A sewer pipe intelligent diagnosis method

Abstract

The application relates to the technical field of municipal drainage pipeline data intelligent detection, and specifically discloses a drainage pipeline intelligent diagnosis method, which comprises the following steps: collecting multi-modal data of a target pipeline to generate a standardized multi-modal data set; acquiring source field data, constructing a deep migration architecture, forming an initial diagnosis model, and outputting a first diagnosis result; actively learning and adjusting the initial diagnosis model to generate a multi-modal collaborative diagnosis model and output a second diagnosis result; setting bidirectional calibration to optimize the multi-modal collaborative diagnosis model; constructing a multiphase flow simulation model to output a virtual sample; setting a reinforcement learning environment to reinforce the multiphase flow simulation model and output a final diagnosis result; combining geographic information, introducing expert diagnosis, and making the multiphase flow simulation model learn again to form a closed loop; and inputting the standardized multi-modal data set into the multi-modal collaborative diagnosis model to realize intelligent diagnosis of the drainage pipeline. The application forms a full-link feedback mechanism by deep migration learning of multiple modes, and has strong engineering practicability.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for municipal drainage pipelines, and specifically to an intelligent diagnostic method for drainage pipelines. Background Technology

[0002] As a core infrastructure of the urban water cycle system, the health of drainage pipelines directly affects urban flood control and drainage capacity, water environment quality, and public safety. With the acceleration of urbanization, the proportion of aging pipe networks is increasing year by year, leading to frequent occurrences of problems such as siltation, cracks, corrosion, and disconnections. This not only reduces drainage capacity and exacerbates the risk of flooding, but may also trigger secondary disasters such as groundwater pollution and pipeline collapses. While current drainage pipeline diagnostic technology has achieved initial intelligentization, it still faces four major bottlenecks. Furthermore, the various technical points are mostly simple linear combinations, lacking deep collaborative mechanisms, making it difficult to meet actual engineering needs. In particular, there are significant shortcomings in the accurate identification of disease types, quantitative assessment of disease severity, and adaptability to different pipe network scenarios.

[0003] Existing methods often rely on single-type data or only perform superficial fusion of multimodal data, failing to leverage dynamic feedback to uncover the correlation value between multi-source data and disease types and levels. This results in low accuracy and poor consistency in disease assessment. Model adaptation for different scenarios depends on a large amount of labeled data, with a lack of synergy between deep transfer learning and active learning. Furthermore, these methods are not designed to take into account the characteristics of drainage pipe networks, such as multiple pipe materials, multiple operating conditions, and imbalanced samples (sufficient samples of common diseases, but scarce samples of rare diseases). They cannot simultaneously solve the problems of cross-scenario generalization and labeling cost control through two-way feedback. Some deep learning models have a "black box" structure, with domain knowledge embedding and data-driven training being disconnected. They cannot guarantee the physical credibility of prediction results through dynamic calibration, and are prone to situations where disease level assessments contradict engineering principles. Diagnostic results form information silos with front-end data collection and back-end optimization. Visualization is only used as a means of presenting results, failing to build a full-link feedback mechanism and providing targeted operation and maintenance guidance based on disease type and level, thus limiting their engineering applicability.

[0004] In existing technologies, some studies have proposed a "knowledge-data" collaborative approach, but they have not combined the synergistic advantages of transfer learning and active learning, and lack specific logic for quantitative assessment of disease severity. Other studies focus on single-modal transfer learning, lack a dynamic fusion mechanism for multimodal data, and have not adapted to the core characteristics of drainage pipe networks with multiple pipe materials and multiple working conditions, resulting in weak cross-scenario generalization ability.

[0005] Therefore, there is an urgent need to develop a technical solution for intelligent diagnosis of drainage pipes that deeply integrates multiple technologies, has bidirectional feedback and dynamic control capabilities, and can accurately identify disease types and assess disease levels. This solution would break through the limitations of existing linear combinations, systematically solve technical bottlenecks, and adapt to the characteristics of drainage pipe networks. Summary of the Invention

[0006] To address the technical problems of existing technologies, such as limited engineering applicability due to the single mode of transfer learning and the failure to construct a full-link feedback mechanism, this invention provides an intelligent diagnostic method for drainage pipelines, comprising:

[0007] Multimodal data of the target pipeline is collected and dynamically preprocessed to generate a standardized multimodal dataset; the multimodal data includes pipeline image data, flow characteristic parameters, environmental parameters, and structural state data;

[0008] Acquire source domain data, construct a deep transfer architecture, train it to form an initial diagnostic model, input the standardized multimodal dataset into the initial diagnostic model, and output the first diagnostic result; the diagnostic result includes the disease type and the severity level of the disease; the source domain data includes standard pipe full-scale test data and existing drainage pipe disease datasets;

[0009] Based on the first diagnostic result, the initial diagnostic model is adjusted through active learning until the accuracy and consistency of the first diagnostic result both reach a preset threshold, generating a multimodal collaborative diagnostic model; a standardized multimodal dataset is input into the multimodal collaborative diagnostic model, and a second diagnostic result is output;

[0010] A two-way calibration is set up to determine whether the second diagnostic result of the sample output of the standardized multimodal dataset violates physical laws; if it does, the sample and the corresponding second diagnostic result are marked as high-priority samples and fed back to the multimodal collaborative diagnostic model for relearning and optimization.

[0011] A multiphase flow simulation model is constructed based on digital twin technology. The multiphase flow simulation model is used to obtain the types and severity levels of defects under rare and extreme operating conditions. The standardized multimodal dataset of the target pipeline is input into the multiphase flow simulation model, and virtual samples are output. The virtual samples include the types and severity levels of defects under rare and extreme operating conditions.

[0012] Based on the virtual samples, the multiphase flow simulation model is calibrated using real-world rare operating condition samples collected. A reinforcement learning environment is set up to optimize the multiphase flow simulation model, outputting optimized virtual samples, which are then fed back to the multimodal collaborative diagnostic model and the multiphase flow simulation model to strengthen the multiphase flow simulation model, update the multimodal collaborative diagnostic model, and update the second diagnostic result as the final diagnostic result. The real-world rare operating condition samples include rare and extreme operating condition samples.

[0013] The final diagnostic results are labeled with geographic information, and expert diagnosis is introduced. The expert diagnosis results are fed back to the multimodal collaborative diagnosis model and multiphase flow simulation model for further learning and updating. This guides the dynamic acquisition strategy of multimodal data for the target pipeline, forming a closed loop. The expert diagnosis results include misdiagnosed samples, missed diagnoses, and operation and maintenance effect data.

[0014] The standardized multimodal dataset is input into the multimodal collaborative diagnostic model to achieve intelligent diagnosis of drainage pipelines.

[0015] Furthermore, the dynamic preprocessing includes the following:

[0016] Eliminate image noise and artifacts in the multimodal data;

[0017] The core sensitive parameters in the multimodal data were screened by grey relational entropy analysis; the core sensitive parameters include the flow velocity / water depth parameter corresponding to siltation and the pH value / pipe wall thickness parameter corresponding to corrosion.

[0018] Cross-modal feature alignment is achieved through comparative learning; the contribution of disease features is fed back through the multimodal collaborative diagnostic model, and the screening threshold of the core sensitive parameters and the cross-modal alignment strategy are dynamically adjusted to generate the standardized multimodal dataset.

[0019] Furthermore, the grey relational entropy analysis process for filtering associations includes the following:

[0020] The grey relational entropy analysis method is used to quantify the correlation between the multimodal data and the disease type and severity level, and the grey entropy correlation degree R is calculated as follows:

[0021] ;

[0022] in, The gray entropy correlation between the i-th data indicator and the target disease is represented; t is the data collection point. is the grey relational coefficient; m is the total number of sample collection points;

[0023] Differentiated gray entropy correlation thresholds are set for different disease types, and the core sensitive parameters are screened by comparing the gray entropy correlation R with the threshold.

[0024] The closer the gray entropy correlation degree R is to the threshold, the higher the correlation degree and the higher the sensitivity of the corresponding multimodal data.

[0025] Furthermore, the deep migration architecture is a three-level deep migration architecture, specifically including: a single-modal encoder, a cross-modal fusion layer, and a domain adaptation layer;

[0026] The single-modal encoder is used to extract morphological features of diseases and temporal sensor features from the image of the source domain data.

[0027] The cross-modal fusion layer is used to dynamically allocate modal weights;

[0028] The domain adaptation layer is used to reduce the feature differences between the source domain data and the standardized multimodal dataset of the target pipeline.

[0029] Furthermore, the active learning includes the following:

[0030] The comprehensive value of the standardized multimodal dataset is calculated based on the initial diagnostic model.

[0031] Based on the comprehensive value of the samples, high-value samples are selected, manually labeled, and an expanded dataset is formed; the high-value samples include rare pipe disconnection samples, high-risk defects in main pipes, and multimodal feature conflict samples;

[0032] The initial diagnostic model and the comprehensive value of the samples are adjusted in reverse based on the expanded dataset, and iterated until the accuracy and consistency of the first diagnostic result both reach a preset threshold.

[0033] Furthermore, the formula for calculating the comprehensive value of the sample is as follows:

[0034] ;

[0035] Where w is the modal weight; d is the feature difference coefficient; x is the sample; H(x) is the sample diagnostic uncertainty measure; and U(x) is the sample comprehensive value.

[0036] Furthermore, the bidirectional calibration includes the following:

[0037] The expanded dataset is input into the multimodal collaborative diagnostic model for training;

[0038] By combining physical knowledge, a hybrid loss function is constructed, using physical knowledge as a constraint on the loss function, and the weight of the physical constraint is dynamically adjusted; the physical knowledge includes fluid dynamics, pipe structure mechanics, and solid-liquid coupling.

[0039] The hyperparameters are optimized using Bayesian optimization, and overfitting is avoided through five-fold cross-validation; the calibration and diagnostic results are then output.

[0040] Determine whether the second diagnostic result contradicts the calibration diagnostic result. If they do, increase the corresponding physical constraint weight, mark the second diagnostic result as a high-priority sample, and feed it back to the multimodal collaborative diagnostic model for further learning and optimization.

[0041] Furthermore, the dynamic adjustment of physical constraint weights includes a full-process dynamic mechanism of presetting, triggering, adjusting, and calibrating; the full-process dynamic mechanism includes the following:

[0042] Based on the characteristics of different disease types, a preset weight gradient is used as the benchmark for dynamic adjustment, and the triggering conditions for dynamic adjustment are determined.

[0043] Based on changes in operating conditions and the accuracy of the multimodal collaborative diagnostic model, auxiliary triggering conditions are set. In conjunction with these triggering conditions, the adjustment requirements of the weights are captured, and the weights are adjusted accordingly.

[0044] After the weights are adjusted, bidirectional calibration is completed, high-priority samples are marked, and the data is fed back to the multimodal collaborative diagnostic model for further learning and optimization. The multiphase flow simulation model simulates the disease response corresponding to the adjusted weights, outputs optimized virtual samples, and performs virtual-real calibration. The adjustment range is corrected based on the virtual-real calibration results, and the weight benchmark value is iteratively optimized to achieve dynamic adjustment of the weights under different disease types and different operating conditions.

[0045] Furthermore, the optimized virtual sample output process includes: constructing a reinforcement learning environment, setting a multi-objective reward function, adjusting the range and action space, using a dual-duel deep Q-network algorithm to iteratively optimize the multiphase flow simulation model, and outputting the optimized virtual samples.

[0046] Furthermore, the multi-objective reward function includes diagnostic accuracy, annotation cost, consistency of disease level assessment, and corresponding weights; the adjustment range is the parameter weights of the multiphase flow simulation model; and the action space is the sample selection strategy.

[0047] The beneficial effects of this invention are:

[0048] (1) Deep coupling of multiple modules improves collaborative efficiency: Breaking the limitations of linear serial connection, a closed loop of the whole link is constructed, and each technical point is dynamically adapted through bidirectional feedback. Compared with the superposition of single technologies, the diagnostic accuracy is improved and the rate of missed diagnosis and misdiagnosis is significantly reduced.

[0049] (2) Dynamic feedback reduces implementation costs: Active learning and deep transfer work together in a two-way collaboration, combined with digital twin virtual samples to supplement the target domain, reducing the amount of labeling required. At the same time, data utilization is improved through data preprocessing and model adaptation feedback.

[0050] (3) Dual guarantee of physical credibility and generalization ability: knowledge embedding and dynamic calibration of model training avoid “black box” prediction bias. Digital twin and reinforcement learning are coupled in the real world, enabling the model to adapt to different pipe materials, climate zones and complex working conditions, and improving long-term operational stability.

[0051] (4) Strong engineering practicality: The combination of visualization and full-link feedback enables the reverse drive of diagnostic results to operation and maintenance, data collection and optimization, adapts to existing operation and maintenance systems, and greatly improves the efficiency and intelligence level of pipeline operation and maintenance. Attached Figure Description

[0052] Figure 1 This is a logic block diagram of the intelligent diagnostic method for drainage pipes provided by the present invention. Detailed Implementation

[0053] The technical solution of the present invention is further described below, but the scope of protection is not limited to what is described.

[0054] This invention provides an intelligent diagnostic method for drainage pipes, such as... Figure 1 As shown, it includes:

[0055] Step S100: Collect multimodal data of the target pipeline and perform dynamic preprocessing to generate a standardized multimodal dataset; the multimodal data includes pipeline image data, flow characteristic parameters, environmental parameters, and structural state data;

[0056] The multimodal data collection of the target pipeline includes: deploying a hierarchical multimodal monitoring network in the target pipeline area; simultaneously collecting pipeline image data (visual features of defects, pipe wall texture), flow characteristic parameters (flow velocity, shear stress, water depth, flow rate), environmental parameters (groundwater level, soil moisture, medium pH value), and structural status data (pipe wall thickness, interface displacement, pipeline deformation) through unmanned pipeline measurement devices, radar detection vehicles, video monitoring robots, and IoT sensor arrays; adaptively capturing visual features of defects such as siltation and cracks using visual data acquisition equipment (UPM, radar detection vehicle); and simultaneously collecting flow parameters, environmental parameters, and structural parameters using sensor arrays, combined with on-site surveys and archive records of basic data such as pipe type, pipeline burial depth, and service life. All devices transmit data synchronously to the comprehensive management database via IoT and can receive feedback instructions from the visualization subsystem, dynamically adjusting the sampling frequency and location for high-level defect areas and specific defect types (e.g., increasing the pH value sensor collection frequency in high-corrosion areas and increasing flow velocity monitoring in siltation areas).

[0057] The dynamic preprocessing process forms a two-way feedback loop with the training of the multimodal collaborative diagnostic model. An enhanced denoising autoencoder is used to eliminate image noise and artifacts. Gray relational entropy analysis is used to screen core sensitive parameters (such as flow velocity / water depth parameters for siltation and pH value / pipe wall thickness parameters for corrosion). Contrastive learning is used to achieve cross-modal feature alignment. Based on the feature contribution feedback from the multimodal collaborative diagnostic model output, the sensitive parameter screening threshold and cross-modal alignment strategy are dynamically adjusted. Finally, a standardized multimodal dataset that adapts to model requirements and correlates with disease features is generated. The dynamic preprocessing specifically includes the following:

[0058] Eliminate image noise and artifacts in the multimodal data;

[0059] The core sensitive parameters in the multimodal data were screened by grey relational entropy analysis; the core sensitive parameters include the flow velocity / water depth parameter corresponding to siltation and the pH value / pipe wall thickness parameter corresponding to corrosion.

[0060] Cross-modal feature alignment is achieved through comparative learning; the contribution of disease features is fed back through the multimodal collaborative diagnostic model, and the screening threshold of the core sensitive parameters and the cross-modal alignment strategy are dynamically adjusted to generate the standardized multimodal dataset.

[0061] The grey relational entropy analysis and correlation screening process includes the following:

[0062] The grey relational entropy analysis method is used to quantify the correlation between the multimodal data and the disease type and severity level, and the grey entropy correlation degree R is calculated as follows:

[0063] (1)

[0064] in, The gray entropy correlation between the i-th data indicator and the target disease is represented; t is the data collection point. is the grey relational coefficient; m is the total number of sample collection points;

[0065] The grey relational coefficient The calculation expression is:

[0066] (2)

[0067] in, The gray relational coefficient between the i-th data indicator (e.g., pH value) and the target disease (e.g., corrosion) is closer to 1, indicating a closer relationship between the corresponding data indicator and the target disease. This represents the minimum difference between all data indicators and the disease. This represents the maximum difference between all data indicators and the disease. ρ represents the actual difference between the i-th indicator and the disease at the t-th collection point; ρ is the adjustment coefficient used to avoid extreme values ​​from affecting the results and to make the calculation more stable; m is the total number of sample collection points.

[0068] Differentiated gray entropy correlation thresholds are set for different disease types, and the core sensitive parameters are screened by comparing the gray entropy correlation R with the threshold.

[0069] The closer the gray entropy correlation degree R is to the threshold, the higher the correlation degree and the higher the sensitivity of the corresponding multimodal data.

[0070] The grey relational entropy analysis method is used to quantify the correlation between data indicators and diseases. Screening criteria are set according to the needs of different diseases. Finally, core data is selected and useless data is eliminated to provide a high-quality data foundation for the intelligent diagnosis of the entire pipeline.

[0071] The feedback mechanism for the feature contribution of grey relational entropy analysis and multimodal collaborative diagnostic model is as follows: the multimodal collaborative diagnostic model is trained to output the feature importance scores of each parameter for different disease types and levels. If the feature importance of a certain parameter is continuously lower than the threshold, its weight in grey relational entropy analysis is automatically reduced, and parameters with high feature contribution (such as flow velocity and water depth parameters for siltation diseases) are retained first, so as to achieve dynamic adaptation between preprocessing and model requirements.

[0072] Step S200: Obtain source domain data, construct a deep transfer architecture, train it to form an initial diagnostic model, input the standardized multimodal dataset into the initial diagnostic model, and output the first diagnostic result; the diagnostic result includes the disease type and the severity level of the disease; the source domain data includes standard pipe full-scale test data and existing drainage pipe disease datasets; the disease types include siltation, cracks, corrosion, disconnection, root intrusion, and scaling; the severity levels of the disease include slight, moderate, severe, and extremely severe, corresponding to the "Technical Specification for Inspection and Evaluation of Urban Drainage Pipelines" CJJ181-2012 standard), and the level determination combines the characteristic quantification indicators in the CJJ181-2012 standard (e.g., crack width ≥ 5mm is severe, siltation thickness ≥ 1 / 3 of pipe diameter is severe).

[0073] The deep migration architecture is a three-level deep migration architecture, specifically including: a single-modal encoder, a cross-modal fusion layer, and a domain adaptation layer; it adapts to the characteristics of multiple pipe materials and multiple working conditions in drainage pipe networks;

[0074] The single-modal encoder is used to extract morphological features of defects and temporal sensor features from the image of the source domain data. The single-modal encoder may be, for example, ResNet-50 or Bi-LSTM. ResNet-50 extracts morphological and texture features of defects from the image, such as linear features of cracks, mottled texture of corrosion, and accumulation patterns of silt. Bi-LSTM extracts temporal sensor features, such as the continuous decrease in flow velocity caused by siltation and stress fluctuations caused by crack propagation. The fully connected network processes structural data, such as the reference values ​​of pipe wall thickness for different pipe materials and the interface displacement threshold.

[0075] The cross-modal fusion layer is used to dynamically allocate modal weights. This layer uses a multi-head attention mechanism to dynamically allocate modal weights (e.g., prioritizing structural data weights for structural defects and prioritizing flow parameter weights for hydraulic defects). This addresses the issue of differing contributions from different modes. The layer's input includes image features, time-series sensor features, and structural features. First, a linear transformation maps these three features to the same dimension. Then, an attention head is constructed, and each attention head calculates the attention weights for different modal features. The weight allocation aligns with the characteristics of pipeline defects, ensuring that multi-modal features collaboratively support defect diagnosis. The expression is:

[0076] (3)

[0077] Here, exp() is an exponential function that makes the data with high matching degree more prominent; score() is the matching score, which calculates the matching degree between the i-th type of data and the diagnostic needs of the j-th type of disease. The feature vector of the i-th type of multimodal data is extracted by a single-modal encoder; h represents the feature mean vector for the diagnostic needs of the j-th type of disease, obtained by statistically analyzing and averaging the typical features of diseases in the source domain disease dataset; u Let h be the feature vector of the u-th class of multimodal data. u correspond ; Let be the feature mean vector representing the diagnostic requirements for the v-th type of disease. correspond The core of the attention weighting mechanism is to give greater weight to a certain type of data as it closely matches the current needs of disease diagnosis, thus giving it more influence in the diagnosis process.

[0078] The domain adaptation layer is used to reduce the feature differences between the source domain data and the standardized multimodal dataset of the target pipeline. The domain adaptation layer introduces a generative adversarial network (GAN) to reduce the feature differences between the source domain data and the standardized multimodal dataset of the target pipeline.

[0079] The training strategy for the initial diagnostic model is "bottom-level freezing + top-level fine-tuning". The bottom-level parameters of ResNet-50 are frozen to retain general visual features, and the top-level and fusion / adaptive layer parameters are fine-tuned according to the characteristics of the drainage pipe network (such as adapting to the texture features of different pipe materials and the morphological features of different diseases). The Adam optimizer is used during the pre-training process, and the annotation results (disease type and grade label) of the actively learned and screened samples are used as the verification basis. Pre-training is stopped when the loss of disease type recognition and the loss of grade evaluation of the labeled samples do not decrease continuously.

[0080] The domain adaptation layer of the GAN is trained adversarially between the generator and the discriminator to minimize the maximum mean difference (MMD) of features between the source domain data and the standardized multimodal dataset of the target pipeline. It focuses on the differences between the source domain data and the standardized multimodal dataset of the target pipeline (such as the difference in operating conditions between the full-scale test in the source domain and the actual pipeline in the target domain, and the difference in structural features of different pipe materials). The adversarial training results are simultaneously fed back to active learning, so that the weight of the domain difference coefficient is dynamically adjusted according to the domain adaptation effect—the MMD weight is increased when the domain adaptation effect is poor, and samples with strong domain specificity are selected first.

[0081] Step S300: Based on the first diagnostic result, the initial diagnostic model is adjusted through active learning until the accuracy and consistency of the first diagnostic result both reach a preset threshold, generating a multimodal collaborative diagnostic model; a standardized multimodal dataset is input into the multimodal collaborative diagnostic model, and a second diagnostic result is output;

[0082] The active learning is based on the diagnostic uncertainty of the initial diagnostic model, cross-modal complementarity, and pipeline network specificity (pipe type, pipeline network importance, and rarity of defects); it also calculates the comprehensive value of samples using weighted averages; the active learning specifically includes the following:

[0083] The comprehensive value of the standardized multimodal dataset is calculated based on the initial diagnostic model.

[0084] Based on the comprehensive value of the samples, high-value samples are selected, manually labeled, and an expanded dataset is formed; the high-value samples include rare pipe disconnection samples, high-risk defects in main pipes, and multimodal feature conflict samples;

[0085] The initial diagnostic model and the comprehensive value of the samples are adjusted in reverse based on the expanded dataset, and iterated until the accuracy and consistency of the first diagnostic result both reach a preset threshold.

[0086] The formula for calculating the comprehensive value of the sample is as follows:

[0087] (4)

[0088] Where w is the modal weight; d is the feature difference coefficient; x is the sample; H(x) is the sample diagnostic uncertainty measure; and U(x) is the sample comprehensive value.

[0089] The formula for calculating the sample diagnostic uncertainty measure H(x) is as follows:

[0090] (5)

[0091] in, The disease type information entropy is used to quantify the uncertainty of the model's determination of the disease type of sample x (the disease types include 6 categories: siltation, cracks, corrosion, detachment, and root intrusion scaling). The information entropy of the severity level of the disease is used to quantify the uncertainty of the model's judgment on the severity level of the disease for sample x (the severity level of the disease is divided into four levels: mild, moderate, severe, and extremely severe).

[0092] The The calculation formula is:

[0093] (6)

[0094] The The calculation formula is:

[0095] (7)

[0096] in, and After fusing multimodal features into the initial diagnostic model for posterior probability, the output sample x is determined as the posterior probability of the k-th disease type and the n-th disease severity.

[0097] Step S400: Set bidirectional calibration to determine whether the second diagnostic result of the sample output of the standardized multimodal dataset violates physical laws; if it does, mark the sample and the corresponding second diagnostic result as high-priority samples and feed them back to the multimodal collaborative diagnostic model for relearning and optimization of the multimodal collaborative diagnostic model.

[0098] The bidirectional calibration includes the following:

[0099] The expanded dataset is input into the multimodal collaborative diagnostic model for training;

[0100] By combining physical knowledge, a hybrid loss function is constructed, using physical knowledge as the constraint of the loss function, and the weight of the physical constraint is dynamically adjusted; the physical knowledge includes fluid dynamics (adapting to hydraulic diseases such as siltation and foreign object blockage), pipeline structural mechanics (adapting to structural diseases such as cracks, corrosion, and deformation), and solid-liquid coupling.

[0101] The hyperparameters are optimized using Bayesian optimization, and overfitting is avoided through five-fold cross-validation; the calibration and diagnostic results are then output.

[0102] Determine whether the second diagnostic result contradicts the calibration diagnostic result (e.g., the predicted siltation level contradicts the flow velocity change pattern). If they do, increase the corresponding physical constraint weight, mark the second diagnostic result as a high-priority sample, and feed it back to the multimodal collaborative diagnostic model for relearning and optimization.

[0103] The expression for the hybrid loss function is:

[0104] (8)

[0105] Where λ represents the dynamic weighting coefficient used to balance the contribution of data fitting and physical constraints, and is dynamically adjusted according to the type of disease, prediction deviation, and changes in working conditions; L is the hybrid loss function; L1 is the dual-task classification loss, L 2i This is for differentiated physical constraint loss.

[0106] The dual-task classification loss L1 is a weighted sum of the disease type classification loss and the disease level regression loss. The weights are dynamically adapted according to the engineering maintenance priority, and the weight coefficients are dynamically adapted according to the task priority. The disease type classification weight is 0.4-0.5, and the disease level regression weight is 0.5-0.6, prioritizing the accuracy of level assessment and meeting the core requirement of level quantification in drainage pipe network disease maintenance decisions. The two sub-losses are adapted to the task attributes of discrete multi-classification and ordered multi-classification, respectively. The core calculation formula is as follows:

[0107] (9)

[0108] in, The types of diseases are classified into six categories: siltation, cracks, corrosion, detachment, root invasion, and scaling. The standard for regressing losses based on the severity level of the disease is divided into four levels: minor, moderate, severe, and extremely severe.

[0109] The types of diseases classified as losses For discrete, unordered multi-class classification problems, multi-class cross-entropy loss is chosen as the core calculation method, which is the classification loss of the disease type. The calculation formula is:

[0110] (10)

[0111] Where N is the total number of multimodal dataset samples used for model training;

[0112] c represents the disease type category index, corresponding to 6 disease types;

[0113] Label the disease type of the Mth sample with a true label, using one-hot encoding (if the Mth sample is disease type c, ...). (and the rest are 0).

[0114] Let be the posterior probability of the c-th disease type output by the model for the M-th sample.

[0115] The regression loss of the severity level of the disease For ordered multi-class classification problems, ordered cross-entropy loss (ordinal regression loss) is chosen as the core calculation method, and the disease severity level regression loss is... The calculation formula is:

[0116] (11)

[0117] Where N is the total number of standardized multimodal dataset samples participating in model training;

[0118] f is the binary classification index for ordered levels, f=1-3 (corresponding to 3 classification thresholds for 4 levels).

[0119] Let f be the ordered true label of the Mth sample, which is a binary value of 0 / 1 (indicating whether the sample has reached the level defined by the fth threshold).

[0120] The model outputs a rank score for the Mth sample;

[0121] S() is the Sigmoid activation function, representing the cumulative probability that the Mth sample reaches the fth level classification threshold.

[0122] The differentiated physical constraint loss L 2i To classify constraints into three categories based on their inherent properties (hydraulic characteristics / structural characteristics), each category corresponds to a specific mathematical expression, with clear parameter definitions and physical meanings:

[0123] For hydraulic defects such as siltation, scaling, and foreign object blockage, the constraint design focuses on the hydraulic correlation between "flow velocity-flow capacity-defect level". The core constraints are constructed based on the Manning formula, and the influence of key parameters is highlighted through weight coefficients. For scaling defects, additional flow section constraints are added, and they are set according to the characteristics of pipe materials. Because HDPE pipes and concrete pipes have different material stiffness and flow capacity tolerance, HDPE pipes are more flexible but more sensitive to flow sections, so the allowable shrinkage rate is lower. Concrete pipes have higher stiffness and can be appropriately relaxed. The setting of these weights and constraint thresholds has been verified by full-scale tests and is adapted to the actual pipe network characteristics.

[0124] ① Hydraulic defects constraint items: For siltation, scaling, and foreign object blockage, the constraint formula is as follows:

[0125] (12)

[0126] in, This represents the physical constraint loss value for hydraulic defects. The magnitude of the value reflects the degree of deviation between the model's prediction of hydraulic defects and the laws of fluid dynamics. The smaller the value, the smaller the deviation, and the higher the physical reliability of the model's prediction. The predicted flow velocity under the target pipeline hydraulic failure scenario is output by a multimodal collaborative diagnostic model combined with a standardized multimodal dataset (flow characteristic parameters, pipeline structure data, etc.); The theoretical value of the flow velocity is derived by combining the Manning formula with pipe diameter, pipe slope and predicted sediment thickness. It is a theoretical reference value that conforms to the laws of fluid dynamics and is the core basis for verifying whether the prediction conforms to the laws of physics. The sedimentation thickness is inferred from the measured water depth and flow velocity; α is the specific weight of sedimentation thickness, highlighting the core role of thickness in grade determination. The thickness of sediment is inferred from the predicted water depth and flow velocity.

[0127] The core hazard of scaling is the reduction of the pipe's cross-sectional area, leading to decreased drainage capacity. An independent constraint on the "cross-sectional reduction rate" is added to avoid the bias of judging the scaling level solely based on flow velocity (e.g., some scaling may not significantly affect flow velocity, but the cross-sectional reduction is already close to the pipe material's tolerance limit). Scaling damage is further constrained by the additional cross-sectional area constraint. ;in, The reduction in cross-sectional area due to scaling. The design, differentiated by pipe material, essentially addresses the dual-dimensional harm of scaling (flow velocity impact + cross-sectional reduction) by adding a specific constraint dimension, making the physical constraints more comprehensive. The scaling-related disease constraint formula is as follows:

[0128] (13)

[0129] in, Loss due to physical constraint caused by scaling;

[0130] For structural defects such as cracks, corrosion, deformation, and delamination, the constraint terms adopt a multi-parameter coupled design, and the weight coefficient rules are set around the "priority of structural stability": the weight of crack width is higher than that of stress, because crack width is the most easily quantifiable indicator in on-site operation and maintenance and is directly related to the degree of structural damage; corrosion defects are subject to dual constraints of thickness reduction and corrosion rate, and the corrosion scenario of metal pipes is further enhanced, the core reason being that the electrochemical corrosion rate of metal pipes is fast and easily leads to perforation and collapse, and the degree of harm is higher than that of concrete pipes, so the constraint strength needs to be strengthened; the weights of deformation and delamination defects are between those of cracks and corrosion, which are adapted to their moderate impact on structural stability, forming a structural defect constraint weight gradient of "cracks > corrosion (metal pipes) > deformation / delamination".

[0131] ② Structural defects constraints: For cracks, corrosion, deformation, and delamination, multi-parameter coupled constraints are used. The crack constraint formula is:

[0132] (14)

[0133] in, For pipe-specific grade thresholds (1mm / 3mm / 5mm / 8mm corresponding to minor / moderate / severe / extremely severe for concrete pipes); β is the width weight; γ is the stress weight.

[0134] The predicted value of the crack width in the target pipeline is output by a multimodal collaborative diagnostic model that integrates multimodal features such as pipeline image data (crack visual features) and structural state data.

[0135] The predicted value of pipe wall stress in the cracked area of ​​the target pipeline is obtained by combining the model with elasticity mechanics, structural strength verification formula, and integrating pipeline structural state data (pipe wall thickness, interface displacement) and environmental parameters (groundwater level, soil pressure), reflecting the actual structural stress state of the cracked area.

[0136] The allowable stress value of the pipe wall corresponding to the target pipeline material is determined by the material properties of the pipe material (such as the compressive strength of concrete and the tensile strength of steel) and the pipeline engineering design standards.

[0137] The corrosion constraint formula is:

[0138] (15)

[0139] in, The physical constraint loss value of corrosion disease directly reflects the deviation between the model's predicted value of the core quantitative indicators of corrosion disease and the pipe-specific allowable / standard value. The smaller the value, the smaller the deviation, and the higher the physical reliability of the model's prediction.

[0140] The allowable reduction in pipe wall thickness corresponding to the target pipe material is determined by the characteristics of the pipe material and engineering design standards (≤20% for steel pipes and ≤30% for concrete pipes). It is the core threshold for ensuring the safety of the pipeline structure and judging the severity of corrosion.

[0141] The predicted value of pipe wall thickness reduction in the target pipeline corrosion area is output by a multimodal collaborative diagnostic model that integrates multimodal data such as pipeline structural status data, environmental parameters (medium pH value, groundwater level), and pipe material characteristics.

[0142] The predicted corrosion rate (unit: mm / year) for the target pipeline corrosion zone is obtained by a multimodal collaborative diagnostic model that integrates data such as pipe material characteristics, environmental parameters, and the trend of pipe wall thickness reduction.

[0143] The allowable corrosion rate for the target pipeline material is determined by the pipe material anti-corrosion design standard and the municipal drainage pipeline engineering operation and maintenance specification.

[0144] For mixed diseases, the weighting coefficient rule adopts a coupling logic of "prioritizing the dominant disease and supplementing the secondary disease" to ensure that the constraint strength of the dominant disease is superior. For example, in the scenario of corrosion as the dominant disease and crack as the secondary disease, the model is calibrated first by reducing the thickness and rate of corrosion, and then supplemented by crack width constraint to avoid mutual interference between the constraints of the two types of diseases. At the same time, the overall mixed disease takes the maximum value of the two single diseases to strengthen the constraint of the mixed scenario and adapt to its characteristics of "multiple diseases superimposed and damage amplified".

[0145] ③ Mixed disease constraints: For scenarios such as corrosion + cracking, and siltation + deformation, a dominant-secondary superposition mode is adopted:

[0146] (16)

[0147] in, θ represents the physical constraint loss for mixed diseases; θ is the coupling weight, which prioritizes strengthening the constraints of the dominant disease. It is the dominant term in the physical constraint loss of disease; This is a secondary term in the physical constraint loss of the disease.

[0148] The aforementioned constraints and weight coefficients are not fixed values, but rather dynamically adjusted through a full-process dynamic mechanism of "preset-trigger-adjustment-calibration." The core is to ensure that the weight coefficients accurately adapt to the characteristics of different disease types while responding to changes in operating conditions and fluctuations in model performance. The specific details of this full-process dynamic mechanism are as follows:

[0149] First, a weight gradient is preset based on the characteristics of the defects to reflect the differences in maintenance priorities: structural defects have a higher weight than hydraulic defects, primarily because structural damage (such as cracks and corrosion) is more likely to cause pipeline collapse, posing a far greater threat than hydraulic stagnation (such as siltation); metal pipe corrosion has an additional weight because of its rapid electrochemical corrosion rate and high risk of perforation, requiring stronger constraints; crack width has a higher weight than stress because width is the most intuitive and quantifiable core indicator for on-site maintenance, aligning with practical engineering needs. This gradient preset is not static but serves as a benchmark for dynamic adjustment, avoiding insufficient adaptability caused by indiscriminate initial weights.

[0150] Secondly, by accurately capturing adjustment needs through multi-dimensional triggering conditions, blind adjustments are avoided: For hydraulic defects, the focus is on monitoring the prediction deviations of flow velocity and sediment thickness. When the patterns of these two changes contradict each other (e.g., sediment thickens but flow velocity does not decrease), hydraulic constraints are strengthened. For structural defects, quantitative indicators (crack width, pipe wall thickness) and mechanical parameters (stress, ring stiffness) are monitored simultaneously. Any deviation of any parameter from the pattern triggers an adjustment, with additional adjustment sensitivity added for metal pipe scenarios. For mixed defects, the proportion of dominant defects is identified through multi-modal features, and dynamic floating coupling weights ensure that the constraints of dominant defects are dominant. At the same time, changes in operating conditions (pipe material, water level) and fluctuations in model accuracy serve as auxiliary triggers, allowing adjustments to cover all scenario variables.

[0151] Finally, adjustments and calibrations are achieved through end-to-end feedback to ensure compatibility: After weight adjustments, the active learning module filters out samples that are still misdiagnosed after the adjustments and adds annotations to correct the weight thresholds in reverse. For example, if scaling defects are still misdiagnosed after adjustments, the allowable shrinkage rate of HDPE pipes is further reduced. The digital twin module simulates the defect response corresponding to the adjusted weights and verifies the effectiveness of the constraints by comparing the deviation between virtual and real samples, thus correcting the adjustment range. Operation and maintenance effect data (such as flow rate recovery after dredging and stress changes after crack repair) iteratively optimizes the weight benchmark value over a long period, extending the adjustment mechanism from "data-driven" to "engineering verification-driven," ultimately achieving accurate adaptation for different defect types and working conditions, while balancing the physical reliability of the model and diagnostic accuracy.

[0152] Structural mechanics formulas (elasticity equations, structural strength verification formulas) are specifically adapted to structural defects such as cracks, corrosion, deformation, and delamination, constructing a "quantitative index-structural stability" constraint logic. For example, when the crack width of a concrete pipe is ≥3mm, combined with the pipe wall stress calculation results, if the stress reaches more than 80% of the design value, the level is directly judged as severe. When the pipe wall thickness reduction is ≥20% (steel pipe) or ≥30% (concrete pipe), the corrosion rate parameter is simultaneously linked. If the corrosion rate exceeds 0.1mm / year, the level is upgraded by one level, ensuring that the corrosion level assessment takes into account both current damage and development trends. Deformation and delamination defects are respectively linked to the ring stiffness design value and the allowable value of interface displacement, forming targeted constraints. Solid-liquid coupling knowledge (sediment initiation condition formula) assists in calibrating the sedimentation level. By matching the water flow shear stress with the sediment particle size, the rationality of the sedimentation level prediction is verified.

[0153] All three types of knowledge are transformed into corresponding physical constraint loss terms and integrated into the hybrid loss function. The constraint strength is dynamically adjusted according to the type of disease: the physical constraint weight is higher for structural diseases than for hydraulic diseases, and for mixed diseases, the constraint of the dominant disease is the core, supplemented by the constraint of the secondary disease. The influence of the two is balanced by coupling weights. A base value is set according to the type of disease, and then adaptively adjusted according to the degree of deviation between the model prediction and the physical law. If the prediction result violates the core law, the constraint is automatically strengthened to ensure that the diagnosis result is both consistent with the data characteristics and conforms to the actual law of engineering.

[0154] Model training and optimization are based on the expanded target domain dataset, which is divided into training, validation, and test sets in a 7:2:1 ratio. Data augmentation is used to expand the training set (e.g., perturbation of time-series parameters to simulate disease characteristics under different shooting angles and working conditions). Hyperparameters such as learning rate, batch size, attention weight, and loss function weight are tuned through Bayesian optimization, with the optimization objective being to maximize the consistency between the accuracy of disease type identification and the grade assessment on the validation set. The Adam optimizer and early stopping strategy are used to avoid overfitting, and 5-fold cross-validation is used to evaluate model stability, ensuring that the model can stably output diagnostic results in different pipeline sections.

[0155] A dynamic calibration mechanism is implemented throughout the entire training process, supporting iterative improvements in the accuracy of disease diagnosis. During training, the consistency between the model's prediction results and physical laws and industry standards is verified in real time. Samples that violate engineering physical laws such as fluid dynamics and pipeline structural mechanics, or that deviate from the disease level judgment standard specified in the "Technical Specification for Inspection and Evaluation of Urban Drainage Pipelines" CJJ181-2012 by more than one level, are marked as high-priority samples and fed back to the active learning module for supplementary annotation. At the same time, the effectiveness of knowledge constraints is verified through model diagnostic accuracy. If a certain type of knowledge constraint leads to a decrease in accuracy (such as the structural constraints of a specific pipe material not being suitable), the constraint form or weight is automatically adjusted. For mixed diseases (such as corrosion accompanied by cracks, and siltation causing deformation), the dominant disease type and comprehensive level are identified through multimodal feature cross-validation and knowledge constraint collaboration, achieving a synergistic unity of data fitting and physical reliability, and finally outputting accurate disease type and level results.

[0156] The coordinated adjustment mechanism of the dynamic weight coefficient λ and parameter weights (α, β, γ, ζ, θ) is based on "the essential characteristics of the disease, the prediction deviation as the trigger, and the full-link feedback as the calibration," to achieve dynamic adaptation of the weight coefficients to different disease types, changes in operating conditions, and model performance. The specific rules are as follows:

[0157] 1) Preset basic weight gradient: Based on the severity of the disease, its quantifiability, and the priority of operation and maintenance, preset weight benchmark values ​​are set to form a basic gradient of "mixed diseases > structural diseases > hydraulic diseases". Among them, the weight of metal pipe corrosion is higher than that of concrete pipe (metal pipe corrosion λ is additionally +0.1), and the weight of cracks is higher than that of deformation / disconnection. Among hydraulic diseases, the weight of siltation thickness is higher than that of flow velocity, and the scale constraint of HDPE pipe is stricter than that of concrete pipe, to ensure that the basic weights are in line with the core needs of pipeline operation and maintenance.

[0158] 2) Dynamically Adjust Triggering Conditions: Three types of triggering scenarios are used to adapt to different disease adaptation needs: Prediction Deviation Trigger: When the prediction result of the multimodal collaborative diagnosis model deviates from the physical law by more than 10% (such as increased siltation but increased flow velocity, or crack width exceeding the standard but stress not reaching the critical level), or deviates from the industry standard level by more than 1 level (such as misjudging a severe crack as a medium one), the weight is immediately increased; Working Condition Change Trigger: When a change in pipe material is detected (such as switching from concrete pipe to steel pipe) or a sudden change in environment (such as a sudden rise in groundwater level causing underwater corrosion), the corresponding disease weight is automatically adjusted (such as strengthening the corrosion constraint weight in the steel pipe scenario); Model Feedback Trigger: When the diagnostic accuracy of a certain type of disease decreases for 3 consecutive rounds, it is determined to be too strong / too weak constraint, and the weight is fine-tuned.

[0159] 3) Differentiated adjustment strategy: Design exclusive adjustment range and direction for different disease types to avoid insufficient adaptability caused by uniform adjustment. When deviation is triggered, λ is increased and parameter weight is finely adjusted. Mixed diseases: Adapt the adjustment rules according to the dominant disease type, and θ dynamically fluctuates in the range of 0.6-0.7 with the proportion of the dominant disease.

[0160] 4) End-to-end feedback calibration: After weight adjustment, a closed-loop calibration is formed through active learning and digital twins. Samples that are still misdiagnosed after weight adjustment are marked as high priority, and the weight threshold is corrected in reverse after supplementing the label. The disease response corresponding to the adjusted weight is simulated by digital twins to verify the effectiveness of the constraints. If the deviation between the virtual sample and the real sample exceeds the threshold, the weight coefficient is further fine-tuned. At the same time, the weight benchmark value is iteratively optimized by combining the feedback of operation and maintenance effect (such as the thickness change after corrosion treatment) to ensure that the adjustment mechanism fits the actual working conditions.

[0161] The weight coefficient adjustment is subject to boundary constraints: the maximum value of λ shall not exceed 0.8 and the minimum value shall not be lower than 0.3. The single adjustment range of the parameter weight shall not exceed ±0.2 to avoid the oscillation of the multimodal collaborative diagnostic model caused by sudden weight changes, and to ensure the balance between the stability of the adjusted model and the diagnostic accuracy. All adjustment rules comply with the "Technical Specification for Testing and Evaluation of Urban Drainage Pipelines" CJJ181-2012 standard.

[0162] Step S500: Based on digital twin technology, a multiphase flow simulation model is constructed (a three-dimensional geometric model is constructed using Ravit, and a multiphase flow simulation model is constructed using Abaqus and EDEM). The multiphase flow simulation model is used to obtain the types and severity levels of defects under rare and extreme operating conditions. The standardized multimodal dataset of the target pipeline (including target pipeline material, burial depth, operating parameters, and interface type) is input into the multiphase flow simulation model, and virtual samples are output. The virtual samples include the types and severity levels of defects under rare and extreme operating conditions. The multiphase flow simulation model simulates the multimodal data response under different defect types, different levels, and complex operating conditions (high groundwater level, low temperature freeze-thaw).

[0163] Step S600: Based on the virtual samples and combined with the collected real scarce operating condition samples, perform virtual-real calibration on the multiphase flow simulation model; set up a reinforcement learning environment to optimize the multiphase flow simulation model, output optimized virtual samples, feed them back to the multimodal collaborative diagnostic model and the multiphase flow simulation model, strengthen the multiphase flow simulation model, update the multimodal collaborative diagnostic model, and update the second diagnostic result as the final diagnostic result; the real scarce operating condition samples include rare and extreme operating condition samples;

[0164] The optimized virtual sample output process includes: constructing a reinforcement learning environment, setting a multi-objective reward function, adjusting the range and action space, using a dual-duel deep Q-network algorithm to iteratively optimize the multiphase flow simulation model, and outputting the optimized virtual samples.

[0165] The multi-objective reward function includes diagnostic accuracy, annotation cost, consistency of disease level assessment, and corresponding weights; the adjustment range is the weight of the parameters of the multiphase flow simulation model; and the action space is the sample selection strategy.

[0166] The virtual-real calibration includes calculating the virtual sample through MMD, combining the characteristic deviations of the collected real scarce working condition samples, and adjusting the fluid dynamics and structural parameters of the simulation model when the difference exceeds a threshold, so as to ensure that the virtual sample can accurately match the disease characteristics and grade correlation rules of the real pipeline network.

[0167] The reinforcement learning optimization aims at Pareto optimality of "diagnostic performance - annotation cost," constructing a complete reinforcement learning environment adapted to pipeline network operation and maintenance: the state space includes the current model parameters, sample selection strategy, diagnostic performance indicators (type recognition accuracy, level evaluation consistency), and digital twin simulation state; the action space includes model hyperparameter adjustment, sample selection K-value adjustment, and digital twin working condition simulation direction adjustment (such as strengthening the simulation of specific pipe materials and complex working conditions); the Dueling Double DQN algorithm is used to reduce overestimation problems and improve optimization stability; the reward function integrates four indicators: diagnostic accuracy, equipment stability, annotation cost, and virtual-real sample bias, and incorporates pipeline network operation and maintenance priorities (such as giving higher weight to the diagnostic accuracy of main pipe diseases than to branch pipes).

[0168] Virtual-real collaboration forms a closed-loop iteration: Reinforcement learning interacts with the digital twin virtual environment to iteratively adjust model parameters and screening strategies. The optimized parameters are synchronously updated to the multimodal collaborative diagnostic model, improving the model's diagnostic capabilities for complex working conditions and rare diseases. The model's diagnostic results and operation and maintenance feedback in real pipeline network scenarios (such as misdiagnosis cases and post-repair effect verification) inversely correct the weights of the reinforcement learning reward function, while guiding the digital twin model to add targeted working condition simulations (such as enhanced simulations for high-frequency disease types in operation and maintenance). After each iteration, the model performance is verified with real data. Optimization stops when the diagnostic accuracy is stable above 90% and the consistency of the grade assessment is stable above 85%, thus achieving continuous improvement in the model's generalization ability and long-term stability.

[0169] Step S700: The final diagnostic results are labeled with geographic information, expert diagnosis is introduced, and the expert diagnosis results are fed back to the multimodal collaborative diagnosis model and multiphase flow simulation model for relearning and updating. This guides the dynamic acquisition strategy of multimodal data of the target pipeline, forming a closed loop. The expert diagnosis results include misdiagnosed samples, missed samples, and operation and maintenance effect data.

[0170] The labeling involves marking the location, type, and severity of the disease on the geographic information using different colors and icons.

[0171] Step S800: Input the standardized multimodal dataset into the multimodal collaborative diagnostic model to achieve intelligent diagnosis of drainage pipelines.

[0172] This invention constructs a full-link feedback control system and establishes a multi-directional interactive channel to optimize the entire process around the disease diagnosis results: Maintenance personnel directly input misdiagnosed and missed cases (such as misjudgment of disease type or grade deviation) into the multimodal collaborative diagnosis model as high-value supplementary annotations, allowing for targeted model optimization; Operation and maintenance implementation effect data (such as flow velocity recovery after dredging, structural stability after crack repair, and pH changes after corrosion treatment) are fed back to the multiphase flow simulation model to correct simulation parameters, while simultaneously inputting them into the multimodal collaborative diagnosis model to adjust and optimize target weights (such as increasing the consistency weight of grade assessment); The development trend of diseases and high-risk areas (such as extremely severe disease areas and rapidly expanding disease areas) are visualized to guide the multimodal monitoring network to dynamically adjust sampling strategies, achieving a full-link closed loop of "diagnosis output - operation and maintenance feedback - data collection - model optimization," continuously improving disease diagnosis accuracy and operation and maintenance efficiency.

[0173] The above-disclosed embodiments are merely specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for intelligent diagnosis of drainage pipes, characterized in that, include: Multimodal data of the target pipeline is collected and dynamically preprocessed to generate a standardized multimodal dataset; the multimodal data includes pipeline image data, flow characteristic parameters, environmental parameters, and structural state data; Acquire source domain data, construct a deep transfer architecture, train it to form an initial diagnostic model, input the standardized multimodal dataset into the initial diagnostic model, and output the first diagnostic result; the diagnostic result includes the disease type and the severity level of the disease; the source domain data includes standard pipe full-scale test data and existing drainage pipe disease datasets; Based on the first diagnostic result, the initial diagnostic model is adjusted through active learning until the accuracy and consistency of the first diagnostic result both reach a preset threshold, generating a multimodal collaborative diagnostic model; a standardized multimodal dataset is input into the multimodal collaborative diagnostic model, and a second diagnostic result is output; A two-way calibration is set up to determine whether the second diagnostic result of the sample output of the standardized multimodal dataset violates physical laws; if it does, the sample and the corresponding second diagnostic result are marked as high-priority samples and fed back to the multimodal collaborative diagnostic model for relearning and optimization. Based on digital twin technology, a multiphase flow simulation model is constructed; the standardized multimodal dataset of the target pipeline is input into the multiphase flow simulation model, and virtual samples are output; the virtual samples include the types and severity levels of defects under rare and extreme conditions; Based on the virtual samples, the multiphase flow simulation model is calibrated to be both virtual and real based on the collected real and scarce working condition samples. A reinforcement learning environment is set up to optimize the multiphase flow simulation model, output optimized virtual samples, and feed them back to the multimodal collaborative diagnostic model and the multiphase flow simulation model to strengthen the multiphase flow simulation model, update the multimodal collaborative diagnostic model, and update the second diagnostic result as the final diagnostic result; the real scarce working condition samples include rare and extreme working condition samples; The final diagnostic results are labeled using geographic information, and expert diagnosis is introduced. The expert diagnosis results are fed back to the multimodal collaborative diagnostic model and the multiphase flow simulation model for further learning and updating. The expert diagnosis results include misdiagnosed samples, missed diagnoses, and operation and maintenance effect data. The standardized multimodal dataset is input into the multimodal collaborative diagnostic model to achieve intelligent diagnosis of drainage pipelines; The deep transfer architecture is a three-level deep transfer architecture, specifically including: a single-modal encoder, a cross-modal fusion layer, and a domain adaptation layer; the single-modal encoder is used to extract morphological features of diseases and temporal sensor features from the images of the source domain data; the cross-modal fusion layer is used to dynamically allocate modality weights; the domain adaptation layer introduces a generative adversarial network (GAN) to reduce the feature differences between the source domain data and the standardized multimodal dataset of the target pipeline; The active learning includes the following: The comprehensive value of the standardized multimodal dataset is calculated based on the initial diagnostic model; the expression for calculating the comprehensive value of the samples is: Where w is the modal weight and d is the feature difference coefficient; For the sample; H(x) is the measure of sample diagnostic uncertainty; U(x) is the overall value of the sample; Based on the comprehensive value of the samples, high-value samples are selected, manually labeled, and an expanded dataset is formed; the high-value samples include rare pipe disconnection samples, high-risk defects in main pipes, and multimodal feature conflict samples; The initial diagnostic model and the comprehensive value of the samples are adjusted in reverse based on the expanded dataset, and iterated until the accuracy and consistency of the first diagnostic result both reach a preset threshold.

2. The intelligent diagnostic method for drainage pipes as described in claim 1, characterized in that, The dynamic preprocessing includes the following: Eliminate image noise and artifacts in the multimodal data; The core sensitive parameters in the multimodal data were screened by grey relational entropy analysis; the core sensitive parameters include the flow velocity or water depth parameters corresponding to siltation, and the pH value or pipe wall thickness parameters corresponding to corrosion. Cross-modal feature alignment is achieved through contrastive learning; The standardized multimodal dataset is generated by dynamically adjusting the screening threshold and cross-modal alignment strategy of the core sensitive parameters by using the feedback of the disease feature contribution of the multimodal collaborative diagnostic model.

3. The intelligent diagnostic method for drainage pipes as described in claim 2, characterized in that, The grey relational entropy analysis and correlation screening process includes the following: The grey relational entropy analysis method is used to quantify the correlation between the multimodal data and the disease type and severity level, and the grey entropy correlation degree R is calculated as follows: ； in, The gray entropy correlation between the i-th data indicator and the target disease is represented; t is the data collection point. is the grey relational coefficient; m is the total number of sample collection points; Differentiated gray entropy correlation thresholds are set for different disease types, and the core sensitive parameters are screened by comparing the gray entropy correlation R with the threshold. The closer the gray entropy correlation degree R is to the threshold, the higher the correlation degree and the higher the sensitivity of the corresponding multimodal data.

4. The intelligent diagnostic method for drainage pipes as described in claim 1, characterized in that, The bidirectional calibration includes the following: The expanded dataset is input into the multimodal collaborative diagnostic model for training; By combining physical knowledge, a hybrid loss function is constructed, using physical knowledge as a constraint on the loss function, and the weight of the physical constraint is dynamically adjusted; the physical knowledge includes fluid dynamics, pipe structure mechanics, and solid-liquid coupling. The hyperparameters are optimized using Bayesian optimization, and overfitting is avoided through five-fold cross-validation; the calibration and diagnostic results are then output. Determine whether the second diagnostic result contradicts the calibration diagnostic result. If they do, increase the corresponding physical constraint weight, mark the second diagnostic result as a high-priority sample, and feed it back to the multimodal collaborative diagnostic model for further learning and optimization.

5. The intelligent diagnostic method for drainage pipes as described in claim 4, characterized in that, The dynamic adjustment of physical constraint weights includes a full-process dynamic mechanism of pre-setting, triggering, adjusting, and calibrating; the full-process dynamic mechanism includes the following: Based on the characteristics of different disease types, a preset weight gradient is used as the benchmark for dynamic adjustment, and the triggering conditions for dynamic adjustment are determined. Based on changes in operating conditions and the accuracy of the multimodal collaborative diagnostic model, auxiliary triggering conditions are set. In conjunction with these triggering conditions, the adjustment requirements of the weights are captured, and the weights are adjusted accordingly. After the weights are adjusted, bidirectional calibration is completed, high-priority samples are marked, and the data is fed back to the multimodal collaborative diagnostic model for further learning and optimization. The multiphase flow simulation model simulates the disease response corresponding to the adjusted weights, outputs optimized virtual samples, and performs virtual-real calibration. The adjustment range is corrected based on the virtual and real calibration results, and the weight benchmark value is iteratively optimized to achieve dynamic adjustment of the weight under different disease types and different working conditions.

6. The intelligent diagnostic method for drainage pipes as described in claim 1, characterized in that, The optimized virtual sample output process includes: constructing a reinforcement learning environment, setting a multi-objective reward function, adjusting the range and action space, using a dual-duel deep Q-network algorithm to iteratively optimize the multiphase flow simulation model, and outputting the optimized virtual samples.

7. The intelligent diagnostic method for drainage pipes as described in claim 6, characterized in that, The multi-objective reward function includes diagnostic accuracy, annotation cost, consistency of disease level assessment, and corresponding weights; the adjustment range is the weight of the multiphase flow simulation model parameters; and the action space is the sample selection strategy.

Images