Decision method for sewer rehabilitation process and related devices

Abstract

The application relates to a decision-making method for a sewer pipe repair process and a related device. The method comprises: acquiring a defect parameter set and a pipe section feature set output by in-pipe video detection; determining a main recommended process and an alternative list for each defect in a preset repair process set through a three-layer architecture of hard rule screening, scoring function and machine learning anomaly review; querying a preset process parameter library based on the main recommended process to obtain process parameter recommendations; and outputting the main recommended process, the alternative list and the process parameter recommendations according to the anomaly review results. The application has the advantages of a hierarchical mixed architecture, which makes the decision-making remain interpretable and supports continuous optimization of the decision-making model based on feedback of post-construction review results.

Description

Technical Field

[0001] This application relates to the field of drainage pipeline repair technology, and in particular to a decision-making method and related apparatus for drainage pipeline repair process. Background Technology

[0002] Drainage pipes are prone to various defects such as cracks, ruptures, deformation, and collapses during long-term service, requiring repair without large-scale road excavation. Currently, commonly used trenchless repair techniques include several independent processes such as point-based in-situ curing, ultraviolet light pipe curing, and pipe fragmentation, each with different applicable ranges for different defect types, severity levels, and pipe section conditions.

[0003] However, when faced with specific defect parameters and pipe segment characteristics output by in-pipe video inspections, the industry currently relies primarily on the experience-based judgment of field engineers to determine which repair process to choose. Because this experience-based judgment is highly dependent on individual professional expertise, different engineers may offer significantly different process recommendations under the same defect conditions, leading to a lack of unified standards for repair process selection within the same project. Furthermore, since the basis for experience-based judgment is implicit in the engineer's individual understanding and cannot be structured and transmitted, the results of post-repair inspections are difficult to use for iterative optimization of decision-making, further limiting the decision-making efficiency and quality stability of large-scale pipeline repair projects. Summary of the Invention

[0004] To address the problem that drainage pipeline repair processes rely on the experience and judgment of on-site engineers, resulting in a lack of unified standards for process selection within the same project and difficulty in using re-inspection results as a basis for iterative optimization, this application provides a decision-making method and related apparatus for drainage pipeline repair processes.

[0005] Firstly, this application provides a decision-making method for drainage pipeline repair processes, which adopts the following technical solution: A decision-making method for drainage pipe repair technology includes the following steps: S1. Obtain the defect parameter set and pipe segment feature set output by the video detection inside the pipe. The defect parameter set shall at least include the defect type, defect level, defect distance and clock position, and the pipe segment feature set shall at least include the pipe diameter and pipe material.

[0006] S2. Perform hard rule screening on the defect parameter set and pipe segment feature set to obtain a candidate process set. The hard rule screening excludes unsuitable processes from the preset repair process set according to preset safety constraints. The preset repair process set includes at least point-like in-situ curing process, ultraviolet light pipe curing process and pipe fragmentation method process.

[0007] S3. Based on the defect parameter set and pipe segment feature set, obtain the score value corresponding to each candidate process in the candidate process set through the scoring function, and sort the candidate process set from high to low according to the score value to obtain the main recommended process and the alternative list.

[0008] S4. Perform machine learning anomaly verification on the main recommended process to obtain the anomaly verification results of the main recommended process. The anomaly verification results shall include at least anomaly labels and output confidence scores.

[0009] S5. Based on the main recommended process, query the preset process parameter library to obtain the recommended process parameters corresponding to the main recommended process, and output the main recommended process, alternative list and recommended process parameters according to the anomaly review results.

[0010] By adopting the above technical solution, hard rule screening, scoring functions, and machine learning anomaly verification are organized into sequential steps. Hard rule screening uses preset safety constraints as a safety net to avoid construction failure caused by recommending incorrect processes. The scoring function scores and ranks the remaining candidate processes to obtain an interpretable main recommended process and alternative list, making the decision-making process transparent to the construction party. Machine learning anomaly verification verifies the main recommended process to identify anomalies that are difficult to detect by rules and scoring alone. The three work together in a causal chain of "first eliminate, then rank, then verify" to achieve a balance between the accuracy, interpretability, and coverage of the decision results.

[0011] Optionally, the method employs a two-stage execution approach for S2, S3, and S4, comprising an initial judgment stage and a verification stage: The initial judgment stage is performed during the in-pipe video inspection process. This stage uses the acquired partial defect parameter set and partial pipe segment feature set as input, sequentially executing S2 and S3 to obtain a preliminary recommended process. S4 is not executed during the initial judgment stage. The verification stage is performed after the in-pipe video inspection is completed. This stage uses the full-segment defect parameter set and full-segment feature set as input, sequentially executing S2, S3, and S4 to obtain the main recommended process, a candidate list, and anomaly verification results. The output of the verification stage serves as the input for S5.

[0012] By adopting the above technical solution, the three-layer action is divided into an initial judgment stage and a review stage. This allows the construction party to obtain a preliminary recommended process during the inspection process to allocate equipment in advance, and to obtain the main recommended process after anomaly review based on the data of the entire section after the inspection is completed. This achieves a balance between the timeliness of decision response and the overall optimality.

[0013] Optionally, the review phase also includes: when identifying multiple defects within the same pipe segment from the set of full-segment defect parameters input during the review phase, clustering the multiple defects according to process affinity based on the primary recommended processes corresponding to each defect to obtain at least one batch of the same process. The process affinity is jointly determined by the degree of shared construction equipment and the continuity of construction sequence among the primary recommended processes corresponding to each defect. When the score values ​​corresponding to all candidate processes in the candidate process set are lower than the preset single process upper limit threshold, a process combination scheme is obtained by assembling at least the top two candidate processes in the candidate process set based on their score values, and the process combination scheme is included in the candidate list.

[0014] By adopting the above technical solution, multiple defects in the same pipe section are clustered into batches of the same process according to process affinity, so that all defects in the batch can be solved in one construction, reducing the fixed cost of repeated entry and exit of inspection wells. When all candidate processes are insufficient in score, the candidate processes with the highest scores are assembled into a process combination scheme to deal with complex defects that are difficult to solve by a single process.

[0015] Optionally, after S3 and before S4, a two-sided approach alarm judgment is also performed: when the candidate process set is not empty, the score value corresponding to all candidate processes in the candidate process set is lower than the preset alarm lower limit threshold, and the number of unsuitable processes excluded from the preset repair process set by hard rule screening reaches the preset exclusion quantity threshold, an alarm is triggered, and the main recommended process is determined to be transferred to manual processing.

[0016] By adopting the above technical solution, alarm judgment must meet two conditions simultaneously: the scores of all candidate processes on the positive side are insufficient and the scores of the excluded processes on the negative side reach the threshold before it is triggered. This avoids false alarms caused by insufficient input features on one side and ensures that manual processing is only initiated when it is confirmed that none of the four processes are sufficient to resolve the issue.

[0017] Optionally, in S2, hard rule screening employs a dual-threshold mechanism, which includes a hard threshold and a soft threshold: In response to a defect level in the defect parameter set reaching the hard threshold, hard rule screening determines unsuitable processes from the preset repair process set according to the preset exclusion rule corresponding to the hard threshold. In response to a defect level in the defect parameter set reaching the soft threshold but not the hard threshold, hard rule screening marks some processes in the preset repair process set as processes awaiting manual review according to the preset marking rule corresponding to the soft threshold; these processes are not temporarily identified as unsuitable. When a process awaiting manual review is determined to be unsuitable after manual review, it is officially designated as an unsuitable process. When a process awaiting manual review is determined to be suitable after manual review, it is retained in the candidate process set.

[0018] By adopting the above technical solution, the determination of hard rules is divided into two levels: strong exclusion of hard thresholds and manual review of soft thresholds. This allows serious defects to be directly converged to the limited process to ensure safety, while allowing boundary conditions to retain the flexibility of expert judgment through manual review.

[0019] Optionally, in S2, hard rule screening performs a two-layer split based on the defect type in the defect parameter set. The first layer determines the defect branch based on the defect type, with each branch including at least a local defect branch and a full-segment defect branch. The second layer marks each process in the preset repair process set as either a preferred or non-preferred marker based on the defect branch. Specifically, the point-like in-situ curing process under the local defect branch is marked as the preferred marker, while the UV-curing process and the fragmentation method under the full-segment defect branch are marked as preferred markers. The remaining processes in the preset repair process set are marked as non-preferred markers. Hard rule screening determines unsuitable processes in the preset repair process set based on the non-preferred markers and the defect level in the defect parameter set.

[0020] By adopting the above technical solution, process families are first divided according to the local or whole-section attributes of the defect type, and then unsuitable processes are determined according to the defect level within the family. This makes the judgment structure of hard rule screening naturally aligned with the applicable objects of the two types of repair processes: point repair and whole-section renewal.

[0021] Optionally, in S4, machine learning anomaly review is performed based on a machine learning model and a training library. The machine learning model is trained on historical decision samples stored in the training library. The training library stores multiple historical decision samples, each containing at least a historical defect parameter set, a historical pipe segment feature set, historical decision results, and a re-inspection success rate. After S5, the following feedback write-back operation is performed on the main recommended process: After completing the repair according to the main recommended process, the repaired pipe segment is re-performed with in-pipe video inspection, and the re-inspection success rate is determined based on the results of the re-performed in-pipe video inspection. The defect parameter set, pipe segment feature set, main recommended process, and re-inspection success rate obtained in S1 are written back to the training library as new historical decision samples. The machine learning model is then updated on a rolling basis based on the written-back training library.

[0022] By adopting the above technical solution, the results of the in-pipe video re-inspection after each repair are paired with the decision recommendation as new samples and written back to the training library to continuously update the machine learning model. This enables the decision-making method to form a closed loop of "decision-making, construction, re-inspection, writing back, and re-decision-making", continuously calibrating the machine learning model's ability to predict the actual repair effect.

[0023] Optionally, when the output confidence of the machine learning anomaly review in S4 is lower than a preset confidence threshold, the following cross-segment similarity analogy correction operation is performed on the main recommended process: Based on the historical decision samples stored in the training library, the comprehensive similarity between each historical decision sample and the current defect is calculated from two sets of features: the historical defect parameter set and the historical segment feature set. The historical decision samples are sorted from largest to smallest based on their comprehensive similarity, and a preset number of historical decision samples are selected as analogy samples. The re-inspection success rate corresponding to the analogy samples is used as a secondary weighting coefficient to perform weighted fusion on the historical decision results corresponding to the analogy samples, resulting in the analogy recommended process. The main recommended process is then corrected based on the analogy recommended process.

[0024] By adopting the above technical solution, when the machine learning model has low confidence in the current defect, the decision results of similar historical pipe sections are introduced as a reference. The historical re-inspection success rate is used for secondary weighting so that the historical decisions with good re-inspection results have greater influence, thereby mitigating the risk of model overfitting when there is insufficient data in the early stage of the training library.

[0025] Optionally, the point-like in-situ curing process in the preset repair process set is implemented by impregnating pre-cut glass fiber fabric with resin, wrapping it on an inflatable airbag, guiding it to the point to be repaired in the pipeline by a pulling device or robot, inflating the airbag to press the resin-impregnated fiber cloth onto the pipeline defect, maintaining pressure and allowing the resin to cure under normal temperature or heating conditions to form an inner lining patch.

[0026] By adopting the above technical solution, a local lining patch is formed by combining a fiberglass felt cylinder with an inflatable airbag, enabling local defects such as pitted cracks and pitted corrosion to be repaired with high strength under excavation conditions.

[0027] Optionally, the UV-curing process in the preset repair process set is implemented by pulling a pre-impregnated photosensitive resin-coated fiberglass hose from the inspection well into the pipe to be repaired using a pulling device, filling the hose with compressed air to make the hose fit tightly against the inner wall of the original pipe, and moving the UV lamp holder along the pipe to irradiate the photosensitive resin with UV light of a specific wavelength to trigger cross-linking and curing of the photosensitive resin within a few hours to form a pipe-in-pipe composite structure.

[0028] By adopting the above technical solution, a corrosion-resistant pipe-in-pipe composite lining is formed inside the original pipeline through the ultraviolet curing reaction of photosensitive resin, so that the structural integrity of the entire severely defective pipeline can be restored and its service life extended without excavation.

[0029] Optionally, the pipe-breaking method in the preset repair process set is implemented by inserting a conical pipe-breaking head with breaking teeth at the front end into the old pipe using a hydraulic jacking device, mechanically breaking the old pipe completely and squeezing the fragments into the surrounding soil to form a new channel, and simultaneously pulling in a high-density polyethylene short pipe that has been pre-assembled on the ground behind the pipe-breaking head.

[0030] By adopting the above technical solution, new pipes are pulled in simultaneously while old pipes are being broken, allowing pipe sections with severe structural defects such as deformation and collapse to be replaced without excavation. At the same time, the diameter of the new pipe can be larger than that of the old pipe to achieve a small expansion.

[0031] Optionally, in-pipe video inspection uses a remote-controlled tracked crawler equipped with a camera and lighting system to enter the drainage pipe and collect real-time images of the pipe wall condition. Based on the encoder, the defect distance is output in real time, and the defect type, defect level and clock position are determined based on image recognition.

[0032] By adopting the above technical solution, a standardized detection input terminal is constructed by combining a remote-controlled crawler with an encoder and image recognition, so that the defect parameter set and pipe segment feature set can be produced in a consistent field format for subsequent decision-making methods to consume directly.

[0033] Optionally, the pipe materials with concentrated pipe segment characteristics include at least reinforced concrete pipes, high-density polyethylene corrugated pipes, polyvinyl chloride pipes, and ductile iron pipes.

[0034] By adopting the above technical solutions, a variety of common pipe types in drainage pipe networks are covered, enabling the decision-making method to provide targeted process recommendations based on the mechanical properties and repair process adaptability of different pipe materials.

[0035] Optionally, the pipe diameters in the pipe segment feature set shall include at least DN300, DN400, DN500, and DN800.

[0036] By adopting the above technical solutions, a variety of common pipe diameters in drainage pipe networks are covered, enabling the decision-making method to provide targeted process recommendations based on the spatial compatibility between pipe diameter and repair process equipment.

[0037] Optionally, the preset process parameter library may store at least the number of rings and airbag holding time for the dot-shaped in-situ curing process, the hose section length and UV lamp holder moving speed for the UV tube curing process, and the diameter expansion amount and hydraulic jacking speed for the tube fragmentation process.

[0038] By adopting the above technical solution, the key parameters of each repair process are pre-structured and stored in the process parameter library, so that after the decision-making method determines the main recommended process, it can directly look up the corresponding construction parameters from the process parameter library to provide the construction party with a complete and executable solution.

[0039] Optionally, the anomaly label in the anomaly review result is a binary label, which is determined by comparing the output probability of the main recommended process by the machine learning model with the preset anomaly judgment threshold; the output confidence in the anomaly review result is the probability value of the judgment result of the main recommended process by the machine learning model.

[0040] By adopting the above technical solution, the meaning of the two fields in the abnormality review result is clarified, enabling downstream steps to decide whether to enter an alarm based on the abnormality flag and whether to initiate cross-pipeline analog correction based on the output confidence level.

[0041] Optionally, manual review is implemented by pushing the process to be manually reviewed and its corresponding defect parameter set and pipe segment feature set to the review terminal, having the reviewer enter the appropriate or inappropriate review results, and then returning the review results to hard rule filtering.

[0042] By adopting the above technical solution, a standardized closed-loop interface for manual review of process markers triggered by soft thresholds is provided, enabling expert judgment to be integrated into the decision-making process in a structured manner.

[0043] Optionally, the machine learning model is a classification model trained based on historical decision samples. The input of the machine learning model is the feature vector of the defect parameter set and the pipe segment feature set. The output of the machine learning model is the anomaly label and output confidence corresponding to the main recommended process.

[0044] By adopting the above technical solution, the input and output structure of the machine learning model is clearly defined, so that the sample format of the training library corresponds one-to-one with the model inference interface, which facilitates the continuous training and replacement of the machine learning model.

[0045] Optionally, the method also includes pushing the main recommended process, alternative list and process parameter recommendations output by S5 to the urban pipeline construction resource scheduling system, which will automatically reserve the corresponding construction team and construction equipment according to the process type and lock in the construction time window that avoids peak traffic.

[0046] By adopting the above technical solution, the decision-making results are integrated with the construction resource scheduling system, enabling the output of the decision-making method to directly drive the reservation and scheduling of construction resources, thereby reducing the manual connection between decision-making and execution.

[0047] Optionally, the output of the machine learning model is further processed by reverse rule extraction to generate an accompanying explanatory rule set. The accompanying explanatory rule set is fitted to the output of the machine learning model on historical decision samples in the form of a decision tree and is updated synchronously with the rolling update of the machine learning model.

[0048] By adopting the above technical solution, an approximate explanation of the decision-making basis of the machine learning model is given in the form of a readable decision tree, so that the interpretability of the decision-making method can still be maintained after the introduction of machine learning.

[0049] Optionally, the defect parameter set also includes defect width, defect boundary roughness, and pipe wall background texture features automatically derived from the original video frame of the in-pipe video detection. The defect width is obtained by detecting the pixel boundary of the defect area in the original video frame, the defect boundary roughness is obtained by curvature analysis of the defect boundary pixels, and the pipe wall background texture features are obtained by statistical analysis of the texture around the defect area.

[0050] By adopting the above technical solution, finer-grained defect morphology features can be obtained from the original video frames of in-pipe video detection without the need for additional detection equipment, thereby improving the sensitivity of the scoring function to differences in defect morphology.

[0051] Optionally, the preset safety constraints used for hard rule filtering are managed by version number, and new repair processes are registered to the preset repair process set in the form of process description plugins. The process description plugins include at least the preset safety constraints corresponding to the new repair process, the scoring row corresponding to the new repair process in the scoring function, and the recommended field of process parameters for the new repair process.

[0052] By adopting the above technical solution, the integration of new repair processes does not require rewriting the main program of the decision-making method. It can be automatically integrated into the three stages of hard rule screening, scoring function and process parameter recommendation simply by supplementing it in the form of a process description plugin.

[0053] Optionally, the execution of the scoring function in S3 is divided into two levels: coarse screening and fine evaluation. Coarse screening performs preliminary scoring on each candidate process in the candidate process set according to the preset rapid discrimination rules to exclude obviously unsuitable candidate processes. Fine evaluation calculates the score value of the candidate processes retained by coarse screening according to the complete scoring model.

[0054] By adopting the above technical solution, the execution of the scoring function is divided into two levels: rapid exclusion and fine scoring, thereby reducing the computational load of the complete scoring model.

[0055] Optionally, the fields in the defect parameter set are divided into required fields and optional fields. Required fields must include at least the defect type, defect level, defect distance, and clock orientation, while optional fields must include at least the defect axial length and the number of defects. Hard rule filtering and scoring functions can still make decisions based on the required fields even when optional fields are missing.

[0056] By adopting the above technical solution, the decision-making method becomes fault-tolerant to the completeness of the fields in the defect parameter set, and can still provide recommended processes when optional fields are missing, thereby improving the usability of the decision-making method when actual field data is incomplete.

[0057] Secondly, the computer device provided in this application adopts the following technical solution: A computer device comprising: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to: The decision-making method for implementing the above-mentioned drainage pipe repair process.

[0058] Thirdly, this application provides a computer-readable storage medium that adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and execute the methods described above.

[0059] The storage medium stores at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the following: The decision-making method for drainage pipe repair processes, as described above.

[0060] In summary, this application includes at least one of the following beneficial technical effects: 1. This application provides an automated decision-making process for the selection of drainage pipeline repair technology based on defect parameters and pipe segment characteristics through a three-layer architecture of hard rule screening, scoring function and machine learning anomaly verification, so as to achieve a balance between the decision results in terms of accuracy, interpretability and coverage.

[0061] 2. This application refines each step of the decision-making method through mechanisms such as two-stage decision-making, dual-sided approach alarm, dual threshold mechanism, two-layer diversion of defect types, and feedback writing back to the training library, thereby enhancing the timeliness of decision response, the rigor of safety net, the integration of expert judgment, and the ability of machine learning models to continuously calibrate the actual repair effect.

[0062] 3. This application uses supplementary solutions such as cross-segment similarity analogy, urban pipeline network super-system integration, machine learning reverse rule extraction, video frame derived features, rule versioning, and process plug-ins to enable the decision-making method to still rely on historical experience when training data is insufficient, the decision results can directly drive construction scheduling, machine learning decisions maintain readable accompanying explanations, and support the dynamic access of new repair processes. Attached Figure Description

[0063] Figure 1 This is an overall flowchart of the decision-making method for the drainage pipeline repair process provided in the embodiments of this application.

[0064] Figure 2 This is a schematic diagram of a three-layer hybrid architecture provided in an embodiment of this application.

[0065] Figure 3This is a schematic diagram of the internal structure of the hard rule filtering provided in the embodiments of this application.

[0066] Figure 4 This is a flowchart illustrating the initial judgment stage in the two-stage execution method provided in the embodiments of this application.

[0067] Figure 5 This is a flowchart illustrating the review stage in the two-stage execution method provided in this application embodiment.

[0068] Figure 6 This is a flowchart of the cross-segment similarity analogy correction mechanism provided in the embodiments of this application.

[0069] Figure 7 This is a schematic diagram of the feedback write-back closed loop provided in the embodiments of this application.

[0070] Figure 8 This is a schematic diagram of the structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0071] The present application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the application and are not intended to limit the scope of the application.

[0072] This application provides a decision-making method for drainage pipeline repair processes, such as... Figure 1 As shown, the method comprises steps S1 to S5 executed sequentially. Based on the defect parameter set and pipe segment feature set output from the in-pipe video detection, this method employs three sequentially executed steps: hard rule filtering, a scoring function, and machine learning anomaly verification. Figure 2 As shown, for each defect in the preset repair process set, a primary recommended process and a list of alternative processes are determined, and the corresponding process parameters are recommended by querying the preset process parameter library based on the primary recommended process for the construction party to execute.

[0073] For ease of understanding, a typical urban drainage pipe network is used as the baseline scenario below. This scenario involves a DN500 reinforced concrete main drainage pipe section, 80 meters long, and 15 years of service. Video inspection within the pipe section revealed three defects: a localized crack at 22 meters from the manhole (defect level 2, clock position 9 o'clock, marked as defect 1); a section of corrosion deformation from 45 to 58 meters from the manhole (defect level 3, axial length 13 meters, marked as defect 2); and a localized misalignment at the joint at 76 meters from the manhole (defect level 1, clock position 6 o'clock, marked as defect 3). The steps of this method are explained in detail below.

[0074] S1. Obtain the defect parameter set and pipe segment feature set from the video inspection output within the pipe. The defect parameter set is a collection of fields describing the characteristics of each defect, including at least four items: defect type, defect level, defect distance, and clock position. The defect type identifies the morphological attributes of the defect, such as cracks, deformation, misalignment, corrosion, etc.; the defect level identifies the severity of the defect, classified from level 1 to level 4 according to industry convention, with higher levels indicating more severe defects; the defect distance identifies the axial position of the defect from the starting inspection well within the pipe segment; and the clock position identifies the circumferential position of the defect on the pipe wall, expressed in clock-scale format.

[0075] A pipe segment feature set is a collection of fields describing the inherent attributes of a pipe segment, including at least pipe diameter and pipe material. Pipe diameter identifies the nominal inner diameter of the pipe, while pipe material identifies the type of material used to build the pipe. Together, they determine the spatial adaptability of the pipe segment to process equipment and the mechanical compatibility with repair materials during repair.

[0076] The output of S1 serves as the common input for the subsequent three steps S2 to S4. The defect parameter set is used for hard rule screening to determine incompatible processes under preset safety constraints, for scoring functions to score candidate processes, and for machine learning anomaly verification as feature vector input. The pipe segment feature set is also involved in all three stages, serving as the engineering context for the defect parameter set.

[0077] For example, in the baseline scenario, S1 outputs three sets of records in the defect parameter set: Defect 1 corresponds to a local crack, defect level 2, defect distance 22 meters, and clock position 9 o'clock; Defect 2 corresponds to a whole-section corrosion deformation, defect level 3, defect distance 45 to 58 meters, and clock position full circumference; Defect 3 corresponds to a local interface misalignment, defect level 1, defect distance 76 meters, and clock position 6 o'clock. In the pipe segment feature set, one set of records is output: pipe diameter DN500, pipe material reinforced concrete.

[0078] In some embodiments, in-pipe video inspection uses a remote-controlled tracked crawler equipped with a camera and lighting system to enter the drainage pipe and collect real-time images of the pipe wall condition. The encoder outputs the defect distance in real time, and image recognition determines the defect type, defect level, and clock position. After the crawler is lowered into the pipe section to be inspected through the inspection well, it is remotely controlled by a ground control console to move along the pipe axis. The camera continuously collects visible light images of the pipe wall in front and around the perimeter, and transmits them in real time to a ground-based industrial control computer via a communication cable. The image recognition algorithm performs defect determination on each frame of the image and links the determination result with the cumulative travel distance of the encoder at that moment.

[0079] In the baseline scenario, when detecting a localized fracture 22 meters from the wellhead, the crawler entered the pipeline from the initial inspection well and moved forward. The encoder accumulated the travel distance in real time as the crawler's tracks rolled. When it reached the 22-meter position, the camera captured an image of the fracture on the pipe wall. The control console determined the fracture morphology based on an image recognition algorithm and identified the defect type as a localized fracture. Based on the fracture opening size and extension range, the defect level was determined to be Level 2. The clock position was determined to be 9 o'clock based on the circumferential position of the fracture in the image. The same detection process was subsequently applied to the entire section of corrosion deformation from 45 to 58 meters from the wellhead and the localized interface misalignment at 76 meters from the wellhead.

[0080] In other embodiments, in-pipe video detection can also employ non-contact laser scanning or low-frequency acoustic detection as alternative inputs. Laser scanning generates cross-sectional profile data by measuring the distance distribution of reflection points on the pipe wall and infers the defect morphology based on the profile deformation; low-frequency acoustic detection infers the defect location and type by analyzing echo delay and attenuation spectrum. Both alternative methods can produce a defect parameter set and pipe segment feature set conforming to the S1 field format, and therefore also fall within the scope of in-pipe video detection implementation methods.

[0081] In some embodiments, the defect parameter set further includes defect width, defect boundary roughness, and pipe wall background texture features automatically derived from the original video frames of the in-pipe video detection. These three types of derived features do not require additional detection equipment and can be extracted simply by performing pixel-level image processing algorithms on the existing original video frames of the in-pipe video detection.

[0082] The defect width is obtained by detecting the pixel boundaries of the defect area in the original video frame. For example, the Canny edge detection operator is used to locate the contour of the defect in the image, and then the maximum span of the contour is taken and converted into physical size according to the camera calibration parameters.

[0083] Defect boundary roughness is obtained by curvature analysis of defect boundary pixels. For example, by sampling several equidistant points along the contour and calculating the variance of the angle change between adjacent points, the larger the variance, the more irregular the boundary shape.

[0084] The background texture features of the pipe wall are obtained by statistical analysis of the texture around the defect area. For example, the contrast, energy, correlation and other statistical quantities are calculated using the gray-level co-occurrence matrix to characterize the corrosion severity of the pipe wall body.

[0085] After incorporating these derived features, the scoring function becomes more sensitive to differences in defect morphology. For the same defect type and grade, defects with wider widths and more irregular boundaries are often recommended for repair processes with greater spatial margins in the scoring function. Defects with high roughness in the pipe wall background texture are often accompanied by extensive substrate corrosion, making whole-section UV curing more advantageous than point-based repair. For example, in the baseline scene, defect 1 might have the following derived feature output from the original video frame: a defect width of approximately 45 mm, a defect boundary roughness of 0.12 radians, and a pipe wall background texture contrast value of 28.

[0086] In some embodiments, the pipe materials with the characteristic set of pipe segments include at least reinforced concrete pipes, high-density polyethylene corrugated pipes, polyvinyl chloride pipes, and ductile iron pipes. Reinforced concrete pipes are the most common type of pipe in urban drainage networks, exhibiting high rigidity but moderate adhesion to linings. High-density polyethylene corrugated pipes are flexible, corrosion-resistant, but have limited wall smoothness. Polyvinyl chloride pipes are semi-rigid and have good adhesion to various resin linings. Ductile iron pipes are highly rigid and impact-resistant, but require specific hardness from expansion-type repair equipment. The reinforced concrete pipe material in the baseline scenario represents a common case in this enumeration.

[0087] In some embodiments, the pipe diameters in the pipe segment feature set include at least DN300, DN400, DN500, and DN800. DN300 and DN400 are common pipe diameters for urban secondary trunk lines, imposing strict size constraints on the repair process equipment; equipment such as crawlers and UV lamp holders require miniaturized variants. DN500 and DN800 are common pipe diameters for main and manifold pipes, allowing standard versions of the repair process equipment to be used directly. In the baseline scenario, DN500 falls within the compatibility range of standard repair process equipment, and equipment for all three preset repair processes can be used normally during construction.

[0088] In some embodiments, the fields in the defect parameter set are divided into mandatory fields and optional fields. The mandatory fields include at least four items: defect type, defect level, defect distance, and clock orientation, which are the judgment criteria that must be used in each step from S2 to S4; the optional fields include at least the defect axial length and the number of defects. The defect axial length is used to refine the extension range of the entire segment of the defect, and the number of defects is used to refine the density of multiple defects within the same segment.

[0089] Hard rule filtering and scoring functions can still make decisions based on required fields even when optional fields are missing. For example, when the axial length of a defect is missing, hard rule filtering can treat this field as unknown and make a preliminary judgment on whether it is unsuitable for the process based on the defect level; the scoring function can apply a preset default weight to the scoring dimension corresponding to this field, so that the scoring process is not interrupted due to the absence of a single field.

[0090] This field-level mechanism ensures the usability of decision-making methods even when faced with incomplete field data. It is particularly important for scenarios where early data is missing in older pipeline networks, and can provide reasonable process recommendations even for historical pipeline sections lacking detailed axial length measurements.

[0091] In some embodiments, the method employs a two-phase execution approach for S2, S3, and S4, which includes an initial judgment phase and a review phase, such as... Figure 4 and Figure 5 As shown. If decision-making only begins after the entire pipe segment has been inspected via video, the construction team can only begin allocating construction crews and equipment after the inspection is complete, resulting in a long waiting time between decision-making and construction. The two-stage execution method divides the decision-making process into an initial assessment stage for responsiveness and a verification stage for overall optimality. This allows the construction team to obtain preliminary recommended processes during the inspection process, enabling them to allocate equipment in advance. After the inspection is completed, the final recommended process is obtained based on the data from the entire pipe segment.

[0092] In the video detection process within the pipe, a preliminary judgment stage is performed. This stage uses a partially acquired set of defect parameters and a partial set of pipe segment features as input, sequentially executing steps S2 and S3 to obtain a preliminary recommended process. Step S4 is not executed in this stage. In other words, the preliminary judgment stage only performs hard rule filtering to obtain a set of candidate processes, and then uses a scoring function to score these candidate processes to obtain the preliminary recommended process, skipping the steps of machine learning anomaly verification and dependence on the training library. This simplification is reasonable, as the defects acquired while the crawler is still moving within the pipe segment are only a part of the overall defects. Performing rigorous machine learning anomaly verification on a single defect lacks both the context of the entire segment and is unnecessary.

[0093] After the video inspection within the pipe is completed, a review phase is executed. This phase uses the entire segment's defect parameter set and feature set as input, sequentially executing steps S2, S3, and S4 to obtain the primary recommended process, alternative lists, and anomaly review results. The output of the review phase serves as the input for step S5. The difference between the review phase and the initial judgment phase lies in the expanded scope to the entire segment. Once all defects and features of the entire pipe segment are collected, hard rule filtering, scoring functions, and machine learning anomaly reviews are all performed based on the complete data. The resulting primary recommended process serves as the final construction decision basis for that pipe segment.

[0094] In the baseline scenario, the crawler is lowered into the initial inspection well and moves forward. When it reaches the 22-meter mark, defect 1 is identified. In the initial assessment stage, the characteristics of defect 1 (local rupture, level 2) and the pipe diameter (DN500) and reinforced concrete structure are used as inputs to run steps S2 and S3. This yields a preliminary recommended process for defect 1: point-based in-situ curing. This result is then pushed to the construction scheduling terminal. Upon receiving the push, the construction team can pre-book the point-based in-situ curing construction team and the airbag traction equipment. The crawler continues forward and identifies defects 2 and 3 at the 45-58 meter and 76-meter marks, respectively. The initial assessment stage also provides preliminary recommended processes for each defect. After the entire pipe section is inspected, the verification stage uses the full-section data from the three defects as inputs to run steps S2, S3, and S4. This yields the primary recommended process, a candidate list, and anomaly verification results, which serve as the basis for the final decision.

[0095] In contrast, the limitation of the single-stage complete decision-making approach is that the decision-making process is concentrated after the inspection is completed, and the allocation of construction resources can only begin after the decision is made. This is less efficient for large-scale projects with tight schedules or requiring the scheduling of multiple processes. The two-stage approach trades the early results of the initial assessment stage for the early allocation of construction resources, while retaining the optimal decision-making capability for the entire process in the verification stage.

[0096] In some embodiments, the review phase also includes two types of processing. One type is batch clustering of multiple defects within the same pipe section, and the other type is constructing a combined solution when the score of a single process is insufficient.

[0097] When identifying multiple defects within the same pipe section based on the comprehensive defect parameters input during the review phase, clustering is performed on the multiple defects according to their respective primary recommended processes, resulting in at least one batch with the same process. Process affinity is jointly determined by the degree of equipment sharing and the continuity of construction sequence among the primary recommended processes for each defect. The higher the degree to which the primary recommended processes for two defects use the same type of construction equipment, and the more natural the connection between the two construction steps in terms of sequence, the greater the process affinity, and the more likely they are to be clustered into the same batch. For example, in the baseline scenario, defects 1 and 3 both have point-based in-situ curing processes. They use the same set of airbags and traction equipment, and the construction sequence does not require interruption, resulting in high process affinity, and they are clustered into the same batch with the same process. Defect 2's primary recommended process is either ultraviolet light pipe curing or pipe fragmentation, which does not share equipment with defects 1 and 3, and is therefore a separate batch. After clustering, the construction team can complete the repair of all defects within each batch with the same process by entering and exiting the inspection well in one go, without needing to repeatedly enter and exit between different batches.

[0098] When the scores of all candidate processes in the candidate process set are lower than the preset single-process upper limit threshold, a process combination scheme is obtained by assembling at least two candidate processes with the highest scores in the candidate process set, and this process combination scheme is included in the candidate list. The single-process upper limit threshold is used to indicate whether a single process is good enough. If the scores of all candidate processes are lower than this threshold, it means that the defect conditions of this pipe section cannot be solved by any single process, and multiple processes need to be combined. For example, the substrate can be reinforced by first using ultraviolet light to cure the entire pipe and then the local mutation points can be treated by point-curing in situ. At least two candidate processes for the combination scheme are selected from the top-scoring candidates to ensure that the feasibility of the combination scheme is better than that of a single low-scoring process.

[0099] This synergy between batch clustering and combined scheme construction means that the output of the review stage is no longer limited to a simple mapping of one process for each defect, but rather focuses on the simultaneous optimization of construction costs and construction quality for the entire pipeline network. This significantly reduces the fixed costs of repeated manhole visits in large-scale pipeline repair projects.

[0100] S2. Perform hard rule screening on the defect parameter set and pipe segment feature set to obtain a candidate process set. The hard rule screening excludes unsuitable processes from the preset repair process set according to preset safety constraints. The preset repair process set includes at least point-like in-situ curing process, ultraviolet light pipe curing process and pipe fragmentation method process.

[0101] Reference Figure 3 Hard rule screening is the first layer of the three-layer hybrid architecture of the decision-making method, serving as a safety fallback. Pre-defined safety constraints are a set of fixed hard constraint rules. Each rule maps a combination of defect parameters and pipe segment characteristics to an exclusion conclusion regarding which processes are absolutely unusable under that combination. If hard rule screening is not performed first, and all processes in the pre-defined repair process set are directly scored, the scoring function, while providing a ranking result, cannot guarantee the safety and usability of the lowest-ranked process. A high-scoring process that actually violates a safety constraint might be recommended to the construction team, leading to construction failure or even a safety accident. Hard rule screening first eliminates obviously unsuitable processes, ensuring that the subsequent scoring function only optimizes within the safe process range.

[0102] The input to S2 is the set of defect parameters and the set of pipe segment features output by S1, and the output is the set of candidate processes. The set of candidate processes is the remaining subset of the preset repair process set after excluding unsuitable processes, and it is the scoring object of the scoring function of S3.

[0103] Typical examples of pre-set safety constraints include matching rules between defect level and the load-bearing capacity of the repair process, matching rules between pipe material and resin adhesion, and matching rules between pipe diameter and the space required for repair equipment. For example, in a severe collapse scenario with defect level 4, point-curing in-situ curing is excluded because the strength of the lining patch cured in-situ is insufficient to restore the load-bearing capacity of the collapsed pipe section; in a micro-diameter scenario with a pipe diameter less than DN300, UV whole-pipe curing is excluded because the size of the UV lamp holder exceeds the passable range of that pipe diameter.

[0104] For example, for defect 2 in the baseline scenario, the defect type is whole-section corrosion deformation and the defect level is 3. Hard rule screening excludes point-based in-situ curing processes based on the safety constraint that defects of level 3 or higher are not applicable. For pipe diameter DN500 and reinforced concrete pipes, there are no exclusion constraints for either UV whole-pipe curing or the pipe-fragmentation method. Therefore, the candidate process set output for defect 2 consists of two items: UV whole-pipe curing and the pipe-fragmentation method.

[0105] This hard rule, which provides a safety net before scoring, significantly reduces the risk of failure of the decision-making method under boundary conditions, giving the method both the flexibility of a scoring function and the reliability of a rule system.

[0106] In some embodiments, the preset repair process set also includes a temporary repair as a fallback option. Temporary repair is not a proactive repair process type, but rather a decision made based on on-site assessment that the defect does not affect the pipeline's flow function at the current level, and therefore will not be repaired, instead proceeding to periodic monitoring and re-inspection. When machine learning anomaly review determines, based on historical decision samples, that the current defect should not be repaired, temporary repair can be output as the decision result.

[0107] In some embodiments, the hard rule screening in S2 employs a dual-threshold mechanism, which includes a hard threshold and a soft threshold. If only a single hard threshold is set, the hard rule either excludes or allows the process, lacking the ability to handle boundary conditions between obvious mismatch and obvious fit. In such boundary conditions, experts might, based on experience, believe that certain processes are still usable, but the hard threshold mechanism would eliminate them outright. The dual-threshold mechanism divides the determination of hard rules into two levels: a hard threshold for strong exclusion and a soft threshold for manual review. This ensures that serious defects directly converge to the specified processes to guarantee a safety net, while allowing for flexibility in expert judgment through manual review of boundary conditions.

[0108] When a defect level in the defect parameter set reaches a hard threshold, a hard rule screening process is used to identify unsuitable processes from the preset repair process set according to the preset exclusion rules corresponding to the hard threshold. This level corresponds to the judgment of obvious incompatibility. When the defect level reaches this threshold, it means that the inherent carrying capacity or repair mechanism of the process can no longer handle the defect, and it must be forcibly excluded.

[0109] If a defect level in the defect parameter set reaches a soft threshold but not a hard threshold, the hard rule screening will mark some processes in the preset repair process set as processes awaiting manual review according to the preset marking rules corresponding to the soft threshold. These processes are not yet determined to be unsuitable. This level corresponds to a boundary judgment where the defect level is in the middle range that experts can discuss, and whether the process is usable needs to be confirmed by experts based on the actual site conditions.

[0110] When a process awaiting manual review is deemed incompatible, it is classified as an incompatible process and is ultimately removed from the candidate process set.

[0111] When a process awaiting manual review is deemed suitable, it will be retained in the candidate process set. At this point, the process awaiting manual review will be retained as a candidate that can participate in the scoring and ranking.

[0112] In the baseline scenario, when determining the level 2 local fracture of defect 1, it is assumed that the hard threshold is defect level 3 and the soft threshold is defect level 2. Defect 1's level 2 defect does not reach the hard threshold but does reach the soft threshold. The hard rule filtering, based on the preset marking rules corresponding to the soft threshold, marks the fragmentation pipe method in the candidate process set as a process awaiting manual review. Whether the fragmentation pipe method's broken old pipe characteristics are necessary in a moderate fracture scenario at defect level 2 depends on expert judgment, such as whether there is an on-site expansion requirement. This process awaiting manual review is pushed to the review terminal for expert confirmation. If the expert confirms its incompatibility based on the lack of on-site expansion requirements, it is included in the incompatible process set and removed from the candidate process set; if the expert confirms it is still usable, it is retained in the candidate process set.

[0113] The dual threshold mechanism enables the decision-making method to remain fully automated under conditions where the defect level is clearly determined, while also incorporating expert judgment under boundary conditions, thus combining decision-making flexibility with safety safeguards.

[0114] In some embodiments, manual review is implemented by pushing the process to be manually reviewed, along with its corresponding set of defect parameters and pipe segment feature sets, to a review terminal. Reviewers then input the appropriate or inappropriate review results, and the results are returned to the hard rule filtering module. The review terminal is typically a tablet controller used on-site by the construction team or a browser interface in an office. After receiving the push notification, reviewers retrieve detailed information about the process to be manually reviewed, make a judgment based on on-site observation and historical experience, and input either an appropriate or inappropriate result into the interface. The system then automatically sends this result back to the hard rule filtering module.

[0115] In some embodiments, hard rule filtering in S2 performs two-level splitting based on the defect type in the defect parameter set.

[0116] The first layer determines the defect branch based on the defect type. The defect branches include at least local defect branches and whole-section defect branches. Local defect branches correspond to defect types that exist only in local locations of the pipe section, such as cracks, misalignments, and pitting corrosion. Whole-section defect branches correspond to defect types that occur over a large area along the axial direction of the pipe section, such as corrosion deformation and collapse of the entire pipe section.

[0117] The second layer marks each process in the preset repair process set as either a preferred or non-preferred process based on the defect branch. Specifically, point-like in-situ curing processes under local defect branches are marked as preferred, while ultraviolet light tube-wide curing and fragmentation methods under whole-section defect branches are marked as preferred. The remaining processes in the preset repair process set are marked as non-preferred. This marking mechanism naturally aligns the decision structure for hard rule filtering with the essential applicable objects of the two types of repair processes: point-like repair and whole-section renewal.

[0118] Hard rule screening identifies unsuitable processes from a pre-defined set of repair processes based on non-preferred flags and defect levels in the defect parameter set. A non-preferred flag means that the process is not the first choice under the current defect type branch. Combined with the defect level determination, processes marked as non-preferred are more likely to be directly excluded when the defect level is higher; processes marked as non-preferred may still be retained in the candidate process set as alternatives when the defect level is lower.

[0119] In the baseline scenario, the local cracking of defect 1 and the local misalignment of the interface of defect 3 both fall into the local defect branch. The point-like in-situ curing process under the branch is marked as the preferred mark, while the UV whole tube curing process and the fragmentation tube method process are marked as the non-preferred mark. The whole-section corrosion deformation of defect 2 falls into the whole-section defect branch. The UV whole tube curing process and the fragmentation tube method process under the branch are marked as the preferred mark, while the point-like in-situ curing process is marked as the non-preferred mark.

[0120] In some embodiments, the preset safety constraints used for hard rule filtering are managed by version number, and new repair processes are registered to the preset repair process set in the form of process description plugins. The process description plugins include at least the preset safety constraints corresponding to the new repair process, the scoring row corresponding to the new repair process in the scoring function, and the recommended field of process parameters for the new repair process.

[0121] Preset safety constraints are continuously revised as industry standards are updated and on-site construction experience is accumulated. Version number management allows decision-making methods to clearly identify which version of safety constraints is currently in use and supports rollback to historical versions, facilitating auditing and accountability. When the industry releases a new version of the standard or new safety experience is accumulated on the same site, the new rules are entered into the database with the new version number, and the old rules are archived after they are no longer referenced.

[0122] When new trenchless repair technologies emerge in the industry, such as spiral winding repair and mechanical point repair, no modifications are needed to the main program of the decision-making method. Only three types of fields need to be submitted as a process description plugin: the preset safety constraints corresponding to the new technology for hard rule filtering and incompatibility determination; the score row of the new technology in the scoring function for participating in candidate technology scoring; and the recommended process parameters field for returning complete construction parameters when S5 queries the preset process parameter library. After loading the new plugin, the main program of the decision-making method automatically incorporates it into the preset repair technology set. The entire integration process requires no code rewriting and no system downtime or restart.

[0123] This combination of versioning and plug-in architecture enables the decision-making method to evolve continuously, providing strong support for the industry's ongoing innovation in repair technology.

[0124] In some embodiments, the point-like in-situ curing process in the preset repair process set is implemented by impregnating a pre-cut glass fiber fabric with resin, wrapping it around an inflatable airbag, guiding it to the point to be repaired in the pipeline by a pulling device or robot, inflating the airbag to press the resin-impregnated fiber cloth onto the pipeline defect, maintaining pressure and allowing the resin to cure under normal temperature or heating conditions to form an inner lining patch.

[0125] The physical process is as follows: after the fiberglass fabric is cut into a shape that matches the size of the defect, it is immersed in two-component epoxy resin or polyester resin to fully saturate it with resin; the saturated fiber cloth is flexibly wrapped around the outside of the inflatable airbag; a pulling device or robot delivers the airbag from the inspection well to the defect location; the airbag is inflated and presses the resin-impregnated fiber cloth onto the pipe wall at the defect location; pressure is maintained for several hours to allow the resin to cure at the ambient temperature inside the pipe or to accelerate curing through a heating device; after curing, the airbag is deflated and extracted, leaving a ring-shaped or rectangular resin composite lining patch that fits the defect location.

[0126] When performing point-like in-situ curing on defect 1 in the benchmark scenario, the construction team first estimated the required size of the fiberglass cloth based on the location parameters of defect 1 at 22 meters from the wellhead and at the 9 o'clock position. After impregnating it with resin, the cloth was wrapped in an airbag and sent to the location. The airbag was inflated, pressed, and cured under pressure to obtain the inner lining patch covering the rupture location.

[0127] Room temperature curing is suitable for applications where the temperature inside the pipe is higher than the resin's critical curing temperature. It has a longer curing time but requires no heating equipment. Heat curing is suitable for applications where the temperature inside the pipe is lower than the resin's critical curing temperature, such as during low-temperature construction in winter. In this case, a heating element on the outside of the airbag increases the curing temperature to shorten the curing time. The two curing methods can be switched on-site based on the actual measured temperature inside the pipe.

[0128] In some embodiments, the UV-curing process in the preset repair process set is implemented by pulling a glass fiber hose pre-impregnated with photosensitive resin from the inspection well into the pipe to be repaired using a pulling device, filling the hose with compressed air to make the hose fit tightly against the inner wall of the original pipe, and moving the UV lamp holder along the pipe to irradiate the photosensitive resin with UV light of a specific wavelength to trigger cross-linking and curing of the photosensitive resin within a few hours to form a pipe-in-pipe composite structure.

[0129] The physical process is as follows: The fiberglass hose is composed of multiple layers of fiberglass mat with photosensitive resin sandwiched inside, forming a layered structure. Before leaving the factory, the hose is impregnated with photosensitive resin and transported to the site in light-proof packaging. The pulling equipment pulls the hose from one end of the inspection well into the section of pipe to be repaired. The hose diameter is slightly smaller than the original pipe's inner diameter to ensure smooth passage. Compressed air is injected into the hose to cause it to expand and adhere tightly to the inner wall of the original pipe. The air pressure is continuously maintained within the process setting range to ensure tight adhesion. Under the guidance of the pulling equipment, the ultraviolet lamp holder moves along the pipe axis. The ultraviolet lamp tubes on the lamp holder irradiate the inner wall of the hose with ultraviolet light of a specific wavelength. The photosensitive resin undergoes a cross-linking and curing reaction under the action of ultraviolet light. After several hours of cross-linking and curing, the hose forms a pipe-in-pipe composite structure that matches the inner diameter of the original pipe. The corroded part of the original pipe is isolated by the newly formed inner lining, and the structural integrity is restored.

[0130] In the baseline scenario, when performing UV curing on defect 2, the construction team first prepared a fiberglass flexible tube, slightly longer than 13 meters, to cover the section from 45 to 58 meters from the manhole to ensure anchorage at both ends within the manhole. The flexible tube was pulled in from the downstream manhole and spread throughout the entire defect 2 section, then filled with compressed air to ensure it adhered tightly to the corroded inner wall of the DN500 reinforced concrete pipe. The UV lamp holder entered from one end and moved along the pipe at a preset speed, triggering the curing of the photosensitive resin on the inner wall of the flexible tube segment by segment. After curing, excess portions were cut off at both ends of the flexible tube and connected to the upstream and downstream pipe sections.

[0131] The timing of the UV lamp holder's movement along the conduit is coordinated with the cross-linking and curing time of the photosensitive resin. If the lamp holder moves too quickly, some resin will not cure fully; if it moves too slowly, cured sections will be repeatedly irradiated, causing uneven stress within the resin. The preset movement speed needs to be calibrated comprehensively based on the hose length, the required curing depth of the photosensitive resin, and the power of the UV lamp tube.

[0132] In some embodiments, the pipe-breaking method in the preset repair process set is implemented by inserting a conical pipe-breaking head with breaking teeth at the front end into the old pipe using a hydraulic jacking device, mechanically breaking the old pipe completely and squeezing the fragments into the surrounding soil to form a new channel, and simultaneously pulling in a high-density polyethylene short pipe that has been pre-assembled on the ground behind the pipe-breaking head.

[0133] The physical process is as follows: The hydraulic jacking equipment is lowered and anchored into one end of the inspection well. The end of the jacking equipment is connected to a conical pipe-breaking head. The breaking teeth at the front end of the pipe-breaking head rotate or advance under hydraulic drive. When the breaking teeth advance along the pipeline axis, the mechanical force acts on the old pipe wall, causing it to break into fragments. The fragments are pushed into the surrounding soil layer under the expansion and compression of the outer diameter of the conical surface of the pipe-breaking head, forming a new channel slightly larger than the outer diameter of the old pipe. A high-density polyethylene short pipe, which is pre-welded and spliced ​​on the ground, is pulled in synchronously behind the pipe-breaking head by steel cables, so that the pulling in of the new pipe and the breaking of the old pipe advance synchronously in the axial direction. When the pipe-breaking head reaches the inspection well at the other end, the entire old pipe is broken into the soil and the entire new pipe is in place. After the excess parts at both ends are cut off, it is connected to the upstream and downstream pipe sections.

[0134] In the baseline scenario, when selecting the pipe-fracture method as an alternative implementation for defect 2, the construction team deploys a hydraulic jacking device and a pipe-fracture head between the manholes at both ends of defect 2. The pipe-fracture head is inserted into the old pipe from one manhole and advanced along the pipe axis, breaking the entire DN500 reinforced concrete old pipe into concrete fragments that are squeezed into the surrounding soil. Simultaneously, a pre-assembled high-density polyethylene short pipe is pulled in from behind the pipe-fracture head. The diameter of the new pipe can be selected as DN500 or DN600 to achieve appropriate capacity expansion, depending on the needs.

[0135] The expansion amount is the difference between the outer diameter of the conical surface of the cracked pipe head and the inner diameter of the old pipe. Adjusting the expansion amount allows for the simultaneous adjustment of the new pipe diameter. In the baseline scenario, if the expansion amount is chosen to make the new pipe diameter reach DN600, it can provide a margin for future drainage flow growth in this pipe section while repairing it, an advantage that the other two processes cannot achieve. However, a larger expansion amount also means more significant disturbance to the surrounding soil layers, which needs to be comprehensively weighed in conjunction with the bearing sensitivity of the surface above the pipe section.

[0136] S3. Based on the defect parameter set and pipe segment feature set, obtain the score value corresponding to each candidate process in the candidate process set through the scoring function, and sort the candidate process set from high to low according to the score value to obtain the main recommended process and the alternative list.

[0137] The scoring function is the second layer of the three-layer hybrid architecture of the decision-making method, and it serves as the primary recommendation function. The scoring function combines the defect parameter set and the field of the pipe segment feature set with preset weights to evaluate the suitability of each candidate process, outputting a score between 0 and 1. The closer the score is to 1, the higher the suitability of the process for the current defect. The input to S3 is the candidate process set output by S2, and the output is a process sequence sorted by score. The first process in the sequence is the primary recommendation process, and the remaining positions constitute the alternative list.

[0138] The scoring function independently scores each candidate process in the candidate process set. The score for each process is based on a multi-dimensional weighted feature set, including the degree of fit between the defect type and the inherent applicable object of the process, the degree of match between the defect level and the repair capability of the process, the degree of match between the pipe diameter and the space margin of the equipment for the process, and the degree of match between the pipe material and the adhesion of the repair material for the process. The final score is obtained by summing the weighted values ​​of each dimension according to preset weights and normalizing the sum. The preset weights can be determined based on industry-standard process selection guidelines or fine-tuned based on statistical analysis of historical decision samples.

[0139] For example, for defect 2 in the baseline scenario, the candidate process set includes two options: UV-curing for the entire pipe and the pipe-fracture method. The scoring function for the UV-curing for the entire pipe comprehensively considers its inherent applicability to corrosion deformation of the entire pipe section, its equipment compatibility with DN500 pipe diameters, and its adhesion to the inner lining of reinforced concrete pipes, resulting in a score of 0.82. For the pipe-fracture method, comprehensively considering its inherent applicability to corrosion deformation of the entire pipe section, its compatibility with DN500 pipe diameter expansion construction, and its feasibility in fracturing reinforced concrete pipes, a score of 0.68 is obtained. After sorting by score from highest to lowest, the UV-curing for the entire pipe section is ranked first as the primary recommended process, while the pipe-fracture method is included in the candidate list.

[0140] The output of the scoring and ranking is interpretable for both the construction and supervision parties. The decision-making method not only provides recommendations but also specific score values ​​for each candidate process, enabling reviewers to trace the basis for the recommended processes and specifically examine the scores of the dimensions with the highest weights in the scoring function for questionable scenarios. This interpretability is a significant advantage compared to pure end-to-end machine learning decision-making.

[0141] In some embodiments, the execution of the scoring function in S3 is divided into two levels: coarse screening and fine evaluation. Coarse screening performs preliminary scoring on each candidate process in the candidate process set according to a preset rapid discrimination rule to exclude obviously unsuitable candidate processes. Fine evaluation calculates the score value of the candidate processes retained by coarse screening according to the complete scoring model.

[0142] The coarse screening and rapid discrimination rule is a simplified version of the scoring function. It uses only two highly sensitive fields—defect type and defect level—to perform a fast decision tree judgment, assigning one of three labels to each candidate process: obviously suitable, boundary, or obviously unsuitable. Candidate processes labeled "obviously unsuitable" are directly excluded at this level and do not proceed to the detailed evaluation. Candidate processes labeled "obviously suitable" or "boundary" proceed to the detailed evaluation. The detailed evaluation uses the complete scoring model to calculate the final score, incorporating all fields such as defect distance, clock orientation, pipe diameter, and pipe material, as well as video frame-derived features for multi-dimensional weighting.

[0143] This two-level scoring system eliminates the need for many clearly unsuitable candidate processes to enter the full scoring model, significantly reducing the computational load on the full scoring model. The rules for rapid coarse screening can be set based on a pre-defined rule base with safety constraints; clearly defined rules can be mapped to coarse screening labels without the need for additional rule definition.

[0144] In some embodiments, a two-sided approach alarm determination is performed after S3 and before S4. The two-sided approach alarm determination is a joint determination of the outputs of the scoring function and hard rule filtering, used to identify extreme conditions where none of the processes in the preset repair process set are applicable and trigger manual processing.

[0145] An alarm is triggered when the candidate process set is not empty, the score values ​​corresponding to all candidate processes in the candidate process set are lower than the preset alarm lower limit threshold, and the number of unsuitable processes excluded from the preset repair process set by hard rules reaches the preset exclusion quantity threshold. The main recommended process is then designated for manual processing. The preset alarm lower limit threshold is the lower bound of the score value; below this value means that although the process has not been forcibly excluded by hard rules, its actual suitability for the current defect is extremely low. The preset exclusion quantity threshold is the lower bound of the number of processes that have been excluded by hard rules; reaching this number means that multiple processes have been excluded by hard rules.

[0146] If only one-sided alarms are set, such as using only a low score from the scoring function as the alarm condition, false alarms may occur due to insufficient input features. The scoring function has limited tolerance for missing or abnormal fields in the defect parameter set, and a single low score may reflect a problem at the input end rather than the inapplicability of the process itself. If only the number of exclusions by hard rules is used as the alarm condition, it may unnecessarily interrupt the decision-making process when the scoring function still provides a reasonable recommendation. Two-sided approximation requires that conditions on both sides be met simultaneously, ensuring that an alarm is triggered only when both sides confirm that the process is generally inapplicable.

[0147] For example, suppose another scenario involves a severe collapse defect in a DN200 ductile iron pipe. The defect type is localized collapse, defect level 4, pipe diameter DN200, and pipe material ductile iron. Hard rule filtering excludes the point-like in-situ curing process because it is inapplicable for defect level 4; it also excludes the UV whole-pipe curing process because the DN200 pipe diameter is too small for the UV lamp holder to access. The number of incompatible processes reaches two, exceeding the preset exclusion threshold. Only the pipe-fracture method remains as a candidate process, but its feasibility for breaking ductile iron pipes is poor, resulting in a score of 0.18, below the preset alarm threshold of 0.30. Since both conditions are met, an alarm is triggered, and the recommended primary process is switched to manual processing.

[0148] The dual-sided approximation avoids false alarms caused by insufficient input features in the single-sided approximation, and can accurately identify extreme working conditions that truly require expert intervention, thereby improving the signal-to-noise ratio of alarms.

[0149] S4. Perform machine learning anomaly verification on the main recommended process to obtain the anomaly verification results of the main recommended process. The anomaly verification results shall include at least anomaly labels and output confidence scores.

[0150] Machine learning anomaly verification is the third layer in the three-layer hybrid architecture of the decision-making method, responsible for anomaly verification. The hard rule filtering and scoring functions in the first two layers are based on manually labeled rules and weights, providing reasonable results for operating conditions covered by the rule base. However, when boundary conditions not covered by the rule base occur on-site, the rules and weights may give seemingly reasonable but actually incorrect recommendations. Machine learning anomaly verification, based on the statistical patterns of a large number of historical decision samples in the training library, performs additional anomaly detection on the current main recommendation process, identifying situations where the rules and scores are deemed reasonable, but the re-examination results were poor under similar historical operating conditions.

[0151] The input to S4 is the main recommended process output from S3, along with the corresponding defect parameter set and pipe segment feature set. The output is the anomaly verification result for the main recommended process. This result includes at least two items: anomaly labeling and output confidence. The anomaly labeling is used for binarization to determine whether the main recommended process appears abnormal, while the output confidence quantifies the machine learning model's grasp of this determination. Together, these constitute the basis for the judgment triggered by the downstream S5 output and the cross-pipe segment similarity analogy correction mechanism.

[0152] In the baseline scenario, the primary recommended process for defect 1 is point-based in-situ curing. S4 inputs this primary recommended process along with the defect parameter set for defect 1, including level 2 localized cracking, 22 meters, and 9 points, as well as the pipe segment feature set, including DN500 and reinforced concrete, into the machine learning model. The model performs anomaly detection on this combination based on the statistical patterns of thousands of historical decision samples in the training library, labeling anomalies as normal and outputting a confidence level of 0.91. This means that the model considers the primary recommended process to be highly consistent with successful decisions under similar historical conditions, indicating a high level of confidence.

[0153] When an anomaly is marked as normal, the primary recommended process proceeds to the S5 output stage after review in S4. When an anomaly is marked as abnormal, it means that the machine learning model has identified, based on historical experience, that the primary recommended process may not be suitable for the current operating conditions. Downstream, a dual-approach alarm judgment or a cross-pipe segment similarity analogy correction mechanism may be triggered. When the output confidence is low, a cross-pipe segment similarity analogy correction mechanism is triggered to use the decisions of historically similar pipe segments as reference weights.

[0154] In some embodiments, the anomaly marker in the anomaly review result is a binary marker, determined by comparing the output probability of the machine learning model for the main recommended process with a preset anomaly judgment threshold. The output confidence level in the anomaly review result is the probability value of the machine learning model's judgment result for the main recommended process.

[0155] The machine learning model outputs a probability value between 0 and 1 for the combination of the main recommended process, defect parameter set, and pipe segment feature set. This probability value represents the probability that the combination is a normal recommendation in the statistical sense of historical decision samples. When the probability value is higher than a preset anomaly judgment threshold, it is marked as normal; when the probability value is lower than or equal to the preset anomaly judgment threshold, it is marked as abnormal. The preset anomaly judgment threshold is usually determined according to the statistical distribution of historical decision samples, for example, taking the lower quintile of the probability distribution of historical successful re-inspection samples.

[0156] Anomaly flags and output confidence scores serve different purposes in downstream processes. Anomaly flags directly determine whether the primary recommended process needs to be interrupted and enter the alarm branch. When the anomaly flag is abnormal, the primary recommended process is considered unreliable, and S5 outputs an alarm or forwards the result to manual processing. Output confidence scores are used to trigger a correction mechanism based on cross-pipe segment similarity analogy. When the output confidence score is lower than a preset confidence threshold, even if the anomaly flag is normal, the primary recommended process is further corrected using historical decisions from similar pipe segments as a reference weight. Both fields jointly support downstream conditional branch decisions.

[0157] In some embodiments, the machine learning anomaly review in S4 is performed based on a machine learning model and a training library. The machine learning model is trained based on historical decision samples stored in the training library. The training library stores multiple historical decision samples, and each historical decision sample contains at least a set of historical defect parameters, a set of historical pipe segment features, historical decision results, and a re-inspection success rate.

[0158] The training library is a structured storage container for historical decision samples. Each historical decision sample corresponds to a past repair decision and construction execution. The historical defect parameter set records all fields of the defect at that time, the historical pipe segment feature set records all fields of the pipe segment at that time, the historical decision result records the recommended main process at that time, and the re-inspection success rate records the percentage of successful re-inspections of the defect repair effect after the repair is completed. The machine learning model uses these fields as input features and supervision signals for training, learning the statistical patterns of which defect, pipe segment, and process combinations correspond to higher re-inspection success rates.

[0159] The relationship between the training library and the machine learning model is as follows: the training library provides the raw data, and the machine learning model is trained based on the training library. After the machine learning model is trained, it is deployed to S4 to perform anomaly verification on the current main recommended process. Specifically, it takes the defect parameter set of the current defect, the pipe segment feature set, and the main recommended process output by S3 as input, and outputs the anomaly label and confidence level of this combination. After the anomaly verification is executed by S4, the anomaly verification result is returned as the basis for the judgment of S5.

[0160] In the baseline scenario, the statistical patterns learned by the machine learning model from thousands of historical decision samples in the training library include a historical re-inspection success rate of over 0.85 for the primary recommended process of point-based in-situ curing for a DN500 reinforced concrete pipe with a localized fracture level 2 defect combination. When S4 receives the current defect 1 as input and the primary recommended process is point-based in-situ curing, the model outputs an anomaly label as normal with a confidence level of 0.91 based on the above statistical patterns. If the primary recommended process for the current defect 1 is the pipe-fracture method, the model will output an anomaly label because the success rate of the localized fracture level 2 plus point-based in-situ curing combination in historical samples is much higher than that of the localized fracture level 2 plus pipe-fracture method combination.

[0161] The larger the training library and the more combinations of working conditions covered by historical decision samples, the higher the statistical confidence of the machine learning model. In the early application stages of the decision-making method, when the training library is limited, the machine learning model may give lower output confidence for unseen combinations of working conditions. In this case, a cross-segment similarity analogy correction mechanism is used to supplement and strengthen the model. A three-layer hybrid architecture is formed by historical machine learning review, hard rule screening based on manual rules, and a scoring function based on manual weights. The machine learning layer does not replace the rule layer and the scoring layer; rather, it performs additional statistical verification based on historical experience, after the rule layer and the scoring layer have determined the results to be reasonable.

[0162] In some embodiments, when the output confidence of the machine learning anomaly review in S4 is lower than a preset confidence threshold, a cross-segment similarity analogy correction operation is also performed on the main recommended process, such as... Figure 6 As shown, when the machine learning model has low confidence in its judgment of the current defect, it means that there are few historical decision samples with similar working conditions in the training library, and the statistical output of the model alone may be unreliable. At this time, cross-pipe section similarity analogy is introduced, and historical decision samples with similar working conditions but not completely identical in the training library are used as reference weights to correct the current main recommended process.

[0163] Based on historical decision samples stored in the training library, a comprehensive similarity is calculated between each historical decision sample and the current defect, using two sets of features: a historical defect parameter set and a historical pipe segment feature set. The comprehensive similarity is obtained by weighting the distance metrics of the two sets of features. For example, in the defect parameter set, defect type is represented by category-coded distance, defect level by numerical distance, and defect distance and clock position by normalized distance; in the pipe segment feature set, pipe diameter is represented by normalized numerical distance, and pipe material by category-coded distance. The normalized distances of the two sets of features are then weighted, summed, and the negative exponent is taken to obtain the final comprehensive similarity, with a similarity value between 0 and 1.

[0164] Sort the historical decision samples by comprehensive similarity from highest to lowest, select a predetermined number of historical decision samples as analogy samples, and use the re-examination success rate of the analogy samples as a secondary weighting coefficient. Weighted fusion is then applied to the historical decision results corresponding to the analogy samples to obtain the analogy recommendation process. The analogy samples are the records most similar to the current defect in the historical decision samples. The secondary weighting coefficient ensures that analogy samples with high re-examination success rates have a greater impact on the fusion result; that is, successful historical experiences are given greater weight, while unsuccessful historical experiences are given less weight, avoiding the reapplication of failed historical decisions to the current defect. The specific method of weighted fusion can be either weighted voting (i.e., the historical decision result that receives the most votes) or weighted averaging (i.e., taking a weighted average when quantifying the decision results); ultimately, a specific analogy recommendation process is output.

[0165] The primary recommended process is modified based on the analogous recommended process. If the analogous recommended process is the same as the original primary recommended process, the modification does not occur, and the primary recommended process remains unchanged. If the analogous recommended process is different from the original primary recommended process, the original primary recommended process is replaced by the analogous recommended process, which becomes the modified primary recommended process and proceeds to S5.

[0166] When performing S4 anomaly review on defect 3 in the baseline scenario, assuming that there are few historical decision samples in the training library for the same working condition of local interface misalignment level 1 plus DN500 reinforced concrete, the confidence score of the machine learning model output for this combination is 0.45, which is lower than the preset confidence threshold of 0.60. The system triggers cross-pipe segment similarity analogy correction. Eight of the most similar historical samples are selected from the training library according to comprehensive similarity as analogy samples. These samples may come from local interface misalignment decisions of other pipe diameters, local defect decisions of the same pipe diameter but different defect types, etc. The historical decision results and re-inspection success rates of each analogy sample are extracted separately. Among them, the historical decision results of 6 samples are point-like in-situ curing process, with an average re-inspection success rate of 0.88; the historical decision results of 2 samples are postponed repair, because the interface misalignment level is low and does not affect the flow, so construction and repair are not carried out temporarily, with an average re-inspection success rate of 0.95. After a second weighting based on the re-inspection success rate, the analogy recommended process is point-like in-situ curing process.

[0167] Cross-pipeline analogy is an effective way to enhance machine learning models when there is insufficient data in the early stages of training. Machine learning models that rely directly on a small number of samples from the same operating conditions in the training library are prone to overfitting or underfitting on those conditions. Cross-pipeline analogy expands the scope of similar operating conditions, so that decisions are not limited to completely identical historical experience. This mitigates the risk of model overfitting when there is insufficient data in the early stages of training, and enables the decision-making method to provide relatively robust results in the early stages of operation.

[0168] In some embodiments, the machine learning model is a classification model trained based on historical decision samples. The input of the machine learning model is the feature vector of the defect parameter set and the pipe segment feature set. The output of the machine learning model is the anomaly label and output confidence score corresponding to the main recommended process.

[0169] The training process of the classification model is as follows: the historical defect parameter set, historical pipeline feature set, and historical decision result of each historical decision sample in the training library are concatenated into a feature vector. The re-examination success rate of the sample is binarized as a supervision label, where the value above the threshold is marked as normal and the value below the threshold is marked as abnormal. The model parameters are iteratively updated by minimizing the cross-entropy loss function so that the model's normal or abnormal classification output of the training samples matches the supervision label. After training, the model also outputs the probability of normal or abnormal for new input feature vectors that have not been seen before, and then obtains the abnormal label and output confidence from the probability.

[0170] The specific type of classification model can be selected during engineering implementation. In addition to the standard training process described above, options include using a support vector machine model for classification via the maximum margin hyperplane, a random forest model for classification via multi-decision tree voting, or a gradient boosting tree model for classification via stacked weak classifiers. These three alternative models also use the feature vectors of the defect parameter set and the pipe segment feature set as input, and output normal or abnormal probabilities. They are compatible with the S4 machine learning anomaly verification interface in this case, and therefore also fall within the scope of machine learning model implementation methods.

[0171] In some embodiments, the output of the machine learning model is further processed by inverse rule extraction to generate an accompanying explanatory rule set. The accompanying explanatory rule set fits the output of the machine learning model to historical decision samples in the form of a decision tree and is updated synchronously with the rolling update of the machine learning model.

[0172] The reverse rule extraction process is as follows: Using historical decision samples from the training library as input, and the normal or abnormal output of the currently deployed machine learning model for each sample as the target, a shallow decision tree is trained to fit the correspondence between the input and the target. Each branch node of the decision tree is a simple judgment that feature X is greater than threshold Y, and each leaf node is a normal or abnormal label. The judgment path of the shallow decision tree, expressed in Chinese characters, constitutes the accompanying interpretation rule set. For example, a rule might be: if the defect level is greater than or equal to 3 and the pipe diameter is less than DN400, then the point-like in-situ curing process is judged as abnormal.

[0173] The AND-OR statements in decision trees are human-readable, providing an approximately interpretable representation of the black-box decision-making process of machine learning models. Supervisors and on-site technical personnel from the construction team can review the accompanying explanatory rule set to understand the basis for the model's anomaly labeling, without needing to access the complex weight matrix of the machine learning model itself. The interpretability of machine learning decisions is maintained even after the introduction of the machine learning layer.

[0174] When the machine learning model is updated in a rolling manner driven by feedback write-back, the interpretation rule set is also synchronously regenerated through decision tree fitting, ensuring that the rule set is always consistent with the currently deployed model version. This synchronous update ensures that interpretability always follows the model's evolution and does not become invalid due to model upgrades.

[0175] S5. Based on the main recommended process, query the preset process parameter library to obtain the recommended process parameters corresponding to the main recommended process, and output the main recommended process, alternative list and recommended process parameters according to the anomaly review results.

[0176] The preset process parameter library is a pre-calibrated table of process parameters, storing the key parameters required for the construction of each process in the preset repair process set. The recommended process parameters are specific construction parameters retrieved from the preset process parameter library by the main recommended process index, serving as a direct basis for the construction team's on-site execution. The input to S5 is the main recommended process and anomaly verification results output by S4, and the output consists of three items: the main recommended process that the construction team can directly execute, a list of alternative processes, and recommended process parameters.

[0177] The execution flow of S5 is as follows: Using the primary recommended process as the index key, the preset process parameter library is queried to obtain the corresponding recommended process parameters; then, the output format is determined by combining the anomaly review results from S4. When the anomaly mark in the anomaly review result is normal, the primary recommended process, alternative list, and recommended process parameters are output for the construction party to execute; when the anomaly mark is abnormal, the result of manual processing is output, and the primary recommended process, alternative list, and recommended process parameters are still output together as a reference for manual review.

[0178] In the baseline scenario, the primary recommended process for defect 1 is point-based in-situ curing. S5 uses point-based in-situ curing as the index key to query the preset process parameter library, obtaining recommended process parameters of 1 ring and 3 hours for airbag pressure holding time. S4 marks the abnormal output of defect 1 as normal. Following the downstream state after expert review and confirmation of the incompatibility of the fragmented tube method in the aforementioned baseline scenario with a dual-threshold mechanism, the candidate process set for defect 1 converges to two items: point-based in-situ curing and UV-curing tube curing. Point-based in-situ curing is the preferred marker for the local defect branch and has a higher score in the scoring function, thus ranking as the primary recommendation. UV-curing tube curing is the secondary option under the non-preferred marker for the local defect branch and is included in the candidate list. Therefore, S5 ultimately outputs: the primary recommended process is point-based in-situ curing, the candidate list includes UV-curing tube curing, and the recommended process parameters are 1 ring and 3 hours for airbag pressure holding time. Upon receiving this output, the construction team can implement the process on-site according to the recommended parameters.

[0179] In some embodiments, the preset process parameter library stores at least the number of rings and airbag holding time corresponding to the dot-shaped in-situ curing process, the hose section length and ultraviolet lamp holder moving speed corresponding to the ultraviolet light tube curing process, and the diameter expansion amount and hydraulic jacking speed corresponding to the tube fragmentation process.

[0180] The key parameter fields for the three processes reflect the key control variables for their respective construction.

[0181] The number of rings in the point-like in-situ curing process indicates the number of layers of the inner lining patch formed by airbag inflation and pressing. Multi-ring structures can improve repair strength. The airbag holding time indicates the time to maintain pressure after airbag inflation while waiting for the resin to cure, which is related to the resin type and the on-site temperature.

[0182] The hose length in the UV curing process refers to the length of the hose section covered by a single pulling operation. If it is too short, multiple splices are required; if it is too long, the pulling resistance will be too high. The UV lamp holder moving speed refers to the speed at which the lamp holder travels along the pipe, which is matched with the curing speed of the photosensitive resin.

[0183] The expansion amount in the pipe-breaking process refers to the expansion of the new pipe diameter relative to the old pipe, which is related to whether expansion is required; the hydraulic jacking speed refers to the speed at which the pipe-breaking head advances along the pipeline, which is related to the difficulty of breaking the old pipe.

[0184] In the baseline scenario, the recommended process parameters for defect 1 are as follows: process type: point-source in-situ curing; number of rings: 1; airbag holding time: 3 hours; impregnation resin type: epoxy resin. This example can be directly implemented after the construction team confirms it on-site.

[0185] The preset process parameter library's table structure supports adding new process parameter rows. When a new repair process is added to the preset repair process set via the process description plugin, the recommended process parameter fields in the plugin are directly written as new rows to the preset process parameter library without modifying the library's table structure definition. The field types of the new rows can differ from those of existing rows because different processes are concerned with different key parameters. For example, the spiral winding repair process is concerned with bandwidth and winding angle, rather than the number of loops or the amount of diameter expansion.

[0186] In some embodiments, the method further includes pushing the main recommended process, alternative list and process parameter recommendations output by S5 to the urban pipeline construction resource scheduling system, so that the urban pipeline construction resource scheduling system can automatically reserve the corresponding construction team and construction equipment according to the process type and lock the construction time window that avoids peak traffic.

[0187] The specific process of the push is as follows: After the decision-making method is output by S5, it packages the main recommended process, the alternative list, and the recommended process parameters into a structured data packet and sends it to the urban pipeline construction resource scheduling system through a preset interface protocol. After receiving the data packet, the scheduling system queries the matching construction teams according to the process type field of the main recommended process, queries the matching construction equipment according to the equipment requirement field in the recommended process parameters, and queries the historical traffic flow data of the surrounding area according to the address of the pipe section to be repaired attached to the data packet. Combining these three factors, a construction time window is determined and the reservation is locked.

[0188] Without this integration, the output of the decision-making process needs to be manually entered into the construction management system by the construction dispatcher. The dispatcher then manually checks the availability of construction teams, equipment inventory, and traffic conditions before making a construction time window arrangement. This manual connection between decision-making and execution is time-consuming. After integration, the entire process is completed automatically by the system within seconds, significantly reducing the manual connection between decision-making and execution, and substantially improving the overall efficiency of large-scale pipeline repair projects.

[0189] This integration can also be extended to interface with the city's integrated pipeline management platform. The city's integrated pipeline management platform is a city-level management platform that coordinates various underground pipelines such as water supply, gas, electricity, telecommunications, and drainage. The output of this decision-making method can be pushed to this platform simultaneously, so that the scheduling of construction time windows can avoid conflicts with the construction of other pipelines in the same area, further reducing the coordination complexity of urban road construction.

[0190] In some embodiments, after S5, a feedback write-back operation is also performed on the primary recommended process, such as... Figure 7 As shown. Feedback writing is the process of feeding back the actual effect data of each repair operation to the training database, so that the decision-making method forms a closed loop of decision-making, construction, re-inspection, writing, and re-decision.

[0191] After the repair is completed according to the recommended process, the repaired pipe section undergoes another video inspection. The success rate of the re-inspection is determined based on the results of this re-inspection. The re-inspection results are determined by comparing the defect parameter set before repair with the pipe wall condition after repair, including multiple indicators such as whether defects have been eliminated, whether new defects have appeared, and whether the repaired area is smooth. The re-inspection success rate is a value between 0 and 1, representing the actual success rate of this repair work. For example, the re-inspection success rate is 1 when the defect is completely eliminated and no new defects have appeared; when the defect is partially eliminated, the re-inspection success rate is the median value based on the degree of defect repair; when new defects appear after repair or the original defects are not eliminated, the re-inspection success rate is a lower value.

[0192] The defect parameter set and pipe segment feature set obtained from S1, the main recommended process output from S5, and the re-inspection success rate are written back to the training library as new historical decision samples. A new historical decision sample is a complete triple consisting of input features, decision results, and actual effects. Its structure is completely consistent with existing historical decision samples in the training library and can be directly appended as a new row. The write-back operation is completed through a preset data interface, eliminating the need for manual data entry.

[0193] The machine learning model is updated on a rolling basis based on the written-back training library. When the training library accumulates a certain number of new samples, the machine learning model re-executes the training process and refits the parameters based on the expanded training library. Typical triggering conditions could be an update every 100 new samples or an update at a fixed time each week. The updated model replaces the old model and continues to perform anomaly verification in S4.

[0194] After completing the point-like in-situ curing repair of defect 1 in the baseline scenario, the construction team again lowered the crawler into the pipe section to perform in-pipe video inspection of the repaired area 22 meters from the wellhead. The re-inspection results showed that the original rupture location had been completely covered by the resin lining patch, and no new defects were found; the re-inspection success rate was set at 0.95. The system used the defect parameter set of defect 1, the pipe section feature set, the main recommended process point-like in-situ curing, and the re-inspection success rate of 0.95 as a new historical decision sample and wrote it back to the training library. This new sample will be used in the training when the machine learning model is updated in subsequent rolling updates, further improving the confidence of anomaly verification for similar future working conditions.

[0195] Without feedback and write-back, the training library for the decision-making method only contains the initial historical data, failing to absorb newly occurring construction experience. The predictive ability of the machine learning model for actual repair effects will become ineffective over time, as factors such as improvements in construction techniques, upgrades in material formulations, and changes in the aging process of the pipeline network gradually render the initial historical data unrepresentative. With feedback and write-back forming a closed loop, the predictive ability of the machine learning model for actual repair effects is continuously calibrated, ensuring the robustness of the decision-making method in long-term operation.

[0196] The decision-making method for the aforementioned drainage pipeline repair process can be executed by computer equipment, such as... Figure 8 As shown. A computer device includes at least one or more processors, memory, and one or more application programs. The application programs are stored in the memory and configured to be executed by the processor to implement all the steps of the methods described above. The processor may be a general-purpose central processing unit or a dedicated graphics processing unit, the memory may be dynamic random access memory or a solid-state drive, and the application programs may be operating system-based executable programs or cloud-based microservice programs.

[0197] All steps of the above method can also be carried on a computer-readable storage medium. A computer program is stored on the computer-readable storage medium, and when the computer program is loaded and executed by a processor, it implements the steps of the above method. The computer-readable storage medium can be a read-only memory, a programmable read-only memory, an electrically erasable programmable read-only memory, flash memory, a magnetic disk, an optical disk, or a combination of the above media.

[0198] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0199] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A decision-making method for drainage pipeline repair technology, characterized in that, Includes the following steps: S1. Obtain the defect parameter set and pipe segment feature set output by the video detection inside the pipe. The defect parameter set includes at least the defect type, defect level, defect distance and clock position. The pipe segment feature set includes at least the pipe diameter and pipe material. S2. Perform hard rule screening on the defect parameter set and the pipe segment feature set to obtain a candidate process set. The hard rule screening excludes unsuitable processes from the preset repair process set according to preset safety constraints. The preset repair process set includes at least point-like in-situ curing process, ultraviolet light whole pipe curing process and pipe fragmentation method process. S3. Based on the defect parameter set and the pipe segment feature set, obtain the score value corresponding to each candidate process in the candidate process set through the scoring function, and sort the candidate process set from high to low according to the score value to obtain the main recommended process and the alternative list; S4. Perform machine learning anomaly verification on the main recommended process to obtain the anomaly verification result of the main recommended process. The anomaly verification result includes at least anomaly label and output confidence. S5. Based on the main recommended process, query the preset process parameter library to obtain the process parameter recommendation corresponding to the main recommended process, and output the main recommended process, the alternative list and the process parameter recommendation according to the anomaly review result.

2. The decision-making method for drainage pipeline repair process according to claim 1, characterized in that, The method employs a two-stage execution approach for steps S2, S3, and S4, comprising an initial judgment stage and a review stage. During the in-pipe video inspection process, the preliminary judgment stage is performed. The preliminary judgment stage takes the acquired partial defect parameter set and partial pipe segment feature set as input, and executes S2 and S3 in sequence to obtain a preliminary recommended process. S4 is not executed in the preliminary judgment stage. After the in-pipe video inspection is completed, the review stage is executed. The review stage takes the defect parameter set and the feature set of the entire pipe segment as input, and executes S2, S3 and S4 in sequence to obtain the main recommended process, the alternative list and the anomaly review result. The output of the review stage is used as the input of S5.

3. The decision-making method for drainage pipeline repair process according to claim 2, characterized in that, The review phase also includes: When multiple defects within the same pipe segment are identified from the set of defect parameters input during the review phase, clustering is performed on the multiple defects according to the main recommended process corresponding to each of the multiple defects to obtain at least one batch of the same process based on process affinity. The process affinity is jointly determined by the degree of commonality of construction equipment and the degree of continuity of construction sequence among the main recommended processes corresponding to each of the multiple defects. When the score value corresponding to all candidate processes in the candidate process set is lower than the preset single process upper limit threshold, a process combination scheme is obtained by assembling at least two candidate processes with the highest score values ​​in the candidate process set, and the process combination scheme is included in the candidate list.

4. The decision-making method for drainage pipeline repair process according to claim 1, characterized in that, A bilateral approach alarm determination is also performed after S3 and before S4: When the candidate process set is not empty, the score value corresponding to all candidate processes in the candidate process set is lower than the preset alarm lower limit threshold, and the number of unsuitable processes excluded from the preset repair process set by the hard rule screening reaches the preset exclusion quantity threshold, an alarm is triggered, and the main recommended process is determined to be transferred to manual processing.

5. The decision-making method for drainage pipeline repair process according to claim 1, characterized in that, The hard rule filtering in S2 employs a dual-threshold mechanism, which includes a hard threshold and a soft threshold: In response to the defect level in the defect parameter set reaching the hard threshold, the hard rule screening determines the unsuitable process from the preset repair process set according to the preset exclusion rule corresponding to the hard threshold; In response to the defect level in the defect parameter set reaching the soft threshold but not the hard threshold, the hard rule screening marks some processes in the preset repair process set as processes to be manually reviewed according to the preset marking rule corresponding to the soft threshold. The processes to be manually reviewed are not temporarily determined as the unsuitable processes. When the process to be manually reviewed is determined to be incompatible after manual review, the process to be manually reviewed is identified as the incompatible process. When a process awaiting manual review is deemed suitable after manual review, the process awaiting manual review is retained in the candidate process set.

6. The decision-making method for drainage pipeline repair process according to claim 1, characterized in that, The hard rule filtering in S2 performs a two-level splitting based on the defect type in the defect parameter set: The first layer determines the defect branch according to the defect type, and the defect branch includes at least a local defect branch and a complete defect branch. The second layer marks each process in the preset repair process set as a non-preferred or preferred mark according to the defect branch. The point-like in-situ curing process under the local defect branch is marked as the preferred mark, the ultraviolet light tube curing process and the fragmentation tube method process under the whole segment defect branch are marked as the preferred mark, and the remaining processes in the preset repair process set are marked as non-preferred marks. The hard rule screening determines the unsuitable process from the preset repair process set based on the non-preferred marker and the defect level in the defect parameter set.

7. The decision-making method for drainage pipeline repair process according to claim 1, characterized in that, The machine learning anomaly review in S4 is performed based on a machine learning model and a training library. The machine learning model is trained based on historical decision samples stored in the training library. The training library stores multiple historical decision samples, and each historical decision sample contains at least a set of historical defect parameters, a set of historical pipe segment features, historical decision results, and a re-inspection success rate. Following step S5, the following feedback write-back operation is also performed on the primary recommended process: After the repair is completed according to the main recommended process, the in-pipe video inspection is performed again on the repaired pipe section, and the re-inspection success rate is determined based on the result of the re-inspection. The defect parameter set, pipe segment feature set, main recommended process, and re-inspection success rate obtained in S1 are written back to the training library as new historical decision samples. The machine learning model is updated on a rolling basis based on the training library after being written back.

8. The decision-making method for drainage pipeline repair process according to claim 7, characterized in that, When the output confidence of the machine learning anomaly verification in S4 is lower than a preset confidence threshold, the following cross-segment similarity analogy correction operation is also performed on the main recommendation process: Based on the historical decision samples stored in the training library, the comprehensive similarity between each historical decision sample and the current defect is calculated from two sets of features: the historical defect parameter set and the historical pipe segment feature set. Sort the historical decision samples from largest to smallest according to the comprehensive similarity, select a preset number of historical decision samples as analogy samples, use the re-examination success rate corresponding to the analogy samples as a secondary weighting coefficient, and perform weighted fusion on the historical decision results corresponding to the analogy samples to obtain the analogy recommendation process; The main recommended process is modified based on the analogy-based recommended process.

9. A computer device, characterized in that, It includes: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to: execute a decision method for a drainage pipe repair process according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement: the decision-making method for the drainage pipeline repair process as described in any one of claims 1 to 8.

Images