CN-121808462-B - Intelligent diagnosis method for drainage pipeline

CN121808462BCN 121808462 BCN121808462 BCN 121808462BCN-121808462-B

Abstract

The invention relates to the technical field of municipal drainage pipeline data intelligent detection, and particularly discloses a drainage pipeline intelligent diagnosis method which comprises the steps of collecting multi-mode data of a target pipeline, generating a standardized multi-mode dataset, obtaining source field data, constructing a deep migration architecture, forming an initial diagnosis model, outputting a first diagnosis result, actively learning and adjusting the initial diagnosis model, generating a multi-mode collaborative diagnosis model, outputting a second diagnosis result, setting bidirectional calibration, optimizing the multi-mode collaborative diagnosis model, constructing a multi-phase flow simulation model, outputting a virtual sample, setting a reinforcement learning environment, reinforcing the multi-phase flow simulation model, outputting a final diagnosis result, combining geographic information, introducing expert diagnosis, re-learning the multi-phase flow simulation model to form a closed loop, and inputting the standardized multi-mode dataset into the multi-mode collaborative diagnosis model to realize drainage pipeline intelligent diagnosis. The invention deep migration learning multiple modes forms a full-link feedback mechanism, and has strong engineering practicability.

Inventors

ZHANG HAICHAO
PING YANG
ZHANG RAN
YANG MENG
WANG NIANNIAN
FENG RUIJIE
ZHANG QING
LI BIN
Di Danyang
LUO CHAO
SUN BAOCHENG
WANG HANTAO

Assignees

中国电建集团贵阳勘测设计研究院有限公司

Dates

Publication Date: 20260512
Application Date: 20260309

Claims (7)

1. An intelligent diagnosis method for a drainage pipeline is characterized by comprising the following steps: Collecting multi-modal data of a target pipeline, and performing dynamic preprocessing to generate a standardized multi-modal data set, wherein the multi-modal data comprises pipeline image data, overcurrent characteristic parameters, environmental parameters and structural state data; acquiring source field data, constructing a depth migration architecture, training to form an initial diagnosis model, and inputting the standardized multi-mode data set to the initial diagnosis model to output a first diagnosis result, wherein the diagnosis result comprises a disease type and a disease severity level; Based on the first diagnosis result, an initial diagnosis model is adjusted through active learning until the accuracy and consistency of the first diagnosis result reach preset thresholds, and a multi-mode collaborative diagnosis model is generated; Setting bidirectional calibration, judging whether a second diagnosis result output by a sample of a standardized multi-mode dataset violates a physical rule, if yes, marking the sample and the corresponding second diagnosis result as high-priority samples, feeding back to the multi-mode collaborative diagnosis model for re-learning, and optimizing the multi-mode collaborative diagnosis model; Inputting the standardized multi-modal data set of a target pipeline into the multi-phase flow simulation model, and outputting a virtual sample, wherein the virtual sample comprises disease types and disease severity levels of rare and extreme working conditions; Setting a reinforcement learning environment, optimizing the multiphase flow simulation model, outputting an optimized virtual sample, feeding back the optimized virtual sample to the multiphase flow collaborative diagnosis model and the multiphase flow simulation model, reinforcing the multiphase flow simulation model, updating the multiphase flow collaborative diagnosis model, and updating the second diagnosis result as a final diagnosis result; Marking the final diagnosis result by combining with geographic information, introducing expert diagnosis, and feeding back the expert diagnosis result to the multi-mode collaborative diagnosis model and the multiphase flow simulation model for re-learning and updating the multi-mode collaborative diagnosis model and the multiphase flow simulation model, wherein the expert diagnosis result comprises a misdiagnosis sample, a missed diagnosis sample and operation and maintenance effect data; Inputting the standardized multi-mode data set into the multi-mode collaborative diagnosis model to realize intelligent diagnosis of the drainage pipeline; The depth migration architecture is a three-level depth migration architecture and specifically comprises a single-mode encoder, a cross-mode fusion layer and a domain self-adaptive layer, wherein the single-mode encoder is used for extracting disease morphological characteristics and time sequence sensor characteristics in an image of source domain data; the active learning includes the following: calculating the sample comprehensive value of the standardized multi-modal data set based on the initial diagnosis model, wherein the calculation expression of the sample comprehensive value is as follows: wherein w is a modal weight, d is a characteristic difference coefficient; H (x) is a sample diagnosis uncertainty measure, U (x) is a sample comprehensive value; Screening high-value samples according to the comprehensive value of the samples, and manually marking to form an extended data set, wherein the high-value samples comprise rare pipeline disjoint samples, main pipe high-risk disease samples and multi-modal characteristic conflict samples; and reversely adjusting the initial diagnosis model and the sample comprehensive value according to the extended data set, and iterating until the accuracy and consistency of the first diagnosis result reach a preset threshold value.
2. The intelligent diagnosis method for the drainage pipeline according to claim 1, wherein the dynamic preprocessing comprises the following steps: Eliminating image noise and artifacts in the multi-modal data; screening core sensitive parameters in the multi-mode data through ash correlation entropy analysis, wherein the core sensitive parameters comprise parameters corresponding to flow speed or water depth of sedimentation and parameters corresponding to pH value or pipe wall thickness of corrosion; And feeding back disease characteristic contribution degree through the multi-mode collaborative diagnosis model, dynamically adjusting a screening threshold value of the core sensitive parameter and a cross-mode alignment strategy, and generating the standardized multi-mode data set.
3. The intelligent diagnosis method for the drainage pipeline according to claim 2, wherein the ash correlation entropy analysis screening correlation process comprises the following steps: And quantifying the association degree of the multi-modal data with the disease type and the disease severity level by adopting an ash association entropy analysis method, and calculating an ash entropy association degree R, wherein the expression is as follows: ; Wherein, the The gray entropy correlation degree of the ith data index and the target disease is represented, wherein t is a collection point; m is the total number of sample acquisition points; setting different gray entropy association thresholds for different disease types, and screening out the core sensitive parameters by comparing the gray entropy association R with the thresholds; the closer the gray entropy correlation degree R is to a threshold value, the higher the characterization correlation degree is, and the higher the corresponding multi-mode data sensitivity degree is.
4. The drain pipeline intelligent diagnostic method of claim 1, wherein the bidirectional calibration comprises the following: inputting the extended data set into the multi-mode collaborative diagnosis model for training; combining physical knowledge, namely constructing a mixed loss function, and taking the physical knowledge as a loss function constraint to dynamically adjust the weight of the physical constraint, wherein the physical knowledge comprises fluid dynamics, pipeline structural mechanics and solid-liquid coupling; adopting Bayes to optimize the tuning super parameters, avoiding overfitting through five-fold cross validation; Judging whether the second diagnosis result is violated with the calibration diagnosis result, if yes, improving the corresponding physical constraint weight, marking the second diagnosis result as a high-priority sample, feeding back to the multi-mode collaborative diagnosis model for learning again, and optimizing the multi-mode collaborative diagnosis model.
5. The intelligent diagnosis method of the drainage pipeline according to claim 4, wherein the dynamic adjustment of the weight of the physical constraint comprises the construction of a full-flow dynamic mechanism of preset-trigger-adjustment-calibration, and the full-flow dynamic mechanism comprises the following contents: based on different disease type characteristics, presetting a weight gradient as a reference for dynamic adjustment, and determining a triggering condition for dynamic adjustment; setting auxiliary triggering conditions according to the working condition change and the accuracy of the multi-mode collaborative diagnosis model, and capturing the adjustment requirement of the weight by combining the triggering conditions to pertinently adjust the weight; The method comprises the steps of completing bidirectional calibration after weight adjustment, marking a high-priority sample, feeding back to the multi-mode collaborative diagnosis model, learning again, optimizing the multi-mode collaborative diagnosis model, simulating disease response corresponding to the adjusted weight by the multi-phase flow simulation model, outputting an optimized virtual sample, carrying out virtual-real calibration, correcting adjustment amplitude according to a virtual-real calibration result, and iteratively optimizing a weight reference value to dynamically adjust the weight under different disease types and different working conditions.
6. The intelligent diagnosis method for the drainage pipeline according to claim 1 is characterized in that the optimized virtual sample output process comprises the steps of constructing a reinforcement learning environment, setting a multi-objective rewarding function, an adjusting range and an action space, performing iterative optimization on the multiphase flow simulation model by adopting a double-resolution depth Q network algorithm, and outputting the optimized virtual sample.
7. The intelligent diagnosis method of the drainage pipeline according to claim 6, wherein the multi-objective rewarding function comprises diagnosis accuracy, labeling cost, disease level evaluation consistency and corresponding weights, the adjustment range is multiphase flow simulation model parameter weights, and the action space is a sample screening strategy.

Description

Intelligent diagnosis method for drainage pipeline Technical Field The invention relates to the technical field of municipal drainage pipeline data intelligent detection, in particular to an intelligent drainage pipeline diagnosis method. Background The drainage pipeline is used as a core infrastructure of the urban water circulation system, and the health state of the drainage pipeline is directly related to urban flood control and drainage capacity, water environment quality and public safety. With the acceleration of the urban process, the old pipe network is increased year by year in proportion, and diseases such as siltation, cracks, corrosion, dislocation and the like are frequently generated, so that not only are drainage capacity reduced and flood disaster risks increased, but also secondary disasters such as groundwater pollution and pipeline collapse can be possibly caused. The current drainage pipeline diagnosis technology is primarily intelligent, but still faces four core bottlenecks, and the technology points are mostly simple linear combinations, lack of depth cooperative mechanisms, and are difficult to meet engineering actual demands, and particularly have obvious defects in accurate identification of disease types, quantitative evaluation of grades and pipe network scene suitability. The existing method depends on single type data or carries out surface fusion on multi-mode data, the related value of multi-source data and disease types and grades cannot be mined through dynamic feedback, so that disease evaluation precision is low and consistency is poor, model adaptation under different scenes depends on a large amount of annotation data, deep migration learning and active learning are lack of coordination, characteristic design of multi-pipe materials of a drainage pipe network, multiple working conditions and unbalanced samples (common disease samples are sufficient and rare disease samples are deficient) is not combined, the problem of cross-scene generalization and labeling cost control cannot be solved through bidirectional feedback at the same time, a part of deep learning model is of a 'black box' structure, field knowledge embedding and data driving training are mutually split, physical credibility of a prediction result cannot be guaranteed through dynamic calibration, the situation that disease grade evaluation is in contradiction with engineering rules easily occurs, a diagnosis result is acquired from the front end, an information island is formed through rear end optimization, a full-link feedback mechanism cannot be built through visualization, and engineering practicability is limited. In the prior art, part of researches provide a 'knowledge-data' collaborative driving thought, but the collaborative advantages of transfer learning and active learning are not combined, and the specific logic of quantitative evaluation of disease grades is lacking, and the research focuses on single-mode transfer learning, lacks a dynamic fusion mechanism of multi-mode data, is not suitable for core characteristics of multiple pipes and multiple working conditions of a drainage pipe network, and has weak cross-scene generalization capability. Therefore, it is needed to construct a technical scheme for intelligent diagnosis of a drainage pipeline, which has multiple-technology deep coupling, bidirectional feedback and dynamic regulation and control capability, can accurately realize disease type identification and grade evaluation, breaks through the limitation of the existing linear combination, systematically solves the technical bottleneck, and adapts to the scene characteristics of the drainage pipeline network. Disclosure of Invention In order to solve the technical problems that the prior art has single transfer learning mode and limited engineering practicability of a full-link feedback mechanism cannot be constructed, the invention provides an intelligent diagnosis method for a drainage pipeline, which comprises the following steps: Collecting multi-modal data of a target pipeline, and performing dynamic preprocessing to generate a standardized multi-modal data set, wherein the multi-modal data comprises pipeline image data, overcurrent characteristic parameters, environmental parameters and structural state data; acquiring source field data, constructing a depth migration architecture, training to form an initial diagnosis model, and inputting the standardized multi-mode data set to the initial diagnosis model to output a first diagnosis result, wherein the diagnosis result comprises a disease type and a disease severity level; Based on the first diagnosis result, an initial diagnosis model is adjusted through active learning until the accuracy and consistency of the first diagnosis result reach preset thresholds, and a multi-mode collaborative diagnosis model is generated; Setting bidirectional calibration, judging whether a second diagnosis result output by a sample of a standardiz