An ai-driven adaptive pipe repair robot system and method

Abstract

This invention relates to an AI-driven adaptive pipeline repair robot system and method, comprising a data processing module, a defect understanding and parametric modeling unit, a fusion localization module, an intelligent decision-making and task graph generation unit, and a control execution unit. The data processing module preprocesses multimodal sensor data. The defect understanding and parametric modeling unit identifies pipeline defect information. The fusion localization module anchors the location of pipeline defects. The intelligent decision-making and task graph generation unit, based on pipeline defect information and anchor point localization, obtains the optimal process and executable flow under rule constraints and constructs a task graph. The control execution unit manages state transitions, timeouts, retry counts, and backtracking branch triggers, while simultaneously executing safety interlocks and converting the defect area on the anchor point localization unfolded surface into an executable path, which is then output to the repair robot. This invention effectively improves the efficiency and reliability of trenchless repair operations.

Description

Technical Field

[0001] This invention relates to the field of pipeline repair, and more particularly to an AI-driven adaptive pipeline repair robot system and method. Background Technology

[0002] As urban underground pipe networks continue to expand, drainage, combined sewer, and industrial transport pipelines are subjected to a combination of factors, including corrosion, sedimentation, foundation settlement, misaligned joints, and external loads, making them prone to defects such as cracks, holes, corrosion thinning, and joint leakage. If these defects develop, they can lead to risks such as leakage, collapse, road subsidence, and secondary pollution, putting pressure on municipal operational safety and maintenance costs. Therefore, frequent pipeline inspections and timely repairs are necessary.

[0003] Current trenchless repair operations typically consist of "manual interpretation of inspection results—experience-based process selection—on-site teaching / manual equipment control—post-inspection," and generally suffer from several problems: First, inspection data comes from diverse sources (video, laser, thickness measurement, acoustics, etc.), lacking unified synchronization, calibration, and fusion processing, leading to unstable defect identification and location results; Second, defect interpretation often remains at the category or general range level, lacking parameterized descriptions that can be directly used for process calculations, with process parameters relying on experience and resulting in poor quality consistency; Third, in-pipe positioning is easily affected by mileage drift, pipe diameter changes, and obstructions, making it difficult to return to the same defect location for comparison and re-inspection before and after repair, hindering acceptance, evidence collection, and traceability; Fourth, uncertain on-site environment and equipment status (excessive gas levels, insufficient lighting, surface contamination, tool blockage, etc.) easily cause task interruptions or rework, and there is a lack of standardized anomaly handling and rollback procedures. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide an AI-driven adaptive pipeline repair robot system and method, which effectively improves the efficiency, reliability, and acceptability of trenchless repair operations.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An AI-driven adaptive pipeline repair robot system includes a data processing module, a defect understanding and parametric modeling unit, a fusion localization module, an intelligent decision-making and task graph generation unit, and a control execution unit. The data processing module preprocesses multimodal sensor data acquired by the inspection robot to obtain processed multimodal sensor data. The defect understanding and parametric modeling unit identifies pipeline defect information based on the multimodal sensor data and converts each defect information into a structured data object. The fusion localization module anchors the pipeline defect location. The intelligent decision-making and task graph generation unit, based on the pipeline defect information and anchor point localization, obtains the optimal process and executable flow under rule constraints and constructs a task graph. The control execution unit schedules nodes according to the task graph, manages state transitions, timeouts, retry counts, and backtracking branch triggers, while simultaneously executing safety interlocks and converting the defect area on the anchor point localization unfolding surface into an executable path, which is then output to the repair robot.

[0006] Furthermore, the data processing module receives multimodal sensor data streams collected by the inspection robot within the pipeline, including visual images, laser line scan data, IMU odometer data, and ultrasonic thickness measurement, acoustic, and gas environment data. In the temporal dimension, the data processing module performs unified timestamp correction and alignment on the data from each sensor, forming synchronous frame packets at a fixed frequency or event-triggered method to ensure that images, point clouds, and odometer information are correlated at the same time. In the spatial dimension, the data processing module loads and manages camera intrinsic parameters, camera-line scan extrinsic parameters, and IMU-machine extrinsic calibration parameters, and performs image processing... The system performs distortion correction and geometric correction, coordinate transformation and scan line resampling on the point cloud to map different modes to a unified coordinate system. Regarding signal quality, the data processing module denoises, deblurs, enhances, equalizes brightness, and suppresses reflections in visual data; performs outlier removal, intensity normalization, and surface smoothing on the point cloud; performs zero-bias compensation and abnormal impulse suppression on the IMU; performs echo gating, coupling quality discrimination, and thickness curve smoothing on ultrasonic thickness measurements; performs bandpass filtering, spectral feature extraction, and noise floor estimation on acoustic data; and performs drift correction and threshold consistency checks on gas and environmental data.

[0007] Furthermore, the defect understanding and parametric modeling unit identifies pipeline defect information based on multimodal sensor data, as detailed below: Define time t k The synchronization frame packet is: ; Among them, I k The corrected image; Π k For imaging parameters; Pk={p j} represents a point cloud; For thickness grids; e is the acoustic feature vector; k The environment state is represented by Δuk, which represents the odometer and attitude increments. Δu,k The uncertainty is... For a point cloud point p=(x,y,z), define the unfolded coordinates (s,θ) with the pipe axis as x and the cross-sectional plane as yz: ; Radius r and topographic relief Defined as: ; Where R0 is the radius inside the tube, the image pixel u=(u,v) is first back-projected to obtain the three-dimensional point, and then mapped to (s,θ) in the same way; Discretize the unfolded surface into a mesh to construct a multi-channel feature map: X(s,θ)=[X I ,X δr ,X th ,X ac ,X e ]; Among them, X I ,X δr ,X th ,X ac ,X e These are visual texture features, radial topography, thickness, acoustic features, and environmental channels; The probability map P of each type of defect is output through visual segmentation. c (s,θ): P c (s,θ)=σ(f c (X I (s,θ))),c∈{cracks, holes, corrosion, misalignment, deposition}; Where fc(·) is the segmentation network and σ is the sigmoid function; Geometric Anomaly Score A g (s,θ) is obtained from the local statistics of δr: ; Where, μ N ,σ N δr(s,θ) represents the mean and standard deviation of the neighborhood N; δr(s,θ) represents the radial offset. It is a constant; Thickness Anomaly Score A t (s,θ): ; Where t0 is the nominal wall thickness or healthy baseline, and t(s,θ) comes from the thickness grid Z. th; Fusing multimodal evidence into a posterior evidence of defect existence: logit P(def|s,θ)=w I logit P c (s,θ)+w g A g (s,θ)+w t A t (s,θ)+w ac A ac (s,θ); ; Among them, w I ,w g ,w t ,w ac These correspond to the credibility weights of the visual, geometric, thickness, and acoustic channels, respectively; A ac Score acoustic anomalies; logit(·) is the logarithmic probability function; For fusion posterior probabilities; Connectivity extraction is performed on P(def|s,θ) to obtain candidate regions Ω. i .

[0008] Furthermore, each defect information is converted into a structured data object, as follows: For each cross-sectional point set {(y) at each position s along the path j ,z j The nominal radius R and center (y0, z0) are obtained by fitting a circle. The circle fitting uses least squares. ; The cross-sectional deformation index is expressed as roundness error E(s): ; Where Varj(·) represents the variance calculated for all circumferential measurement points j of the cross section; Interface misalignment is detected using abrupt changes in radius along the connection direction: ; Among them, S step R(s) is the radius step; R(s) is the fitted radius at position s along the path; Δs is the difference step size along the path. For candidate region Ω i Calculate geometric parameters: ; ; ; Among them, L i For length, Wi For width, A i For area, D i For depth, η i For the thinning rate; Environmental and material related parameters from e k Based on the surface assessment model, including humidity index h, gas exceedance marker g, and surface deposit level a, a repairability judgment is given: ; Where α1, α2, α3, and α4 are weighting coefficients; 1[h>h0] indicates height anomaly, which is 1 when the defect height exceeds the threshold h0; 1[a>a0] indicates area anomaly, which is 1S when the defect area exceeds the threshold a0. i For comprehensive severity scoring; The output confidence level can be taken as the posterior mean within the region and quality gating should be considered. ; Where |Ωi| represents the area of ​​the defect region; γ(q) is the quality modulation factor, and q is the data quality index; C i The average confidence level for the region; γ(q)∈(0,1] is a discount factor determined by data quality q. The final structured defect object is defined as:

[0009] Among them, c i Defect category; For the bounding box along the process; For circumferential bounding box; m i The material state; evidence is the chain of evidence.

[0010] Furthermore, the fusion positioning module receives the mileage and attitude increments and their uncertainty estimates from the synchronization frame packets, the corrected image and imaging parameters, point cloud data, and the defect object output by the defect understanding and parameterized modeling unit, and maintains the anchor point set A={a k}, each anchor point a k Includes anchor point type, feature descriptor, and mileage estimation at the anchor point. With covariance For each defective object d i The fusion positioning module first calculates its center along the path. and the surrounding center Then select the nearest anchor point a. As a reference anchor point, output the anchor point positioning result. .

[0011] Furthermore, the anchor point construction is as follows: During the inspection process, the inspection robot performs anchor point candidate detection on the image and point cloud: On the image side, features such as circumferential stripes and marking patterns are extracted and descriptors are formed; on the point cloud side, geometric features such as misaligned steps, diameter changes, and bifurcations are extracted through cross-section fitting and abrupt change detection along the path; and a candidate set is constructed by combining the motion priors provided by mileage and attitude increments; then, the similarity between the candidate and historical anchor points is calculated using the mileage prediction interval as a spatiotemporal gate, and the anchor point observation is confirmed through threshold and multi-frame consistency checks; once a match is successful, the module uses the anchor point observation as a constraint to update the mileage drift, thereby aligning the along-path coordinates with the anchor point table and reducing cumulative errors.

[0012] Furthermore, in the process decision-making stage, the intelligent decision-making and task graph generation unit selects a set of candidate processes from the process library based on defect type, size parameters, and material compatibility. The objective function includes expected repair quality, operation time cost, resource consumption, and operation risk. The constraints include hard constraints and soft constraints. The constraint satisfaction problem CSP solver is used for optimization and solution, and the optimal process selection, specific parameter configuration, and operation priority ranking are output for each defect.

[0013] Furthermore, in the task graph generation stage, the intelligent decision-making and task graph generation unit converts the process decision results into a directed acyclic graph containing nodes and edges. Node types include condition check nodes, execution action nodes, decision branch nodes, and synchronization nodes. Edges represent the dependencies between nodes, triggering conditions, and data flows, and carry metadata such as timeout thresholds, maximum number of retries, and fallback paths in case of failure. The task graph also embeds safety interlocking logic: a safety check is inserted before each execution node, and once triggered, a forced jump to a safety fallback branch is made; and quality inspection and exception handling branches are set after key process nodes. The final output task graph contains a complete set of nodes, a set of edges, initial and target state definitions, and interface agreements with the anchor point positioning module.

[0014] Furthermore, the control execution unit receives the task graph G=(V,E) output by the intelligent decision-making and task graph generation unit, the anchor point localization results output by the fusion localization module, and the defect area information output by the defect understanding unit. Simultaneously, it subscribes to robot body sensing and execution feedback. The control execution unit maintains the running state and running context of each node. During scheduling, it determines whether a node is enabled based on the task graph dependencies and the node's predicate: when all predecessor nodes of a node are SUCCESS and its predicate check condition is true, it is set to RUNNING and the corresponding control primitive is issued. During node execution, it continuously monitors sensor feedback and process variables; when the completion criterion is met, it is set to SUCCESS and the successor node is advanced along the edge that meets the trigger condition. For timeout management, if the node's runtime exceeds its threshold τ... v or side window τe If violated, it is judged as FAIL and an exception branch is triggered: when k <K v When the task graph is entered, the RETRY edge is entered and the predefined retry preparation actions are executed, while k k+1; when k=K v If consecutive failures reach the policy threshold, the system will enter the backtracking subgraph along the FALLBACK edge and write the failure reason, evidence chain, and node log into the task execution record for subsequent acceptance and traceability.

[0015] Furthermore, the control execution unit performs safety interlock gating at the entry points of all execution action nodes, calculates safety predicates in real time, and immediately preemptively terminates the current action once any interlock is triggered, forcibly jumping to the safe retreat branch in the task diagram to perform actions such as stopping, disabling tools, and retreating to the nearest safe anchor point, until the safety predicate is restored before allowing re-entry into the task diagram; the control execution unit is also responsible for positioning the defect in the area Ω=[s] on the anchor point location unfolding surface. min ,s max ]×[θ min ,θ max Converting to a robot executable path: First, determine the coverage strategy based on process parameters, generate a coverage trajectory Γ={(s(t),θ(t))} on the unfolded surface, and map it to the absolute path s=s(a) according to the anchor points. )+Δs; then map the (s,θ) trajectory to the three-dimensional target point sequence p(t) in the pipeline coordinate system to form a path segment that can be tracked by the underlying controller; finally, the control execution unit outputs the current task graph node ID, path segment p(t), trajectory parameters, process control parameters and safety constraints as a unified instruction to the repair robot for execution, and writes the execution back to the task graph context.

[0016] A control method for an AI-driven adaptive pipeline repair robot system includes the following steps: S1: Preprocess the multimodal sensor data acquired by the inspection robot to obtain the processed multimodal sensor data; S2: Based on multimodal sensor data, identify pipeline defect information and convert each defect information into a structured data object; S3: Anchor point location for pipeline defects; S4: Based on pipeline defect information and anchor point location, obtain the optimal process and executable flow under rule constraints, and construct a task graph; S5: Schedule nodes according to the task graph, manage state transitions, timeouts, retry counts, and backtracking branch triggers, while executing safety interlocks, converting the defect area on the anchor point positioning unfolding surface into an executable path, and outputting it to the repair robot.

[0017] The present invention has the following beneficial effects: 1. This invention unifies, synchronizes, calibrates, and assesses the quality of multimodal raw data to ensure stable and reliable downstream input. It also generates structured objects containing multidimensional parameters such as geometric dimensions, surface condition, material degradation, and environmental factors through the quantification of defect information. This reduces reliance on manual experience and repeated trial and error on-site, and improves the consistency and reproducibility of defect identification, process selection, and repair actions. 2. This invention binds the defect location to a unified coordinate system through anchor point positioning, which significantly suppresses the repeated positioning error caused by long-distance mileage drift and ensures that the same position can be returned for comparison and re-inspection before and after repair. 3. Under rule constraints, the intelligent decision-making and task graph generation unit of this invention outputs the optimal process and executable flow, and solidifies the complete logic in the form of a task graph; it schedules nodes according to the task graph, and makes safety events interlocked, which can automatically degrade, pause or roll back in abnormal situations, reducing the probability of accidents and the interruption rate of tasks; it automatically converts the area of ​​defects on the anchor point positioning unfolding surface into an executable path and sends it to the robot, improving efficiency and consistency, and accumulating high-quality process data for continuous learning and parameter optimization. Attached Figure Description

[0018] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation

[0019] 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, an AI-driven adaptive pipeline repair robot system is provided, including a data processing module, a defect understanding and parametric modeling unit, a fusion positioning module, an intelligent decision-making and task graph generation unit, and a control execution unit. The data processing module preprocesses the multimodal sensor data acquired by the inspection robot to obtain processed multimodal sensor data. The defect understanding and parametric modeling unit identifies pipeline defect information based on the multimodal sensor data and converts each defect information into a structured data object containing multidimensional parameters of geometry, material, and environment. The fusion positioning module performs anchor point positioning of the pipeline defect location. The intelligent decision-making and task graph generation unit obtains the optimal process and executable flow under rule constraints based on the pipeline defect information and anchor point positioning, and constructs a task graph. The control execution unit schedules nodes according to the task graph, manages state transitions, timeouts, retry counts, and backtracking branch triggers, while executing safety interlocks (gas over-limit, emergency stop, mooring abnormality), and converts the area of ​​the defect on the anchor point positioning unfolding surface into an executable path, which is then output to the repair robot.

[0020] In this embodiment, the data processing module receives multimodal sensor data streams collected by the inspection robot within the pipeline, including visual images, laser line scan data, IMU mileage data, and ultrasonic thickness measurement, acoustic, and gas environment data. In the time dimension, the data processing module performs unified timestamp correction and alignment on the data from each sensor, forming synchronization frames at a fixed frequency or event-triggered method to ensure that images, point clouds, and mileage information are correlated at the same time. In the spatial dimension, the data processing module loads and manages camera intrinsic parameters, camera-line scan extrinsic parameters, and IMU-machine extrinsic calibration parameters, performs distortion correction and geometric correction on the images, and performs coordinate transformation and scan line resampling on the point clouds, enabling different modalities to be mapped to a unified coordinate system (e.g., mileage along the pipeline s). (The data processing module performs noise reduction, deblurring enhancement, brightness equalization, and reflection suppression on visual data; outlier removal, intensity normalization, and surface smoothing on point clouds; zero-bias compensation and anomalous impulse suppression on IMUs; echo gating, coupling quality discrimination, and thickness curve smoothing on ultrasonic thickness measurements; bandpass filtering, spectral feature extraction, and noise floor estimation on acoustic data; and drift correction and threshold consistency checks on gas and environmental data, thereby obtaining stable, fusionable, and interpretable processed data. The processed multimodal sensor data output by the data processing module is provided in the form of synchronous frame packets, including: corrected images and their imaging parameters, point cloud data transformed to the machine system or pipeline coordinate system, mileage and attitude increments and their uncertainty estimates, as well as thickness grids, acoustic feature vectors, and environmental conditions.

[0021] In this embodiment, the defect understanding and parametric modeling unit identifies pipeline defect information based on multimodal sensor data, as follows: Define time t k The synchronization frame packet is: ; Among them, I k The corrected image; Π k These are imaging parameters (intrinsic parameters, exposure, gain, etc.). Pk={p j} represents a point cloud; For thickness grids; e is the acoustic feature vector; k Environmental conditions (gas, temperature, humidity); Δuk represents the mileage and attitude increments; Σ Δu,k The uncertainty is... For a point cloud point p=(x,y,z), define the unfolded coordinates (s,θ) with the pipe axis as x and the cross-sectional plane as yz: ; Radius r and topographic relief Defined as: ; Where R0 is the radius inside the tube, the image pixel u=(u,v) is first back-projected to obtain the three-dimensional point, and then mapped to (s,θ) in the same way; Discretize the unfolded surface into a mesh to construct a multi-channel feature map: X(s,θ)=[X I ,X δr ,X th ,X ac ,X e ]; Among them, X I ,X δr ,X th ,X ac ,X e These are visual texture features, radial topography, thickness, acoustic features, and environmental channels; The probability map P of each type of defect is output through visual segmentation. c (s,θ): P c (s,θ)=σ(f c (X I (s,θ))),c∈{cracks, holes, corrosion, misalignment, deposition}; Where fc(·) is the segmentation network and σ is the sigmoid function; Geometric Anomaly Score A g (s,θ) (e.g., the edge of a depression / step / hole) is obtained from the local statistics of δr: ; Where, μ N ,σ N δr(s,θ) represents the mean and standard deviation of the neighborhood N; δr(s,θ) represents the radial offset. It is a constant; Thickness Anomaly Score A t (s,θ): ; Where t0 is the nominal wall thickness or healthy baseline, and t(s,θ) comes from the thickness grid Z. th ; Fusing multimodal evidence into a posterior evidence of defect existence: logit P(def|s,θ)=w I logit P c (s,θ)+w g A g (s,θ)+w t A t (s,θ)+wac A ac (s,θ); ; Among them, w I ,w g ,w t ,w ac These correspond to the credibility weights of the visual, geometric, thickness, and acoustic channels, respectively; A ac Score acoustic anomalies; logit(·) is the logarithmic probability function; For fusion posterior probabilities; Connectivity extraction is performed on P(def|s,θ) to obtain candidate regions Ω. i .

[0022] In this embodiment, each defect information is converted into a structured data object, as follows: For each cross-sectional point set {(y) at each position s along the path j ,z j The nominal radius R and center (y0, z0) are obtained by fitting a circle. The circle fitting uses least squares. ; The cross-sectional deformation index is expressed as roundness error E(s): ; Where Varj(·) represents the variance calculated for all circumferential measurement points j of the cross section; Interface misalignment is detected using abrupt changes in radius along the connection direction: ; Among them, S step R(s) is the radius step; R(s) is the fitted radius at position s along the path; Δs is the difference step size along the path. For candidate region Ω i Calculate the geometric parameters (on the unfolded surface): ; ; ; Among them, L i For length, W i For width, A i For area, D i For depth, η i For the thinning rate; Environmental and material related parameters from e kBased on the surface assessment model, including humidity index h, gas exceedance marker g, and surface deposit level a, a repairability judgment is given: ; Where α1, α2, α3, and α4 are weighting coefficients; 1[h>h0] indicates height anomaly, which is 1 when the defect height exceeds the threshold h0; 1[a>a0] indicates area anomaly, which is 1S when the defect area exceeds the threshold a0. i For comprehensive severity scoring; The output confidence level can be taken as the posterior mean within the region and quality gating should be considered. ; Where |Ωi| represents the area of ​​the defect region; γ(q) is the quality modulation factor, and q is a data quality indicator (image sharpness, point cloud density, etc.); C i The average confidence level for the region; γ(q)∈(0,1] is a discount factor determined by data quality q. The final structured defect object is defined as:

[0023] Among them, c i Defect categories (cracks, holes, corrosion, misalignment, etc.); For the bounding box along the process; For circumferential bounding box; m i Material condition (pipe type, lining, roughness, etc.); evidence is the chain of evidence (original image region, point cloud fragment, thickness distribution, acoustic features, etc.).

[0024] In this embodiment, the fusion positioning module receives the mileage and attitude increments and their uncertainty estimates from the synchronization frame packet, the corrected image and imaging parameters, point cloud data, and the defect object output by the defect understanding and parameterized modeling unit, and maintains the anchor point set A={a k}, each anchor point a k Includes anchor point type, feature descriptor, and mileage estimation at the anchor point. With covariance For each defective object d i The fusion positioning module first calculates its center along the path. (Obtained by integrating the extent of the defect on the unfolded surface with the odometer reading) and the circumferential center. Then select the nearest anchor point a. As a reference anchor point, output the anchor point positioning result. ,in .

[0025] In this embodiment, anchor point construction is as follows: During the inspection process, the inspection robot performs anchor point candidate detection on the image and point cloud: On the image side, features such as circumferential stripes and marking patterns are extracted and descriptors are formed; on the point cloud side, geometric features such as misaligned steps, variable diameters, and bifurcations are extracted through cross-section fitting and abrupt change detection along the path; and a candidate set is constructed by combining the motion priors provided by mileage and attitude increments; then, using the mileage prediction interval as a spatiotemporal gate, the similarity between the candidate and historical anchor points (visual descriptor distance, geometric feature difference, and cross-frame persistence) is calculated, and the anchor point observation is confirmed through threshold and multi-frame consistency checks; once a match is successful, the module uses the anchor point observation as a constraint to update the mileage drift, thereby aligning the path coordinates with the anchor point table and reducing cumulative errors.

[0026] In this embodiment, during the process decision-making stage, the intelligent decision-making and task graph generation unit selects a set of candidate processes from the process library based on the defect type, size parameters and material compatibility. For example, for crack-type defects, "resin injection + surface patch" or "simple UV-cured coating" can be selected, and for hole-type defects, "filling + patch" or "mechanical plugger + winding reinforcement" can be selected. Subsequently, the unit applies a multi-objective optimization framework to select processes and parameters under safety constraints. The objective functions include expected repair quality (based on defect size, process coverage, and material performance suitability scores), operation time cost (including relocation, pretreatment, and curing), resource consumption (material usage and energy consumption), and operation risk (based on environmental parameters and equipment accessibility safety scores). The constraints include hard constraints (prohibition of heating processes due to excessive gas limits, pipe diameter restrictions on the entry of large equipment, and prohibition of drilling due to excessively thin walls) and soft constraints (workspace accessibility, surface treatment preconditions, and material curing time windows). The constraint satisfaction problem CSP solver is used for optimization, outputting the optimal process selection, specific parameter configurations (such as spray thickness, curing temperature and time, patch size and shape), and operation priority ranking for each defect.

[0027] In this embodiment, during the task graph generation stage, the intelligent decision-making and task graph generation unit converts the process decision results into a directed acyclic graph containing nodes and edges. Node types include condition check nodes (anchor point arrival confirmation, environmental parameter detection, surface pretreatment quality verification), execution action nodes (movement, cleaning, spraying, curing, measurement, etc.), decision branch nodes (success / failure / retry based on sensor feedback), and synchronization nodes (multi-device collaboration, parallel operation merging). Edges represent the dependencies between nodes, triggering conditions, and data flow, and carry metadata such as timeout thresholds, maximum number of retry attempts, and fallback paths in case of failure. The task graph also embeds safety interlocking logic: safety checks (gas exceeding limits, emergency stop signals, mooring status) are inserted before each execution node, and once triggered, a forced jump to a safety fallback branch is initiated. Quality inspection and anomaly handling branches are set after key process nodes, such as triggering respraying or re-surface treatment if the thickness detection after spraying is unqualified. The final output task graph contains a complete set of nodes, edges, initial and target state definitions, and interface agreements with the anchor point positioning module (mapping function from defect location to robot path).

[0028] In this embodiment, the control execution unit receives the task graph G=(V,E) output by the intelligent decision-making and task graph generation unit, the anchor point positioning results output by the fusion positioning module, and the defect area information output by the defect understanding unit. Simultaneously, it subscribes to robot body sensing and execution feedback (including motion control, tool status, gas / emergency stop / tether status, thickness / visual re-inspection results). The control execution unit maintains the running state (READY / RUNNING / SUCCESS / FAIL / ABORT) and running context (input data key, output result, cumulative retry count k, and start timestamp) of each node. During scheduling, it determines whether a node is enabled based on the task graph dependencies and the node's predicate: when all predecessor nodes of a node are SUCCESS and its predicate check condition is true, it is set to RUNNING and the corresponding control primitive is issued. During node execution, it continuously monitors sensor feedback and process quantities; when the completion criterion is met, it is set to SUCCESS and the subsequent node is advanced along the edge that meets the trigger condition. For timeout management, if the node's runtime exceeds its threshold τ... v or side window τ e If violated, it is judged as FAIL and an exception branch is triggered: when k <K v Enter the RETRY edge according to the task map and execute predefined retry preparation actions (such as re-anchor point confirmation, deceleration rescanning, supplemental lighting / cleaning), while k k+1; when k=K vIf consecutive failures reach the policy threshold, the system will enter the rollback subgraph along the FALLBACK edge (e.g., "re-surface pretreatment", "use alternative process", "skip the defect and mark it for manual review" or "execute safe rollback"), and write the failure reason, evidence chain and node log into the task execution record for subsequent acceptance and traceability.

[0029] In this embodiment, the control execution unit performs safety interlock gating at the entry point of all execution action nodes, calculates safety predicates (gas over-limit, emergency stop trigger, mooring abnormality, etc.) in real time, and immediately preemptively terminates the current action (sets the relevant node to ABORT) and forcibly jumps to the safe retreat branch in the task graph, performing actions such as stopping, tool de-energizing, and retreating to the nearest safe anchor point, until the safety predicate is restored before allowing re-entry into the task graph; the control execution unit is also responsible for positioning the defect in the area Ω=[s] on the anchor point location unfolding surface. min ,s max ]×[θ min ,θ max Converting to a robot-executable path: First, determine the coverage strategy based on process parameters (e.g., grid scan step distance Δw=b(1-ρ), target tool distance from wall d0, travel speed / flow rate / curing dosage, etc.), generate a coverage trajectory Γ={(s(t),θ(t))} on the unfolded surface, and map it to the absolute path s=s(a) according to the anchor point alignment. )+Δs; then map the (s,θ) trajectory to the three-dimensional target point sequence p(t) in the pipeline coordinate system (and combine it with the body / tool ​​coordinate transformation to obtain the end pose), forming a path segment (including velocity, acceleration, attitude / normal constraints and process synchronization parameters) that can be tracked by the underlying controller; finally, the control execution unit outputs the current task graph node ID, path segment p(t), trajectory parameters, process control parameters and safety constraints as a unified instruction to the repair robot for execution, and writes the execution feedback (arrival error, thickness / quality inspection value, energy dose, anomaly code, etc.) back to the task graph context.

[0030] A control method for an AI-driven adaptive pipeline repair robot system includes the following steps: S1: Preprocess the multimodal sensor data acquired by the inspection robot to obtain the processed multimodal sensor data; S2: Based on multimodal sensor data, identify pipeline defect information and convert each defect information into a structured data object; S3: Anchor point location for pipeline defects; S4: Based on pipeline defect information and anchor point location, obtain the optimal process and executable flow under rule constraints, and construct a task graph; S5: Schedule nodes according to the task graph, manage state transitions, timeouts, retry counts, and backtracking branch triggers, while executing safety interlocks, converting the defect area on the anchor point positioning unfolding surface into an executable path, and outputting it to the repair robot.

[0031] 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.

[0032] 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.

[0033] 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.

[0034] 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.

[0035] 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. An AI-driven adaptive pipeline repair robot system, characterized in that, It includes a data processing module, a defect understanding and parametric modeling unit, a fusion positioning module, an intelligent decision-making and task graph generation unit, and a control execution unit; the data processing module preprocesses the multimodal sensor data acquired by the inspection robot to obtain the processed multimodal sensor data; The defect understanding and parametric modeling unit identifies pipeline defect information based on multimodal sensor data and converts each defect information into a structured data object; the fusion positioning module performs anchor point positioning of the pipeline defect location; the intelligent decision-making and task graph generation unit obtains the optimal process and executable flow under rule constraints based on pipeline defect information and anchor point positioning, and constructs a task graph; the control execution unit schedules nodes according to the task graph, manages state transitions, timeouts, retry counts, and backtracking branch triggers, while executing safety interlocks, converting the defect area on the anchor point positioning unfolding surface into an executable path, and outputting it to the repair robot.

2. The AI-driven adaptive pipeline repair robot system according to claim 1, characterized in that, The defect understanding and parametric modeling unit identifies pipeline defect information based on multimodal sensor data, as detailed below: Define time t k The synchronization frame packet is: ； Among them, I k The corrected image; Π k For imaging parameters; Pk={p j } represents a point cloud; For thickness mesh; e is the acoustic feature vector; k The environment state is represented by Δuk, which represents the odometer and attitude increments. Δu,k For uncertainty; For a point cloud point p=(x,y,z), define the unfolded coordinates (s,θ) with the pipe axis as x and the cross-sectional plane as yz: ； Radius r and topographic relief Defined as: ； Where R0 is the radius inside the tube, the image pixel u=(u,v) is first back-projected to obtain the three-dimensional point, and then mapped to (s,θ) in the same way; Discretize the unfolded surface into a mesh to construct a multi-channel feature map: X(s,θ)=[X I ,X δr ,X th ,X ac ,X e ]; Among them, X I ,X δr ,X th ,X ac ,X e These are visual texture features, radial topography, thickness, acoustic features, and environmental channels; The probability map P of each type of defect is output through visual segmentation. c (s,θ): P c (s,θ)=σ(f c (X I (s,θ))),c∈{cracks, holes, corrosion, misalignment, deposition}; Where fc(·) is the segmentation network and σ is the sigmoid function; Geometric Anomaly Score A g (s,θ) is obtained from the local statistics of δr: ； Where, μ N ,σ N δr(s,θ) represents the mean and standard deviation of the neighborhood N; δr(s,θ) represents the radial offset. It is a constant; Thickness Anomaly Rating A t (s,θ): ； Where t0 is the nominal wall thickness or healthy baseline, and t(s,θ) comes from the thickness grid Z. th ; Fusing multimodal evidence into a posterior evidence of defect existence: logit P(def∣s,θ)=w I logit P c (s,θ)+w g A g (s,θ)+w t A t (s,θ)+w ac A ac (s,θ); ； Among them, w I ,w g ,w t ,w ac These correspond to the credibility weights of the visual, geometric, thickness, and acoustic channels, respectively; A ac Score acoustic anomalies; logit(·) is the logarithmic probability function; For fusion posterior probabilities; Connectivity extraction is performed on P(def|s,θ) to obtain candidate regions Ω. i .

3. The AI-driven adaptive pipeline repair robot system according to claim 2, characterized in that, The process of converting each defect information into a structured data object is as follows: For each cross-sectional point set {(y) at each position s along the path j ,z j The nominal radius R and center (y0, z0) are obtained by fitting a circle. The circle fitting uses least squares. ； The cross-sectional deformation index is expressed as roundness error E(s): ； Where Varj(·) represents the variance calculated for all circumferential measurement points j of the cross section; Interface misalignment is detected using abrupt changes in radius along the connection direction: ； Among them, S step R(s) is the radius step; R(s) is the fitted radius at position s along the path; Δs is the difference step size along the path. For candidate region Ω i Calculate geometric parameters: ； ； ； Among them, L i For length, W i For width, A i For area, D i For depth, η i For the thinning rate; Environmental and material related parameters from e k Based on the surface assessment model, including humidity index h, gas exceedance marker g, and surface deposit level a, a repairability assessment is given: ； Where α1, α2, α3, and α4 are weighting coefficients; 1[h>h0] indicates height anomaly, which is 1 when the defect height exceeds the threshold h0; 1[a>a0] indicates area anomaly, which is 1S when the defect area exceeds the threshold a0. i For comprehensive severity scoring; The output confidence level can be taken as the posterior mean within the region and quality gating should be considered. ； Where |Ωi| represents the area of ​​the defect region; γ(q) is the quality modulation factor, and q is the data quality index; C i The average confidence level for the region; γ(q)∈(0,1] is a discount factor determined by data quality q. The final structured defect object is defined as: ； Among them, c i Defect category; For the bounding box along the process; For circumferential bounding box; m i The material state; evidence is the chain of evidence.

4. The AI-driven adaptive pipeline repair robot system according to claim 3, characterized in that, The fusion positioning module receives the mileage and attitude increments and their uncertainty estimates from the synchronization frame packets, the corrected image and imaging parameters, point cloud data, and the defect object output by the defect understanding and parameterized modeling unit, and maintains the anchor point set A={a k }, each anchor point a k Includes anchor point type, feature descriptor, and mileage estimation at the anchor point. With covariance ; For each defective object d i The fusion positioning module first calculates its center along the path. and the surrounding center Then select the nearest anchor point a. As a reference anchor point, output the anchor point positioning result. .

5. The AI-driven adaptive pipeline repair robot system according to claim 4, characterized in that, The anchor point construction is as follows: During the inspection process, the inspection robot performs anchor point candidate detection on the image and point cloud: On the image side, features such as circumferential stripes and marking patterns are extracted and descriptors are formed; on the point cloud side, geometric features such as misaligned steps, variable diameters, and bifurcations are extracted through cross-section fitting and abrupt change detection along the path; and a candidate set is constructed by combining the motion priors provided by mileage and attitude increments; then, the mileage prediction interval is used as a spatiotemporal gate to calculate the similarity between the candidates and historical anchor points, and the anchor point observation is confirmed through threshold and multi-frame consistency checks; Once a match is successful, the module uses the anchor point observation as a constraint to update the mileage drift, thereby aligning the coordinates along the route with the anchor point table and reducing accumulated errors.

6. The AI-driven adaptive pipeline repair robot system according to claim 1, characterized in that, In the process decision-making stage, the intelligent decision-making and task graph generation unit selects a set of candidate processes from the process library based on defect type, size parameters and material compatibility. The objective function includes expected repair quality, operation time cost, resource consumption and operation risk. The constraints include hard constraints and soft constraints. The constraint satisfaction problem CSP solver is used to optimize the solution and output the optimal process selection, specific parameter configuration and operation priority ranking for each defect.

7. The AI-driven adaptive pipeline repair robot system according to claim 6, characterized in that, In the task graph generation stage, the intelligent decision-making and task graph generation unit converts the process decision results into a directed acyclic graph containing nodes and edges. The node types include condition check nodes, execution action nodes, decision branch nodes, and synchronization nodes. The edges represent the dependencies between nodes, triggering conditions, and data flows, and carry metadata such as timeout thresholds, maximum number of retries, and fallback paths in case of failure. The task graph also embeds safety interlocking logic: a safety check is inserted before each execution node, and once triggered, it forces a jump to a safe rollback branch; And set up quality inspection and anomaly handling branches after key process nodes; the final output task graph includes a complete set of nodes, edges, initial state and target state definitions, as well as interface conventions with the anchor point positioning module.

8. The AI-driven adaptive pipeline repair robot system according to claim 7, characterized in that, The control execution unit receives the task graph G=(V,E) output by the intelligent decision-making and task graph generation unit, the anchor point localization results output by the fusion localization module, and the defect area information output by the defect understanding unit. Simultaneously, it subscribes to robot body sensing and execution feedback. The control execution unit maintains the running state and running context of each node. During scheduling, it determines whether a node is enabled based on the task graph dependencies and the node's predicate: when all predecessor nodes of a node are SUCCESS and its predicate check condition is true, it sets the node to RUNNING and issues the corresponding control primitive. During node execution, it continuously monitors sensor feedback and process variables; when the completion criterion is met, it sets the node to SUCCESS and advances the successor node along the edge that meets the trigger condition. For timeout management, if the node's runtime exceeds its threshold τ... v or side window τ e If violated, it is judged as FAIL and an exception branch is triggered: when k <K v When the task graph is entered, the RETRY edge is entered and the predefined retry preparation actions are executed, while k k+1; when k=K v If consecutive failures reach the policy threshold, the system will enter the backtracking subgraph along the FALLBACK edge and write the failure reason, evidence chain, and node log into the task execution record for subsequent acceptance and traceability.

9. The AI-driven adaptive pipeline repair robot system according to claim 8, characterized in that, The control execution unit performs safety interlock gating at the entry points of all execution action nodes, calculates safety predicates in real time, and immediately preemptively terminates the current action once any interlock is triggered, forcibly jumping to the safe retreat branch in the task diagram to perform actions such as stopping, disabling tools, and retreating to the nearest safe anchor point, until the safety predicate is restored before allowing re-entry into the task diagram; the control execution unit is also responsible for positioning the defect in the area Ω=[s] on the anchor point location unfolding surface. min ,s max ]×[θ min ,θ max Converting to a robot executable path: First, determine the coverage strategy based on process parameters, generate a coverage trajectory Γ={(s(t),θ(t))} on the unfolded surface, and map it to the absolute path s=s(a) according to the anchor points. )+Δs; then map the (s,θ) trajectory to the three-dimensional target point sequence p(t) in the pipeline coordinate system to form a path segment that can be tracked by the underlying controller; finally, the control execution unit outputs the current task graph node ID, path segment p(t), trajectory parameters, process control parameters and safety constraints as a unified instruction to the repair robot for execution, and writes the execution back to the task graph context.

10. A control method for an AI-driven adaptive pipeline repair robot system as described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Preprocess the multimodal sensor data acquired by the inspection robot to obtain the processed multimodal sensor data; S2: Based on multimodal sensor data, identify pipeline defect information and convert each defect information into a structured data object; S3: Anchor point location for pipeline defects; S4: Based on pipeline defect information and anchor point location, obtain the optimal process and executable flow under rule constraints, and construct a task graph; S5: Schedule nodes according to the task graph, manage state transitions, timeouts, retry counts, and backtracking branch triggers, while executing safety interlocks, converting the defect area on the anchor point positioning unfolding surface into an executable path, and outputting it to the repair robot.

Images