A method and system for inspecting urban water supply and drainage pipelines

Abstract

The application discloses a kind of for city supply and drainage pipeline's inspection method and system, specifically relates to city pipeline inspection planning field, for solving the problem of low efficiency of existing artificial inspection, scheduling allocation is uneven and multiple inspection unit operation conflict;The method constructs the state space of multi-agent cooperative inspection scheduling by receiving the real-time state data stream of pipeline network, calculates the risk score of pipeline node based on topological dependence coefficient and state abnormal index;Action strategy sequence of each agent is generated using a pre-trained collaborative learning model, and the inspection node priority and path planning scheme are output;After identifying the agent with resource conflict, the strategy update and optimization are performed according to the global risk score and conflict cost;And after completing the second conflict verification, the actual inspection scheduling is carried out, so as to realize the efficient cooperation and dynamic optimization management of city pipe network inspection.

Description

Technical Field

[0001] This invention relates to the field of urban pipeline inspection and planning technology, and more specifically, to an inspection method and system for urban water supply and drainage pipelines. Background Technology

[0002] Current urban water supply and drainage networks are generally large in scale and complex in distribution. Underground pipelines are densely interwoven and their operating environment is affected by multiple factors such as geographical conditions, climate change, and human construction. Potential hazards such as pipeline leaks, blockages, and ruptures are characterized by suddenness and regional clustering. Traditional manual inspection methods suffer from problems such as long inspection cycles, low dispatch efficiency, and delayed risk identification, making it difficult to meet the requirements of modern cities for the continuity and safety of water supply and drainage systems.

[0003] With the widespread adoption of IoT sensing and intelligent inspection robots or other terminal devices in municipal infrastructure, pipeline operating status can be collected in real time through multi-dimensional data such as pressure and flow. However, how to achieve coordinated scheduling and task allocation of multiple inspection units in a massive data environment has become a core challenge. Especially in scenarios such as dense pipelines in old urban areas, limited underground space, peak nighttime drainage, rainstorm emergencies, road closures, or limited underground space, scheduling of multiple inspection units is prone to path conflicts and overlapping operations, affecting inspection efficiency and safety.

[0004] Therefore, there is an urgent need for an urban water supply and drainage pipeline inspection method based on multi-agent collaborative learning, so as to achieve dynamic optimization allocation of inspection paths and resources, thereby realizing continuous, adaptive, and efficient intelligent inspection scheduling in complex urban pipeline network environments. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method and system for inspecting urban water supply and drainage pipelines to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for inspecting urban water supply and drainage pipelines includes the following steps:

[0008] S1. Receive real-time status data streams from the urban water supply and drainage pipeline network and construct a state space for multi-agent collaborative inspection and scheduling.

[0009] S2. Set the dynamic weights of pipeline nodes in the initial state space of each agent. The dynamic weights are calculated based on the topology dependency coefficient and the anomaly index of the state data, and the risk score of the pipeline nodes is calculated in combination with the environmental data.

[0010] S3. Based on pipeline node risk scoring, generate action strategy sequences for each agent through a pre-trained collaborative learning model, including pipeline node inspection action priority and path planning;

[0011] S4. In the action policy sequence of all agents, identify the conflicting agents with resource conflicts, and update the action policy of the conflicting agents based on the global pipeline node risk score and conflict cost.

[0012] S5. After the action strategy of the conflicting agents is updated, a second verification of resource conflicts is performed on the action strategy sequences of all agents until the verification is passed, and then the actual inspection and scheduling of agents is carried out.

[0013] In a preferred embodiment, step S1, receiving the real-time status data stream of the urban water supply and drainage pipeline network and constructing the state space for multi-agent collaborative inspection and scheduling, specifically includes:

[0014] The pipeline pressure and flow rate are collected at a fixed sampling frequency as real-time status data streams, while environmental data streams are accessed from the urban environmental monitoring center;

[0015] Construct a graph model based on the pipeline network topology, where pipeline nodes represent pipeline intersections and edges represent connecting pipeline segments;

[0016] Each agent is assigned a connected subgraph as its inspection responsibility area, and the set of pipeline nodes covered by the responsibility area constitutes the agent's local observation space.

[0017] The real-time state data values ​​in the local observation space of each agent are arranged in chronological order to form a state vector. The elements of the state vector include the current state data value and the state data of the historical monitoring period.

[0018] The state vectors of individual agents are stored in a shared memory area to form a state space matrix. The rows of the matrix correspond to the pipeline node indices in the local observation space, and the columns correspond to the state vectors.

[0019] In a preferred embodiment, step S2 involves setting dynamic weights for pipeline nodes in the initial state space of each agent. These dynamic weights are calculated based on topological dependency coefficients and anomaly indices from state data, and the calculation of pipeline node risk scores is combined with environmental data. Specifically, this includes:

[0020] Dynamic weights are assigned to pipeline nodes corresponding to the state space of the agent. The initial value of the dynamic weight is set based on the number of users served by the pipeline node.

[0021] The state vectors of pipeline nodes are extracted using a graph convolutional network as feature embeddings, and the topological dependency coefficients between pipeline nodes are calculated using an attention mechanism.

[0022] An anomaly index is calculated based on the deviation between real-time status data and historical baselines of pipeline node status vectors.

[0023] Update the dynamic weights of pipeline nodes based on the normalized result of the Hadamard product operation performed on the topology dependency coefficient and the anomaly index.

[0024] Calculate the mutual information value between environmental data and historical pipeline node failure events, and use the result as the environmental factor weight;

[0025] By combining dynamic weight vectors and environmental factor weights, and smoothing them through linear combination and time window processing, a pipeline node risk score is output.

[0026] In a preferred embodiment, in step S3, based on the pipeline node risk score, a pre-trained collaborative learning model is used to generate action strategy sequences for each agent, including pipeline node inspection action priority and path planning, specifically including:

[0027] The agent's state space and the risk score distribution of pipeline nodes are input into a pre-trained collaborative learning model, and the model outputs a sequence of action strategies for the agent on pipeline nodes within the responsibility area.

[0028] The collaborative learning model adopts a centralized training and distributed execution architecture, and is trained using digital twin simulation data of the agent's state space and the risk score distribution of pipeline nodes, specifically:

[0029] A policy network for the agent is constructed, and a local reward function is built based on the local observation space inspection coverage completeness rate and the pipeline node risk score decay rate. The policy network parameters are iterated through simulation in a digital twin environment.

[0030] The policy network outputs the predicted action priority values ​​of each pipeline node within the agent's inspection responsibility area based on the input risk score distribution and pipeline node topology.

[0031] Based on the distribution of action priority prediction values, an initial inspection pipeline node priority sequence is generated for the agent and input into the path optimizer. The final inspection path output by the path optimizer updates the initial inspection pipeline node sequence, thus constructing the agent's action policy sequence.

[0032] In a preferred embodiment, the path optimizer calculates the final inspection path with the objective function of minimizing the overall cost; the overall cost is calculated by weighting the travel time cost between pipeline nodes and the cost of pipeline node risk scores increasing over time.

[0033] In a preferred embodiment, step S4 involves identifying conflicting agents with resource conflicts from the action policy sequences of all agents, and updating the action policies of these conflicting agents based on global pipeline node risk scores and conflict costs. This specifically includes:

[0034] Analyze all agent action strategy sequences to identify conflicting agents with pipeline node monopoly, path intersection, and resource overrun conflicts. Each conflicting agent broadcasts the target pipeline node and planned arrival time corresponding to its action strategy to each other.

[0035] Each conflicting agent receives the action policies of other agents and calculates the conflict cost based on the overlap rate of pipeline node resource occupancy, path intersection length, and resource load.

[0036] Using the comprehensive performance index, which consists of the global pipeline node risk score decay rate and the conflict cost reduction rate, as the optimization objective, the action strategy of the conflict agent is adjusted in multiple rounds, and the action strategy corresponding to the optimal comprehensive performance index is selected to update the original action strategy sequence of the conflict agent.

[0037] In a preferred embodiment, the multi-round strategy adjustment process involves adjusting the planned arrival time of conflicting agents according to a fixed time step, and recalculating the corresponding comprehensive performance index after each round of adjustment.

[0038] In a preferred embodiment, step S5, after the action strategy of the conflicting agents is updated, performs a secondary verification of resource conflicts on all agent action strategy sequences until the verification is passed, and then performs actual agent inspection and scheduling. This specifically includes:

[0039] After the action strategy of the conflicting agents is updated, a second identification is performed on the conflicting agents in all agent action strategy sequences that have pipeline node monopoly, path intersection and resource over-limit conflict.

[0040] The process of calculating comprehensive performance indicators and adjusting the action strategies of conflicting agents is repeated until the conflict verification of all agent action strategy sequences is passed, at which point the conflicting agent identification and action strategy adjustment process ends.

[0041] Actual inspection and scheduling of agents are carried out based on the final action strategy sequence.

[0042] On the other hand, the present invention provides an inspection system for urban water supply and drainage pipelines, comprising:

[0043] The data acquisition module is used to receive real-time status data streams and environmental data of the urban water supply and drainage pipeline network, and to construct a multi-agent collaborative inspection status space based on topology.

[0044] The risk assessment module is used to calculate the dynamic weight of pipeline nodes based on the topology dependency coefficient and the state anomaly index, and to generate pipeline node risk scores in combination with environmental factors.

[0045] The collaborative decision-making module is used to call a pre-trained collaborative learning model based on the risk score of pipeline nodes to generate action strategy sequences for each agent and inspection node priorities;

[0046] The path optimization module is used to plan the path of the initial inspection node sequence with the goal of minimizing the overall cost, output the final inspection path and update the action strategy.

[0047] The conflict coordination module is used to identify resource conflicts such as node monopoly and path intersection in multi-agent action strategies, and generate a conflict-free strategy sequence through multiple rounds of strategy adjustment.

[0048] The technical effects and advantages of the present invention regarding the inspection method and system for urban water supply and drainage pipelines are as follows:

[0049] By introducing a multi-agent collaborative learning model and a digital twin simulation environment, intelligent allocation and dynamic optimization of urban water supply and drainage pipeline inspection tasks are achieved. The method can generate globally optimal action strategy sequences by comprehensively considering pipeline node risk scores, spatial topological relationships, and environmental factors in complex pipeline network topologies and multi-source data environments, significantly improving inspection coverage and scheduling efficiency. A path optimizer is used to minimize the comprehensive cost of the initial inspection path, reducing ineffective movement and time waste between inspection nodes. A conflict identification and multi-round strategy adjustment mechanism is introduced to enable multiple inspection agents to achieve path decoupling and task division in high-density pipeline areas, avoiding resource conflicts and overlapping tasks. The system can automatically revise action strategies based on real-time data changes, enabling rapid response and dynamic scheduling in areas with sudden risks.

[0050] Compared with traditional manual inspection and single-agent scheduling modes, this invention can improve inspection accuracy and timeliness while ensuring safety, reduce operation and maintenance costs, and has engineering feasibility and economic value for promotion and application in large-scale urban pipe networks. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of an inspection method for urban water supply and drainage pipelines according to the present invention;

[0052] Figure 2 This is a schematic diagram of the structure of an inspection system for urban water supply and drainage pipelines according to the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0054] Example 1

[0055] Figure 1 The present invention provides a method for inspecting urban water supply and drainage pipelines, which includes the following steps:

[0056] S1. Receive real-time status data streams from the urban water supply and drainage pipeline network and construct a state space for multi-agent collaborative inspection and scheduling.

[0057] S2. Set the dynamic weights of pipeline nodes in the initial state space of each agent. The dynamic weights are calculated based on the topology dependency coefficient and the anomaly index of the state data, and the risk score of the pipeline nodes is calculated in combination with the environmental data.

[0058] S3. Based on pipeline node risk scoring, generate action strategy sequences for each agent through a pre-trained collaborative learning model, including pipeline node inspection action priority and path planning;

[0059] S4. In the action policy sequence of all agents, identify the conflicting agents with resource conflicts, and update the action policy of the conflicting agents based on the global pipeline node risk score and conflict cost.

[0060] S5. After the action strategy of the conflicting agents is updated, a second verification of resource conflicts is performed on the action strategy sequences of all agents until the verification is passed, and then the actual inspection and scheduling of agents is carried out.

[0061] In S1, the real-time status data stream of the urban water supply and drainage pipeline network is received, and a state space for multi-agent collaborative inspection and scheduling is constructed.

[0062] Key operating parameters of water supply and drainage pipelines are collected in real time at a fixed sampling frequency. The sampling frequency is determined based on the time-varying characteristics of pipeline operation dynamics; for example, it is set to once per minute in main urban pipelines and once every five minutes in branch pipes or low-flow-velocity areas. The collected real-time data mainly includes pressure and flow parameters within the pipeline. Pressure data reflects the pipeline's stress and fluid transport stability, while flow data reflects the water supply and drainage load and flow uniformity. To supplement the impact of the external environment on pipeline operation, environmental data streams from the urban environmental monitoring center are also connected. Environmental data includes information such as temperature, humidity, rainfall, and groundwater level. These data are integrated into the same timeline in a time-synchronized manner to ensure that pressure, flow, and environmental changes correspond at the same moment.

[0063] Based on the network topology of urban water supply and drainage pipelines, a graph model is constructed. The model uses pipeline intersections as vertices and pipeline segments between adjacent nodes as edges. Each edge records the length, diameter, material, design pressure rating, and geographic coordinates of the corresponding pipe segment. During modeling, engineering archives and construction completion data of the urban underground pipeline network are used to determine the connection relationships and spatial location information between nodes. To ensure the connectivity of the graph model, logical virtual nodes are added to edge segments with inter-regional pipelines, keeping the graph structure closed and free of isolated points at the computational level. After construction, the model is stored in matrix form, with each node having a unique index number, and the connection relationships between nodes are described using an adjacency list. To achieve spatial independence of task allocation, the overall topology model is divided into multiple connected subgraphs based on regional characteristics. Each connected subgraph corresponds to the working range of an intelligent inspection unit. The division principle is based on spatial continuity and task balance, ensuring that the pipeline length and number of nodes covered by each subgraph are within a reasonable range. For example, each subgraph in the main pipeline area contains 30 to 50 nodes, while in densely branched areas, the number is limited to 15 to 30 nodes. The set of connected subgraphs after division is defined as the inspection responsibility area, and the set of nodes within the responsibility area constitutes the local observation space of the intelligent agent, providing a physical boundary for subsequent risk calculation. The intelligent agent is a mobile detection unit with autonomous inspection, data acquisition, and path execution capabilities. It can be a ground mobile robot or a pipeline walking robot equipped with pressure sensors, flow sensors, and positioning modules, or a drone unit equipped with environmental monitoring devices, used to independently perform inspection tasks and participate in collaborative scheduling in different areas of the urban water supply and drainage pipeline network.

[0064] After the responsibility areas are divided, the real-time state data in the local observation space of each agent are arranged in chronological order to construct corresponding state vectors. Each column of the state vector represents the state change sequence of the same pipeline node at different sampling times, and each row represents the set of state values ​​of all nodes at a single point in time. In addition to the pressure and flow values ​​at the current moment, the vector also contains data from the most recent monitoring periods to reflect the dynamic evolution characteristics of the node state. For example, if the sampling period is one minute, the state vector can contain continuous sampling values ​​from the past five minutes to reflect the pressure fluctuation trend. The constructed state vectors are uniformly encoded in the temporal dimension, and a unique index is assigned to each node. The state vectors of all agents are centrally stored and accessed through a shared memory area, forming a state space matrix for multi-agent collaborative inspection. The rows of this matrix correspond to the pipeline node indices in each responsibility area, the columns correspond to the state values ​​in the continuous time series, and the matrix elements are the pressure or flow data of a specific node at a specific moment. To ensure the real-time update performance of the matrix, a sliding mechanism based on a time window is adopted, that is, at the end of each sampling period, the latest sampled data is inserted at the end of the matrix, while the oldest expired data is removed, thereby maintaining the dynamic stability of the data structure.

[0065] In step S2, dynamic weights are set for pipeline nodes in the initial state space of each agent. These dynamic weights are calculated based on the topology dependency coefficient and the anomaly index of the state data, and the risk score of the pipeline nodes is calculated in conjunction with environmental data.

[0066] Dynamic weights are assigned to pipeline nodes in the state space of each agent to characterize the importance of different nodes in inspection and scheduling. The initial values ​​of the dynamic weights are set based on the number of users served by each pipeline node, obtained from water supply household registers or water metering records. For example, main water supply nodes typically serve over 1,000 users, corresponding to an initial dynamic weight of 1.0; branch nodes or terminal branch nodes serve fewer users, generally less than 100, corresponding to an initial dynamic weight of 0.3 to 0.5. To avoid the magnitude difference between node weights causing bias in subsequent calculations, the initial weights of all nodes are linearly normalized, limiting the weight values ​​to between 0 and 1. After weight assignment, the dynamic weights are written into the node attribute field of the agent's state space matrix.

[0067] Based on the constructed state space matrix, a graph convolutional network is used to extract features from the state vectors of pipeline nodes, obtaining high-dimensional feature embeddings of nodes within the topological structure. The graph convolution process uses the topological connectivity of the pipeline network as input and node state data such as pressure and flow rate as input features. Through multiple convolutional operations, it aggregates the state information of adjacent nodes, thereby extracting the operational correlation features of nodes within the local network. Each convolutional operation weights and summarizes the node's own features with those of its neighboring nodes; the weight coefficients are determined by the node's connection strength and spatial distance. For example, when the pipe segment length between two nodes is short and the flow rate trends are consistent, a higher convolutional aggregation weight is used; otherwise, a lower weight is used. The embedded feature vectors obtained after multiple convolutions comprehensively reflect the topological dependencies and state coupling features between nodes. To enhance the accuracy of node correlation modeling, an attention mechanism is introduced during the aggregation process of the graph convolutional network to adaptively allocate the degree of influence between nodes. The attention mechanism compares the contributions of different neighboring nodes to the current node's state change, assigning higher attention coefficients to highly relevant nodes and lower attention coefficients to lowly relevant nodes, thereby generating a coefficient matrix that reflects the strength of topological dependencies.

[0068] For each pipeline node's state vector, an anomaly index is calculated based on the deviation between real-time state data and the historical baseline. The historical baseline data consists of the node's average pressure and flow rate or stable range during its normal operating cycle, obtained by establishing a historical sample database during the initial phase of pipeline network operation. Real-time state data represents the actual monitored values ​​within the current sampling period; the difference between the two reflects whether the node's state has deviated abnormally. The anomaly index is calculated using the relative deviation method, measured by the ratio of the difference between the real-time data and the baseline value to the baseline value. For example, if the current pressure value of a node drops by more than 10% compared to the historical average, its anomaly index is set to 0.1; if the drop reaches 30%, the corresponding anomaly index is increased to 0.3. To suppress the interference of instantaneous fluctuations on the results, a time window smoothing process is used, taking a moving average of the anomaly index over several consecutive sampling periods to reflect the persistence of anomalies within a short period. The higher the anomaly index, the greater the deviation of the node's operating state from the normal range during that time period.

[0069] The dynamic weights of pipeline nodes are updated using the calculated topology dependency coefficient and anomaly index. To achieve feature fusion of the two types of data, element-wise multiplication is used to perform Hadamard product operations, ensuring that the topology dependency strength and anomaly offset of each node correspond on the same dimension. Specifically, when a node has a high topology dependency coefficient and a large anomaly index, its corresponding product result will be significantly higher than that of other nodes, reflecting that the node possesses both structural criticality and operational anomaly in the overall network. After the operation, the result vector is normalized to limit the weight values ​​of all nodes to the range of 0 to 1, ensuring the comparability and numerical stability of the weight distribution. The normalization method uses extreme value mapping, i.e., linear scaling is performed with the maximum and minimum values ​​of the product result as boundaries.

[0070] After the dynamic weights are updated, the mutual information value between environmental data and historical pipeline node failure events is calculated to quantify the correlation between external environmental changes and failure occurrence. Environmental data includes variables such as temperature, humidity, rainfall, and surface water level, while historical failure event records include the occurrence time and node location of leaks, ruptures, and blockages. Mutual information calculation is performed on a time window basis, with each window potentially set to 24 hours, to capture the impact of short-term climate change on pipeline operation. During calculation, the magnitude of environmental data changes is paired with the frequency of failure events within the corresponding time period to form a joint distribution, thereby determining the information dependence between the environment and the failure. For example, during the rainy season, the mutual information value between rainfall and failure events is typically higher than 0.6, indicating that rainwater infiltration and changes in ground load are important factors inducing risk. The mutual information calculation results, after normalization, are input as environmental factor weights into the subsequent risk assessment model to correct the node dynamic weights, ensuring that the risk assessment reflects both structural status and external conditions.

[0071] The updated dynamic weight vector and environmental factor weights are combined to calculate the final pipeline node risk score. The combination process uses a linear combination method, weighting the two types of weights at a fixed ratio. This ratio is preset based on historical validation results; for example, dynamic weight accounts for 0.7, and environmental weight accounts for 0.3, to ensure that the internal state characteristics of the node dominate the evaluation. After combination, to avoid short-term fluctuations affecting the stability of the results, a time window smoothing process is performed on the risk score values ​​over multiple consecutive sampling periods. The smoothing operation uses a moving average, calculating the average value based on the data from the most recent five sampling periods, so that the risk score changes gradually over time and reflects the trend characteristics. The final risk score represents the node's risk level. The risk scores of all nodes are stored in a risk distribution table, providing a decision-making basis for subsequent agent inspection strategy generation and path optimization.

[0072] In S3, based on the pipeline node risk score, a pre-trained collaborative learning model is used to generate action strategy sequences for each agent, including the priority of pipeline node inspection actions and path planning.

[0073] The constructed agent state space and the corresponding risk score distribution of pipeline nodes are fed into the pre-trained collaborative learning model as input data. The input data includes the state vector, risk score, and adjacency information of each pipeline node within the topology. The state space records the pressure and flow change sequences of each node over a continuous monitoring period, while the risk score reflects the potential risk level of the node under multiple factors including structure, operation, and environment. The collaborative learning model operates with a centralized training and distributed execution architecture. During the training phase, the central control unit uniformly schedules the state data of each agent to update the model parameters. During the execution phase, each agent independently uses the pre-trained model weights to generate its own action decisions. The training data comes from a digital twin simulation environment built based on real pipeline topology and historical operating data, capable of simulating state changes and risk propagation processes under different operating conditions. During training, the simulation time step is set to 1 minute, and the duration is no less than 72 hours to ensure the model learns the temporal dependencies between nodes. The model continuously adjusts its policy parameters through multiple rounds of simulation to maintain efficient decision-making performance under different risk distributions. After training, the model, upon inputting the current state space and risk score distribution, outputs a sequence of action strategies for each agent within its corresponding responsibility area. This sequence includes the access order and action priority of pipeline nodes. Furthermore, the action strategy sequence also includes the planned arrival time of pipeline nodes, the path space coordinates between nodes, and corresponding resource occupancy constraints. This guides the agent to complete inspection, detection, and data collection tasks sequentially within a specified time window, ensuring optimal spatiotemporal coordination and risk coverage during the inspection process.

[0074] During the training phase of the collaborative learning model, a core policy network is constructed to generate action decisions. The policy network takes the agent's local observation space as input and establishes a nonlinear mapping relationship between input states and action outputs. In the design of the policy network, the inspection coverage completeness rate and the pipeline node risk score decay rate are the main indicators for constructing a local reward function. The inspection coverage completeness rate measures the proportion of nodes visited within the agent's responsibility area within a certain time window, with a target value set above 90%. The risk score decay rate evaluates the degree to which the node risk value decreases after the inspection task is executed, with a target value set between 10% and 30%. During training, the digital twin environment simulates different node risk mutations, traffic anomalies, and environmental changes using random perturbations. The policy network obtains reward feedback through continuous interaction with the environment and updates the parameter weights after each simulation. The update method adopts an experience replay and parameter aggregation mechanism, that is, after each round of simulation, the agent's state-action-reward sequence is collected, and the policy parameters are adjusted by comprehensively adjusting multiple batches of data. After multiple iterations, the policy network outputs a stable action tendency based on local state information and assigns priority weights to each node. After the policy network is trained, it is deployed in the execution phase of the collaborative learning model. During execution, the policy network receives the risk score distribution and topological connectivity of pipeline nodes within the responsibility area as input, comprehensively analyzes the risk level, spatial location, and adjacency influence of the nodes, and outputs the predicted action priority value of each node.

[0075] Based on the distribution of action priority predictions output by the policy network, an initial inspection pipeline node priority sequence is generated for each agent. The predicted values ​​of each node are sorted in descending order to generate a node access order table, which serves as the agent's initial inspection priority sequence within its responsibility area. To reduce path duplication caused by excessive node concentration, the sorting results are processed by adjacency sparsification, i.e., spatially close node groups are merged, and the access order is rearranged according to the average priority of the groups. After processing, the resulting node sequence balances the continuity of the inspection path with the coverage requirements of high-risk nodes. This initial priority sequence is input into the path optimizer to calculate the optimal inspection path under spatial and temporal constraints. The input data for each agent includes parameters such as the responsibility area topology, node spatial coordinates, node priority values, and inspection speed limits. All parameters are input to the optimizer in a uniform format for a comprehensive cost calculation process.

[0076] The path optimizer, during the execution phase, calculates the final inspection path for the agent within its responsibility area, with the objective function of minimizing the overall cost. The overall cost consists of two parts: the movement time cost between nodes, measuring the time taken for the agent to travel between nodes; and the cost of the pipeline node risk score increasing over time, measuring the impact of delayed inspections on risk accumulation. The movement time cost is calculated based on the spatial distance between nodes and the agent's movement speed. Distance data comes from node coordinates in the pipeline topology model, and speed values ​​are set according to the agent's physical movement characteristics, for example, an average inspection speed of 1.2 meters per second. The risk growth cost is calculated based on the rate of change of a node's risk score. When a node's risk score increases by more than 10% within a continuous time window, the corresponding cost increases proportionally. To balance the importance of the two types of costs, weighted coefficients are used for linear combination. The default weight for movement time cost is set to 0.4, and the weight for risk growth cost is set to 0.6, ensuring both timeliness and risk control in path planning. The path optimization process employs a heuristic search strategy, selecting the path with the minimum overall cost as the final output after traversing the candidate path set. The optimizer outputs an ordered list of node access sequences, including the corresponding time order. Finally, the inspection paths output by the path optimizer are used to update the agent's initial node priority sequence, forming the corresponding adjusted action policy sequence.

[0077] In step S4, among all agents' action policy sequences, conflicting agents with resource conflicts are identified, and their action policies are updated based on global pipeline node risk scores and conflict costs.

[0078] A centralized analysis of the action strategy sequences of all agents was conducted to identify agents with resource conflicts. The action strategy sequences recorded the order of each agent's access to inspection nodes, planned arrival time, and task duration. By comparing the action sequences of different agents, three typical conflicts were identified: pipeline node monopoly conflict, path intersection conflict, and resource overload conflict. Pipeline node monopoly conflict refers to multiple agents planning to access the same node within the same time period, potentially causing duplicate work or detection interference; path intersection conflict refers to two inspection paths having spatially overlapping sections and similar access times, potentially leading to collision risks; resource overload conflict refers to the number of agents operating simultaneously in the same area exceeding the safety limit, causing traffic congestion or energy overload. The identification method involved pairwise comparison of node indices and access time intervals in the agent's action sequences. If the nodes were identical and the time overlap exceeded 50%, it was determined to be an monopoly conflict; if the path overlap distance exceeded 30% of the total path length, it was determined to be a path intersection conflict. The identification results form a set of conflict agents. Each conflict agent sends its current action strategy's target node, planned arrival time, and path node sequence to the conflicting objects in the same group through a broadcast mechanism, ensuring that each agent shares each other's action intention information within the same time window.

[0079] After conflict identification and information broadcasting, each conflicting agent calculates its own conflict cost based on the received action strategy information from other agents. The conflict cost quantifies the degree of competition between agents in terms of time, space, and resources. The calculation involves three indicators: pipeline node resource overlap rate, path intersection length, and resource load. The resource overlap rate represents the proportion of time overlap when different agents access the same node. When the overlap rate exceeds 70%, the conflict cost is set to a high-level value of 1.0; when the overlap rate is less than 30%, the cost is set to 0.3. The path intersection length is obtained by calculating the overlapping distance of different inspection paths in geographic space. If the overlap length exceeds 20% of the total path length, the corresponding conflict cost is added to that portion. The resource load is determined by comparing the number of agents conducting parallel inspections in the area with a set safety limit. For example, a maximum of two agents are allowed to operate simultaneously in the same area. When three or more parallel operations are detected, the conflict cost is increased proportionally for the excess. The combined results of these three indicators form a single conflict cost value, used to characterize the conflict intensity between the current agent and other agents. This generation value is input into the subsequent optimization process to guide the adjustment of action strategies.

[0080] After the conflict cost calculation is completed, the comprehensive performance index, composed of the global pipeline node risk score decay rate and the conflict cost reduction rate, is used as the optimization objective to perform multiple rounds of adjustments to the action strategy of the conflict agent. The comprehensive performance index reflects the overall optimality of the inspection task in terms of both risk reduction and resource coordination. The risk decay rate represents the proportion by which the risk score of high-risk nodes in the responsibility area decreases over time under the current action strategy; the conflict cost reduction rate represents the magnitude of the decrease in conflict intensity after multiple rounds of adjustments. To achieve a balance between the two indicators, the risk decay rate is weighted at 0.6 and the conflict cost reduction rate at 0.4, and the comprehensive performance index is formed by weighted summation. The optimization process adopts a multi-round strategy adjustment method with a fixed time step, set at 30 seconds. Each round of adjustment is achieved by modifying the planned arrival time of the conflict agent. During the adjustment process, the node access order remains unchanged, and only the arrival time is changed to ensure the stability of the inspection coverage. After each round of time adjustment, the global comprehensive performance index is recalculated and compared with the results of the previous round. When the overall performance index improves by less than 2% for three consecutive rounds, the adjustment process is terminated, and the action strategy corresponding to the round with the highest overall performance index is selected as the final update result. After the update is completed, the original action strategy sequence is replaced to ensure that the conflicting agents reallocate task timing under globally optimal conditions, thereby achieving coordinated execution of inspection tasks.

[0081] In S5, after the action strategy of the conflicting agents is updated, a secondary verification of resource conflicts is performed on the action strategy sequences of all agents until the verification is passed, and then the actual inspection and scheduling of agents is carried out.

[0082] After the action strategies of conflicting agents are optimized and updated, a conflict check is performed again on the action strategy sequences of all agents to ensure that there are no new resource conflicts or path interferences globally. During this stage, the latest action strategy sequences of all agents are synchronized to ensure that the planned arrival time, node access order, and inspection duration of each sequence are all under a unified time benchmark. After synchronization, a secondary conflict identification process is executed. The judgment rules for secondary identification are the same as those for the initial identification, but the analysis scope is expanded to all updated strategy sequences, focusing on identifying indirect conflicts caused by adjustments to the strategies of individual agents. This secondary identification mechanism can detect the cascading effects caused by adjustments to the strategies of some agents, avoiding global inconsistencies caused by local optimization.

[0083] After the new set of conflict agents is determined, the comprehensive performance index calculation and action strategy adjustment process is re-executed. This process uses the comprehensive performance index, composed of the global pipeline node risk score decay rate and conflict cost reduction rate, as the optimization objective to ensure that the adjusted action strategy maintains a balance between risk control and resource coordination. In each loop, the planned arrival time of the conflict agents is fine-tuned with a fixed time step of 30 seconds. After each adjustment, the comprehensive performance index is recalculated and compared with the previous round's result. If the comprehensive performance index continues to improve, it indicates that the adjustment direction is correct. If the index change is less than 2% for two consecutive rounds, it indicates that the system is stabilizing. At this point, the group of conflicts is marked as resolved and removed from the conflict set. The entire loop process continues until all agents' action strategy sequences pass the conflict verification, indicating that global resource allocation and timing arrangements have reached a coordinated state. After the loop ends, the final version of the action strategy sequence is solidified into the agent scheduling instruction set and enters the actual inspection and scheduling phase. Actual inspection scheduling is executed according to the finalized action strategy, including node access order, path direction, and time allocation, ensuring that each agent completes the inspection task of the designated area within the predetermined time window. Data records from this stage are synchronously transmitted back to the system log for subsequent verification and model retraining, providing empirical data support for the next inspection cycle.

[0084] Example 2

[0085] The difference between Embodiment 2 and Embodiment 1 is that this embodiment introduces an inspection system for urban water supply and drainage pipelines.

[0086] Figure 2 A schematic diagram of an inspection system for urban water supply and drainage pipelines according to the present invention is provided. The inspection system for urban water supply and drainage pipelines includes:

[0087] The data acquisition module is used to receive real-time status data streams and environmental data of the urban water supply and drainage pipeline network, and to construct a multi-agent collaborative inspection status space based on topology.

[0088] The risk assessment module is used to calculate the dynamic weight of pipeline nodes based on the topology dependency coefficient and the state anomaly index, and to generate pipeline node risk scores in combination with environmental factors.

[0089] The collaborative decision-making module is used to call a pre-trained collaborative learning model based on the risk score of pipeline nodes to generate action strategy sequences for each agent and inspection node priorities;

[0090] The path optimization module is used to plan the path of the initial inspection node sequence with the goal of minimizing the overall cost, output the final inspection path and update the action strategy.

[0091] The conflict coordination module is used to identify resource conflicts such as node monopoly and path intersection in multi-agent action strategies, and generate a conflict-free strategy sequence through multiple rounds of strategy adjustment.

[0092] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0093] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0094] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0095] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0096] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0097] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0098] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0099] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0101] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for inspecting urban water supply and drainage pipelines, characterized in that, Includes the following steps: S1. Receive real-time status data streams from the urban water supply and drainage pipeline network and construct a state space for multi-agent collaborative inspection and scheduling. S2. Set the dynamic weights of pipeline nodes in the initial state space of each agent. The dynamic weights are calculated based on the topology dependency coefficient and the anomaly index of the state data, and the risk score of the pipeline nodes is calculated in combination with the environmental data. S3. Based on pipeline node risk scoring, generate action strategy sequences for each agent through a pre-trained collaborative learning model, including pipeline node inspection action priority and path planning; S4. In the action policy sequence of all agents, identify the conflicting agents with resource conflicts, and update the action policy of the conflicting agents based on the global pipeline node risk score and conflict cost. S5. After the action strategy of the conflicting agents is updated, a second verification of resource conflicts is performed on the action strategy sequences of all agents until the verification is passed, and then the actual inspection and scheduling of agents is carried out. In step S3, based on pipeline node risk scores, a pre-trained collaborative learning model is used to generate action strategy sequences for each agent, including pipeline node inspection priority and path planning, specifically including: The agent's state space and the risk score distribution of pipeline nodes are input into a pre-trained collaborative learning model, and the model outputs a sequence of action strategies for the agent on pipeline nodes within the responsibility area. The collaborative learning model adopts a centralized training and distributed execution architecture, and is trained using digital twin simulation data of the agent's state space and the risk score distribution of pipeline nodes, specifically: A policy network for the agent is constructed, and a local reward function is built based on the local observation space inspection coverage completeness rate and the pipeline node risk score decay rate. The policy network parameters are iterated through simulation in a digital twin environment. The policy network outputs the predicted action priority values ​​of each pipeline node within the agent's inspection responsibility area based on the input risk score distribution and pipeline node topology. Based on the distribution of action priority prediction values, an initial inspection pipeline node priority sequence is generated for the agent and input into the path optimizer. The final inspection path output by the path optimizer updates the initial inspection pipeline node sequence and constructs the agent's action policy sequence. In step S4, identifying conflicting agents with resource conflicts from the action policy sequences of all agents, and updating the action policies of these conflicting agents based on global pipeline node risk scores and conflict costs, specifically includes: Analyze all agent action strategy sequences to identify conflicting agents with pipeline node monopoly, path intersection, and resource overrun conflicts. Each conflicting agent broadcasts the target pipeline node and planned arrival time corresponding to its action strategy to each other. Each conflicting agent receives the action policies of other agents and calculates the conflict cost based on the overlap rate of pipeline node resource occupancy, path intersection length, and resource load. Using the comprehensive performance index, which consists of the global pipeline node risk score decay rate and the conflict cost reduction rate, as the optimization objective, the action strategy of the conflict agent is adjusted in multiple rounds, and the action strategy corresponding to the optimal comprehensive performance index is selected to update the original action strategy sequence of the conflict agent.

2. The method for inspecting urban water supply and drainage pipelines according to claim 1, characterized in that, In step S1, receiving the real-time status data stream of the urban water supply and drainage pipeline network and constructing the state space for multi-agent collaborative inspection and scheduling specifically includes: The pipeline pressure and flow rate are collected at a fixed sampling frequency as real-time status data streams, while environmental data streams are accessed from the urban environmental monitoring center; Construct a graph model based on the pipeline network topology, where pipeline nodes represent pipeline intersections and edges represent connecting pipeline segments; Each agent is assigned a connected subgraph as its inspection responsibility area, and the set of pipeline nodes covered by the responsibility area constitutes the agent's local observation space. The real-time state data values ​​in the local observation space of each agent are arranged in chronological order to form a state vector. The elements of the state vector include the current state data value and the state data of the historical monitoring period. The state vectors of individual agents are stored in a shared memory area to form a state space matrix. The rows of the matrix correspond to the pipeline node indices in the local observation space, and the columns correspond to the state vectors.

3. The method for inspecting urban water supply and drainage pipelines according to claim 1, characterized in that, In step S2, dynamic weights are set for pipeline nodes in the initial state space of each agent. These dynamic weights are calculated based on topological dependency coefficients and anomaly indices of state data, and the risk score for pipeline nodes is calculated in conjunction with environmental data. Specifically, this includes: Dynamic weights are assigned to pipeline nodes corresponding to the state space of the agent. The initial value of the dynamic weight is set based on the number of users served by the pipeline node. The state vectors of pipeline nodes are extracted using a graph convolutional network as feature embeddings, and the topological dependency coefficients between pipeline nodes are calculated using an attention mechanism. An anomaly index is calculated based on the deviation between real-time status data and historical baselines of pipeline node status vectors. Update the dynamic weights of pipeline nodes based on the normalized result of the Hadamard product operation performed on the topology dependency coefficient and the anomaly index. Calculate the mutual information value between environmental data and historical pipeline node failure events, and use the result as the environmental factor weight; By combining dynamic weight vectors and environmental factor weights, and smoothing them through linear combination and time window processing, a pipeline node risk score is output.

4. The method for inspecting urban water supply and drainage pipelines according to claim 1, characterized in that, The path optimizer calculates the final inspection path with the objective function of minimizing the overall cost. The overall cost is calculated by weighting the travel time cost between pipeline nodes and the cost of the pipeline node risk score increasing over time.

5. The method for inspecting urban water supply and drainage pipelines according to claim 1, characterized in that, The multi-round strategy adjustment process involves adjusting the planned arrival time of conflicting agents according to a fixed time step, and recalculating the corresponding comprehensive performance index after each round of adjustment.

6. The method for inspecting urban water supply and drainage pipelines according to claim 1, characterized in that, In step S5, after the action strategy of the conflicting agents is updated, a secondary verification of resource conflicts is performed on all agent action strategy sequences until the verification is passed. The actual agent inspection and scheduling then includes: After the action strategy of the conflicting agents is updated, a second identification is performed on the conflicting agents in all agent action strategy sequences that have pipeline node monopoly, path intersection and resource over-limit conflict. The process of calculating comprehensive performance indicators and adjusting the action strategies of conflicting agents is repeated until the conflict verification of all agent action strategy sequences is passed, at which point the conflicting agent identification and action strategy adjustment process ends. Actual inspection and scheduling of agents are carried out based on the final action strategy sequence.

7. An inspection system for urban water supply and drainage pipelines, used to implement the inspection method for urban water supply and drainage pipelines as described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to receive real-time status data streams and environmental data of the urban water supply and drainage pipeline network, and to construct a multi-agent collaborative inspection status space based on topology. The risk assessment module is used to calculate the dynamic weight of pipeline nodes based on the topology dependency coefficient and the state anomaly index, and to generate pipeline node risk scores in combination with environmental factors. The collaborative decision-making module is used to call a pre-trained collaborative learning model based on the risk score of pipeline nodes to generate action strategy sequences for each agent and inspection node priorities; The path optimization module is used to plan the path of the initial inspection node sequence with the goal of minimizing the overall cost, output the final inspection path and update the action strategy. The conflict coordination module is used to identify resource conflicts such as node monopoly and path intersection in multi-agent action strategies, and generate a conflict-free strategy sequence through multiple rounds of strategy adjustment.

Images