circle onto the window point cloud, the ellipticity is obtained, the depression depth is calculated, and these are combined to form a point cloud anomaly score, which is then mapped to a probability. ; ; ; ; Where a(i) is the major axis obtained by fitting the pipe cross-section at window i; b(i) is the minor axis obtained by fitting the cross-section at window i; e(i) is the ellipticity / deformation index; the condition a≥b specifies the order of major and minor axes; P wi For window w i A subset of the point cloud; x is a 3D point in the point cloud; R fit (i) is the fitting radius of the pipe wall in window i; c(i) is the center point of the fitting circle axis to the cross section; Let x be the distance from point x to the center of the circle; d(i) be the maximum concavity depth of window i; step(i) be the step strength index of window i; r p (i) represents the original evidence score of the point cloud at window i; β1, β2, β3 are the weight coefficients of each geometric index; p p (i) represents the point cloud modal leakage probability of window i; α p θ is the slope coefficient of the point cloud fraction to probability mapping. p The threshold for point cloud score mapping; Perform STFT on the acoustic segment to obtain the power spectrum P(f,i), and define the leakage band energy ratio: ; Where P(f,i) is the acoustic power spectral density at frequency f of window i; f1 and f2 are the upper and lower limits of the calibrated leakage characteristic frequency band, respectively; f max ρ is the highest frequency in the spectral analysis; ρ(i) is the energy proportion of the leakage band; And spectral flatness is added (to suppress single-frequency mechanical noise): ; Where SF(i) is the spectral flatness; P k (i) represents the spectral energy of window i at the k-th frequency point; K is the number of frequency points involved in the calculation; Acoustic evidence score: ; Where, r a(i) represents the acoustic raw evidence score of window ii; γ1 and γ2 are the weighting coefficients of ρρ and SF, respectively. Mapped to probability: ; Where, p a (i) represents the acoustic modal leakage probability of window i; α a θ is the slope coefficient; a The threshold value is used.

[0022] In this embodiment, based on the leak candidate list and confidence information, and combined with the precise pose data provided by the robot SLAM system, each candidate point is precisely located, as follows: For each leak candidate point c i Known: Mileage range Fusion confidence level P i SLAM provides the robot's pose at time t: ; Where W is the network map coordinate system and B is the robot base coordinate system. (t) represents the rigid body pose transformation from coordinate system B to coordinate system W at time t; SE(3) represents the three-dimensional rigid body motion group; R(t) represents the rotation matrix at time t (rotating the vector in B to W); p(t) represents the translation vector at time t (the position of the origin of B in W); It is a 1×3 zero vector; Map candidate mileage intervals to time periods (using the mileage-time mapping s(t)): ; In T i Internally, pixel observations are extracted from the ROIs (such as seepage points and spray atomization areas) output by the visual network. Given the camera intrinsic parameter K, the normalized line-of-sight direction is obtained: ; Transform it to world coordinates to obtain the ray: ; Where k is the index of the kth observed ray; t k To generate the timestamp for the kth observation; o W,k Let r be the starting point of the k-th ray in the world system W; W,k Let p(t) be the direction vector of the k-th ray in the world system W; k (t) represents time t k The location of the machine's origin in the world system; R(t) k(t) represents time t k Rotation from machine body to world system The position of the origin of the camera coordinate system C in the body coordinate system B Let C be the extrinsic rotation matrix from camera system C to machine system B; Let be the normalized imaging ray direction of the k-th ray in the camera system (obtained by backprojection of pixel (uk,vk) through intrinsic parameters, usually taken as a unit vector); Corresponding ray equation: ; in, Let λ be the parameterized line of the k-th ray in the world system; λ is the scale parameter along the ray direction. Simultaneously, a set P of local pipe wall points is extracted from the point cloud in the neighborhood at the same time. k (Or use an existing local map) transfer the radar point from the radar system to the world system: ; Where, x L Let the coordinates of the point be in the lidar coordinate system L. The extrinsic pose of the lidar system L to the machine system B; x W Let W be the coordinates of the same point in the world system W. A local plane (or quadratic surface) is fitted to the local point cloud of the pipe wall as an approximate surface. The planar form is as follows: ; Where, π k The equation for the k-th plane; n k dk is the plane normal vector; x is an arbitrary point in space; dk is the signed distance parameter of the plane offset; Intersection of ray and plane: ; in, For the k-th ray and the plane π k Intersection parameters; It is the dot product of the ray direction and the plane normal; To obtain the three-dimensional coordinates of the intersection point; Using the nearest neighbor method from ray to point cloud: ; in, Let be the 3D matching point corresponding to the k-th ray; be the point cloud set in the world system; For the k-th ray; dist(x, From point x to the line The shortest Euclidean distance; For candidate point i, sum all single-frame estimates and use a weighted robust mean: ; Where, q k For data quality; w k P represents the weight of the k-th observation; i The prior weights associated with the i-th candidate leak segment; q k Let k be the quality factor of the kth observation itself; The final location was obtained through optimization: ; Where ρ(·) is the robust loss function; The location covariance can then be approximated by a weighted distribution: ; Where, Σ i Let ∑ be the weighted covariance matrix of the i-th localization result; k w k The weighted sum is used for normalization; the output is the precise localization result. With uncertainty Σ i .

[0023] In this embodiment, geometric reconstruction is performed as follows: For each candidate point i, within its neighborhood mileage range Internal extraction of point cloud frames By utilizing SLAM poses, they are unified into the world system and integrated: ; Among them, P j Let x be the original point cloud set of the j-th frame in the radar system LL; L Let L be the coordinates of a point in the point cloud in the radar coordinate system L. To achieve the external parameter position of the radar Let time t j The body's position in the world; t j x is the timestamp corresponding to the j-th frame; W Let x be the coordinates of point x after transformation to the world system; P W,j Let j be the point set of the point cloud in the world frame; Initial assembly: ; Among them, P W For all frames / global point cloud / map after transformation to world system; To perform a union over all j; To eliminate the small drift of SLAM and the sensor delay residual, registration fine-tuning is performed on adjacent frames, and the incremental value for each frame is calculated. : ; Among them, C j For the set of corresponding point pairs, C p C q For local covariance; ΔR j For ΔT j Rotation increment in; Δt j For ΔT j Translation increment in C; j Let be the set of correspondences (matching point pairs) for the j-th frame, obtained by nearest neighbor, feature matching, or projection association; (p,q) is a pair of matching points; p is the source point; q is the target point; The information matrix representing the uncertainty of synthesis; This indicates rotating the source point covariance to the target coordinate direction; Updated point cloud fusion: ; in, For P W,j Corrected point cloud after applying the optimal increment; The corrected global map is the result of merging all corrected point clouds; Using TSDF voxel fusion, the truncation distance d for voxel v in each frame j (v) and weight w j (v) Accumulation: ; Where D(v) is the fused depth value of voxel v; The local geometric model M is obtained by extracting the mesh at D(v)=0 using Marching Cubes and reconstructing it. i .

[0024] In this embodiment, based on the precise positioning results and the three-dimensional geometric reconstruction model, combined with the sealing process database and on-site constraints, the optimal sealing strategy is generated through a multi-objective optimization algorithm, as follows: With precise location results of each leak point and local three-dimensional geometric reconstruction model M iBased on this, a structured feature set for process matching is extracted within the defect neighborhood: defect type (pitting, crack, thinning, misalignment / ellipticization, composite damage), dimensional parameters (equivalent leak diameter, defect length / width / maximum depth, cross-sectional area loss rate), morphological parameters (local curvature, roughness, pipe wall ellipticization rate, step misalignment height), and geometric boundaries related to construction (operable space, minimum clearance, recommended range of target sealing length); simultaneously, on-site constraints and operational constraints (allowing uninterrupted supply / pressure reduction, maximum allowable pressure drop, operating window duration, environmental explosion-proof rating, humidity and pollution levels, ventilation conditions, robot accessibility and energy consumption limits, end effector capability boundaries, material arrival and personnel scheduling) are incorporated to form a constraint set for that point; A two-level search of the sealing process database is performed, involving hard constraint filtering and soft constraint scoring. First, solutions that do not meet the applicable range of medium and temperature, pipe diameter range, curing time and window mismatch, or insufficient equipment capacity are eliminated. Then, the solutions are ranked according to historical success rate, impact on traffic cross-section, reversibility of subsequent maintenance, and stringency of surface treatment requirements. Finally, several feasible candidate process packages are generated for each leak point (e.g., glue injection sealing, local lining / short section lining, mechanical expansion / internal support, composite material reinforcement, temporary leak sealing + permanent repair combination, etc.). Adjustable parameters (sealing length, thickness / number of layers, glue injection pressure and flow rate, curing time, construction posture and path, re-inspection method, etc.) are specified for each candidate process as decision variables for subsequent optimization. The process selection and process parameters are modeled as a multi-objective optimization problem: under the premise of satisfying all hard constraints, the remaining risk after repair, cost, construction time, and disturbance to operation are minimized. The cost objective includes materials, equipment and manpower, transportation and re-inspection. The schedule objective consists of site preparation, actual construction, curing and waiting, and acceptance and re-inspection. The operational impact is measured by pressure drop, flow reduction or operational constraints during construction. The NSGA-II multi-objective evolutionary algorithm, which is suitable for the mixed optimization of discrete process ID and continuous parameters, is adopted, and feasible solutions are prioritized for constraints to obtain the optimized output.

[0025] Preferably, the optimized output is typically a set of Pareto optimal plugging strategies (different solutions have different trade-offs between safety, cost, time, and operational impact), with secondary decisions made based on operational strategy preferences: for example, in a "safety first" scenario, the solution with the lowest remaining risk and the largest safety margin is prioritized; in a "non-stop operation first" scenario, the solution with the least operational impact and the ability to complete curing and acceptance within the specified window is prioritized; in a "cost-constrained" scenario, the lowest-cost solution is selected while meeting risk thresholds and regulatory constraints. Finally, for each leak point, an optimal plugging strategy that can be directly deployed and executed is output, including: the selected process item ID and applicable basis, plugging location (mileage / circumferential angle / spatial coordinates) and recommended plugging length range, key construction parameters (thickness / number of layers, injection pressure and flow rate, curing time, expansion force or reinforcement layer configuration), robot operation path and posture requirements, a list of required equipment and materials, construction time plan and personnel arrangement suggestions, and acceptance and re-inspection plans (such as post-plugging pressure regression testing, acoustic regression confirmation, and deviation thresholds between local 3D re-scanning and the pre-construction model). Meanwhile, the system forms a closed loop of "strategy-execution-re-inspection": if the re-inspection indicators fail to meet the standards, the system will automatically trace back the cause (parameter deviation, surface conditions not met, operating condition fluctuations exceeding limits, etc.), and after the constraints are updated, it will re-trigger the optimization to generate alternative solutions or secondary reinforcement solutions, ensuring that the entire process from positioning and assessment to sealing and disposal is controllable, traceable, and iterative.

[0026] The intelligent leak sealing system for underground pipelines based on robot learning includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent leak sealing method for underground pipelines based on robot learning as described above. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent plugging of leaks in underground pipe networks based on robot learning, characterized in that, Includes the following steps: S1: Obtain abnormal alarm information, historical operation and maintenance data, and basic pipeline archives from the pipeline monitoring system, and construct work plans and resource configurations; S2: Based on the work plan and resource allocation, the pipeline robot performs inspections along a predetermined route, while collecting data from multiple sensors and preprocessing it. S3: Based on the preprocessed multi-sensor data and a multi-modal leak detection model, a list of leak candidate points is obtained, as detailed below: Based on the preprocessed three-modal data stream, visual frame I k Point cloud block P k Acoustic waveform segment a k SLAM is used to map it to uniform mileage coordinates s, and samples are constructed according to mileage windows: ； Among them, w i For the i-th mileage window; s i The mileage corresponding to the center of the window; W is the window length; x v (i) represents the visual modal input features of the i-th window; x p (i) represents the modal input features of the point cloud in the i-th window; x a (i) represents the acoustic modal input features of the i-th window; Φv, Φp, and Φa are the encoding mapping functions for the three modalities, which transform the original data set within the window into features usable by the model; s k Let {k} be the mileage position corresponding to the k-th sample; {·} represents the set of all samples within the window; Calculate the leakage probability p for each window i mode m. m (i)∈(0,1), m∈{v,p,a}; v,p,a represent the visual modality, point cloud modality, and acoustic modality, respectively, and a reliability q is calculated for each modality. m (i)∈[0,1], obtain the normalized weights: ； Among them, w m (i) represents the fusion weights of mode m in the i-th window; q m (i) represents the quality score of mode m in window i; It is a constant; We use weighted log-probability fusion to obtain the fusion confidence sequence P along the mileage window. leak (i): ； Where logit(·) is the log-odds function, mapping the probability p∈(0,1) to the real number domain; p is the event probability; L(i) is the fused log-odds score for the i-th window; b is the bias term; p m (i) The single-mode leakage prediction probability of mode m in window i; P leak (i) represents the fusion leakage probability of window i; σ(·) is the Sigmoid function; P leak (i) Consider the leakage probability thermodynamic curve to generate candidate point centers and their locations: ； For each candidate peak i Given the position range: ； Where I represents the candidate peak i The corresponding set of candidate leakage intervals For the i-th The central mileage location corresponding to each peak; These represent the minimum and maximum mileage, respectively; λ' is the contour line scale factor. S4: Based on the leak candidate list and confidence information, combined with the precise pose data provided by the robot SLAM system, each candidate point is precisely located and geometrically reconstructed to obtain the precise location results and three-dimensional geometric reconstruction model. S5: Based on the precise positioning results and the three-dimensional geometric reconstruction model, combined with the sealing process database and on-site constraints, the optimal sealing strategy is generated through a multi-objective optimization algorithm.

2. The intelligent sealing method for underground pipeline leaks based on robot learning according to claim 1, characterized in that, The process of acquiring abnormal alarm information, historical operation and maintenance data, and basic pipeline archives from the pipeline monitoring system, and constructing work plans and resource configurations, is as follows: Obtain real-time abnormal alarm information from the pipeline network SCADA monitoring system, and extract historical operation and maintenance data of relevant pipeline sections as well as basic pipeline network information; Through data cleaning and standardization, false alarms and redundant information are eliminated, a complete data profile of the target area is established, and based on the frequency, intensity and duration characteristics of abnormal alarms, combined with the statistical patterns of similar historical failures and pipeline aging models, the preliminary probability and risk level of leakage are calculated to obtain preliminary risk assessment results. Based on the preliminary risk assessment results, and taking into account the geometric characteristics, environmental constraints, and safety requirements of the target pipe section, a work plan is generated. Construct a hierarchical risk control matrix, and based on the hazard of the medium, pressure level, and environmental sensitivity, clarify the operational constraints, safety protection requirements, emergency evacuation triggering conditions, and response measures under different risk levels.

3. The intelligent sealing method for underground pipeline leaks based on robot learning according to claim 2, characterized in that, By combining statistical patterns of similar historical failures and pipeline aging models, the preliminary probability and risk level of leakage are calculated, yielding preliminary risk assessment results, as follows: Extract multi-dimensional feature vectors from the cleaned alarm data, including the time features, intensity features, spatial features, and correlation features of abnormal alarms; By analyzing historical data on similar failures using big data mining techniques, a failure feature pattern library is constructed to identify typical combinations of leakage symptoms and establish feature fingerprint spectra for different defect types. Combined with pipeline aging models, the degree of deterioration and baseline failure probability of the target pipe section are analyzed. A leakage probability calculation model is established using a Bayesian inference framework. Historical fault statistics are used as prior probabilities, current abnormal features are used as observational evidence, and the posterior probability of leakage is obtained by updating the likelihood function. The probability calculation comprehensively considers the weighted contribution of multiple factors, including the matching degree of abnormal alarms, the degree of pipeline aging, the impact of operating conditions, and external environmental factors. Based on probability calculations and combined with the severity assessment of leakage consequences, a risk level assessment matrix is ​​established, dividing the initial risks into four levels.

4. The intelligent sealing method for underground pipeline leaks based on robot learning according to claim 1, characterized in that, The multi-sensor data includes data from visual sensors, point cloud sensors, and acoustic sensors. The visual sensors continuously acquire images to obtain visual clues such as cracks, corrosion, deformation, and water seepage traces. The point cloud sensors acquire three-dimensional point clouds of the inner wall of the pipe to obtain the geometric morphology of pipe diameter changes, ellipticization, misalignment, settlement, obstacles, and attachments. The acoustic sensors acquire raw waveforms at a high sampling rate to capture broadband noise, turbulent sound, and their energy distribution as a function of location caused by leakage.

5. The intelligent sealing method for underground pipeline leaks based on robot learning according to claim 1, characterized in that, A confidence map q(u,v) of leakage-related appearance cues is obtained using a defect detection network. The ROIs are then aggregated and mapped to probabilities. ； ； Where ROI(i) is the suspected leakage region obtained within window i; (u,v) are the pixel coordinates; r v (i) represents the strength of the original visual evidence for window i; p v (i) represents the visual modal leakage probability of window i; α v θ is the slope coefficient; v The threshold for visual evidence; By fitting a cross-sectional circle onto the window point cloud, the ellipticity is obtained, the depression depth is calculated, and these are combined to form a point cloud anomaly score, which is then mapped to a probability. ； ； ； ； Where a(i) is the major axis obtained by fitting the pipe cross section at window i; b(i) is the minor axis obtained by fitting the cross section at window i; e(i) is the ellipticity rate or deformation index; the condition a≥b specifies the order of major and minor axes. For window w i A subset of the point cloud within; x is any point in space; R fit (i) is the fitting radius of the pipe wall in window i; c(i) is the center point of the fitting circle axis to the cross section; Let x be the distance from any spatial point x to the center of the circle; d(i) is the maximum concavity depth of window i; step(i) is the step strength index of window i; r p (i) represents the original evidence score of the point cloud at window i; β1, β2, β3 are the weight coefficients of each geometric index; p p (i) represents the point cloud modal leakage probability of window i; α p θ is the slope coefficient of the point cloud fraction to probability mapping. p The threshold for point cloud score mapping; Perform STFT on the acoustic segment to obtain the power spectrum P(f,i), and define the leakage band energy ratio: ； Where P(f,i) is the acoustic power spectral density at frequency f of window i; f1 and f2 are the upper and lower limits of the calibrated leakage characteristic frequency band, respectively; f max ρ is the highest frequency in the spectral analysis; ρ(i) is the energy proportion of the leakage band; And add spectral flatness: ； Where SF(i) is the spectral flatness; P k (i) represents the spectral energy of window i at the k-th frequency point; K is the number of frequency points involved in the calculation; Acoustic evidence score: ； Where, r a (i) represents the acoustic raw evidence score for window i; γ1 and γ2 are the weighting coefficients of ρ(i) and SF(i), respectively. Mapped to probability: ； Where, p a (i) represents the acoustic modal leakage probability of window i; α a θ is the slope coefficient; a The threshold value is used.

6. The intelligent sealing method for underground pipeline leaks based on robot learning according to claim 1, characterized in that, The process involves precisely locating each candidate point based on the leak candidate list and confidence information, combined with the accurate pose data provided by the robot SLAM system, as detailed below: For each leak candidate point c i Known: Mileage range Fusion confidence level P i SLAM provides the robot's pose at time t: ； Where W is the network map coordinate system and B is the robot base coordinate system. (t) represents the rigid body pose transformation from the B coordinate system to the W coordinate system at time t; SE(3) represents the three-dimensional rigid body motion group; R(t) represents the rotation matrix at time t; p(t) represents the translation vector at time t; It is a 1×3 zero vector; Map candidate mileage intervals to time periods: ； In T i Internally, pixel observation is performed on the ROI output by the visual network. Given the camera intrinsic parameter K, the normalized line-of-sight direction is obtained: ； Transform it to world coordinates to obtain the ray: ； Where k is the index of the kth observed ray; t k To generate the timestamp for the kth observation; o W,k Let r be the starting point of the k-th ray in the world system W; W,k Let p(t) be the direction vector of the k-th ray in the world system W; k (t) represents time t k The location of the machine's origin in the world system; R(t) k (t) represents time t k Rotation from machine body to world system The position of the origin of camera coordinate system C in body coordinate system B Let C be the extrinsic rotation matrix from camera system C to machine system B; Let be the normalized imaging ray direction of the k-th ray in the camera system; Corresponding ray equation: ； in, Let λ be the parameterized line of the k-th ray in the world system; λ is the scale parameter along the ray direction. Simultaneously, a set P of local pipe wall points is extracted from the point cloud in the neighborhood at the same time. k Transfer the radar point from the radar system to the world system: ； Where, x L Let the coordinates of the point be in the lidar coordinate system L. The extrinsic pose of the lidar system L to the machine system B; x W Let W be the coordinates of the same point in the world system W. A local plane is fitted as an approximate surface using the local pipe wall point cloud. The plane form is: ； Where, π k The equation for the k-th plane; n k d is the plane normal vector; x is any point in space; d k The signed distance parameter for plane offset; Intersection of ray and plane: ； in, For the k-th ray and the plane π k Intersection parameters; It is the dot product of the ray direction and the plane normal; To obtain the three-dimensional coordinates of the intersection point; Using the nearest neighbor method from ray to point cloud: ； in, For point clouds and point sets within the world system; For the k-th ray; dist(x, For any point x in space, the line from x to x is... The shortest Euclidean distance; For candidate point i, sum all single-frame estimates and use a weighted robust mean: ； Where, q k For data quality; w k P represents the weight of the k-th observation; i The prior weights associated with the i-th candidate leak segment; q k Let k be the quality factor of the kth observation itself; Optimization yields precise positioning results: ； Where ρ(·) is the robust loss function; The location covariance is then approximated by the weighted dispersion: ； Where, Σ i Let ∑ be the weighted covariance matrix of the i-th localization result; k w k The weighted sum is used for normalization; the output is the precise localization result. With the weighted covariance matrix Σ i .

7. The intelligent sealing method for underground pipeline leaks based on robot learning according to claim 6, characterized in that, The geometric reconstruction is as follows: For each candidate point i, within its neighborhood mileage range Internal extraction of point cloud frames By utilizing SLAM poses, they are unified into the world system and integrated: ； Among them, P j Let x be the original point cloud set of the j-th frame in radar coordinate system L; L Let L be the coordinates of a point in the point cloud in the radar coordinate system L. To achieve the external parameter position of the radar Let time t j The body's position in the world; t j x is the timestamp corresponding to the j-th frame; W Let x be the coordinates of point x after transformation to the world system; P W,j Let j be the point set of the point cloud in the world frame; Initial assembly: ； Among them, P W A global point cloud map after all frames have been transformed to the world frame; To perform a union over all j; To eliminate the small drift of SLAM and the sensor delay residual, registration fine-tuning is performed on adjacent frames, and the incremental value for each frame is calculated. : ； Among them, C j For the set of corresponding point pairs, C p C q For local covariance; ΔR j For ΔT j Rotation increment in; Δt j For ΔT j Translation increment in C; j Let be the set of correspondences for the j-th frame, obtained by nearest neighbor, feature matching, or projection association; (p,q) is a pair of matching points; p is the source point; q is the target point; The information matrix representing the uncertainty of synthesis; This indicates rotating the source point covariance to the target coordinate direction; Updated point cloud fusion: ； in, For P W,j Corrected point cloud after applying the optimal increment; The corrected global map is the result of merging all corrected point clouds; Using TSDF voxel fusion, the truncation distance d for voxel v in each frame j (v) and weight w j (v) Accumulation: ; Where D(v) is the fused depth value of voxel v; The local geometric model M is obtained by extracting the mesh at D(v)=0D using Marching Cubes and reconstructing it. i .

8. The intelligent sealing method for underground pipeline leaks based on robot learning according to claim 7, characterized in that, Based on the precise positioning results and the three-dimensional geometric reconstruction model, combined with the sealing process database and on-site constraints, the optimal sealing strategy is generated through a multi-objective optimization algorithm, as follows: With precise location results of each leak point and local three-dimensional geometric reconstruction model M i Based on this, a structured feature set for process matching is extracted within the defect neighborhood: defect type, size parameters, morphological parameters, and geometric boundaries related to construction; at the same time, on-site constraints and operational constraints are incorporated to form a constraint set for that point. A two-level search of the sealing process database is performed, consisting of hard constraint filtering and soft constraint scoring: First, schemes that do not meet the applicable range of medium and temperature, pipe diameter range, curing time and window mismatch, or insufficient equipment capacity are eliminated; then, they are sorted according to historical success rate, impact on traffic cross section, reversibility of subsequent maintenance, and stringency of surface treatment requirements, and finally, several feasible candidate process packages are generated for each leak point, and adjustable parameters are specified for each candidate process; The process selection and process parameters are modeled as a multi-objective optimization problem: under the premise of satisfying all hard constraints, the remaining risk after repair, cost, construction time and disturbance to operation are minimized in a comprehensive manner. Cost targets include materials, equipment and manpower, transportation and re-inspection; The project schedule consists of site preparation, actual construction, curing and waiting, and acceptance and re-inspection; the operational impact is measured by pressure drop, flow reduction or operational restrictions during construction. The NSGA-II multi-objective evolutionary algorithm, which is suitable for the mixed optimization of discrete process ID and continuous parameters, is adopted, and feasible solutions are prioritized for constraints to obtain the optimized output.

9. A smart leak sealing system for underground pipelines based on robot learning, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent sealing method for underground pipeline leaks based on robot learning as described in any one of claims 1-8.

Images