Unmanned ship swarm cooperative monitoring task allocation method and system based on water area clustering partition and abnormal confidence bidding

Abstract

The application provides an unmanned ship swarm cooperative monitoring task allocation method and system based on water area clustering partition and abnormal confidence bidding, which divides a target river basin into multiple sub-water areas through clustering and constructs a distributed prediction model to obtain a predicted concentration and a predicted uncertainty in real time; calculates abnormal confidence and a time warming factor of each sub-water area, and generates a ship-sub-water area combined bid according to the unmanned ship sailing distance and the remaining power; solves a designation scheme under constraints with the maximum total bid as the target, and each ship executes a task according to a reinforcement learning strategy; and re-clusters online when the sub-water area centroid drifts or the confidence distribution divergence exceeds a threshold. The application integrates water quality abnormal urgency and resource state into the task allocation decision, considers abnormal priority response and monitoring blind area elimination, and improves the cooperative monitoring response speed and resource utilization efficiency.

Description

Technical Field

[0001] This application relates to the field of unmanned surface vessel (USV) water environment monitoring and multi-USV task allocation technology, and more specifically, to a method and system for collaborative monitoring task allocation of USV swarms based on water area clustering and anomaly confidence bidding. Background Technology

[0002] Currently, unmanned surface vessels (USVs) have been applied to water quality monitoring, anomaly tracing, and sampling operations in rivers and lakes. When multiple USVs collaborate on large-scale watershed monitoring tasks, the watershed is typically divided into several operational areas based on fixed grids, administrative boundaries, or human experience, and these areas are then surveyed in sections. Task allocation often employs a general bidding function or fixed priority ranking based on resource status such as travel distance and remaining battery power. These bidding functions do not establish a correlation with water quality prediction factors such as predicted concentration, prediction uncertainty, and its rate of change. When water quality anomalies occur, it is difficult to adjust the allocation order according to the urgency of the anomaly, resulting in a delayed anomaly response. Furthermore, there is a lack of priority enhancement mechanisms for areas that have not been surveyed for extended periods, easily creating monitoring blind spots. In addition, once the above zoning method is determined, it remains fixed and cannot adaptively adjust to changes in the spatial distribution of water quality anomalies. This can easily lead to a mismatch between the prediction model within the zone and the current water quality characteristics, thus affecting prediction accuracy and the rationality of task allocation.

[0003] Therefore, in the dynamic operation scenario of large watersheds, how to unify and quantify water quality prediction information and the resource status of unmanned vessels and incorporate them into the same task allocation decision framework to achieve anomaly priority response, elimination of monitoring blind spots, and adaptive updating of water area zones has become an urgent technical problem to be solved. Summary of the Invention

[0004] This application provides a method and system for allocating tasks for collaborative monitoring of unmanned vessel swarms based on water area clustering and anomaly confidence bidding. This solves the problems in existing collaborative monitoring of unmanned vessel swarms, such as the difficulty in effectively integrating and unifying dynamic environmental perception information with resource status in task allocation, and the inability of task area division to adaptively adjust with environmental changes, resulting in low response timeliness and low resource utilization efficiency.

[0005] To solve the above-mentioned technical problems, the technical solution of this application is as follows: Firstly, this application provides a method for allocating collaborative monitoring tasks for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, including: The target watershed is clustered and divided into multiple sub-water areas, and an independently trained distributed prediction model is constructed for each sub-water area; each distributed prediction model outputs the predicted concentration and prediction uncertainty of the corresponding sub-water area in real time. Based on the predicted concentration, the predicted uncertainty, and the rate of change obtained from the predicted concentration, the anomaly confidence level of each sub-water area is calculated; and based on the interval since the last service of each sub-water area, the time warming factor of each sub-water area is calculated. For each combination of an unmanned surface vessel (USV) and each sub-water area, the bid for the combination is calculated based on the anomaly confidence level, prediction uncertainty, and time warming factor of the sub-water area, combined with the travel distance of the USV to the sub-water area and the remaining battery power of the USV. With the objective of maximizing the sum of bids for all unmanned vessels and all sub-water areas, the assignment scheme is solved under preset constraints, and conflicts between multiple vessels in the same sub-water area are resolved according to the suboptimal assignment rule, thus completing the task allocation of the unmanned vessel swarm to all current sub-water areas. Each unmanned vessel obtains its assigned target sub-water area and task type according to the assignment scheme, and performs monitoring tasks according to a reinforcement learning strategy. The centroid drift of sub-water areas and the divergence of the spatial distribution of anomaly confidence are monitored in real time. When the centroid drift of the sub-water area or the divergence of the spatial distribution of anomaly confidence exceeds the corresponding threshold, online re-clustering is performed to update the water area division of the target watershed, and the distributed prediction model is returned to re-output the predicted concentration and prediction uncertainty until all sub-water areas have been served and no new anomalies trigger re-clustering.

[0006] Furthermore, the target watershed is clustered and divided into multiple sub-water areas, including: Collect historical data from all monitoring points within the target watershed and perform data cleaning. The historical data includes visual samples and non-visual samples. For the cleaned historical data, the elbow method is used to determine the optimal number of clusters, and the probability sampling method is used to determine the initial cluster centers. For each data sample in the cleaned historical data, calculate the distance from the data sample to all cluster centers, and perform soft clustering according to the correlation weight between samples to obtain the contribution weight of each data sample. The cluster centers are updated based on the contribution weight of each data sample. After iterating to a preset standard, the silhouette coefficient is used to evaluate and generate the final clustering result. The multiple sub-water areas are divided according to the final clustering result.

[0007] Furthermore, the formula for calculating the anomaly confidence level is as follows:

[0008] In the formula, This refers to the predicted concentration output by the distributed prediction model for this sub-water area. The prediction uncertainty is the output of the distributed prediction model for this sub-water area. These are the limits for water quality indicators. To predict the rate of change in concentration; This is a reference value for the rate of change; This is the first calibration coefficient; This is the second calibration coefficient.

[0009] Furthermore, the formula for calculating the time-heating factor is as follows:

[0010] In the formula, For the last time that sub-water area k was served; The required inspection cycle is specified in the regulations; This is the heating rate coefficient; This refers to the current moment.

[0011] Furthermore, the formula for calculating the bid for the combination is as follows:

[0012] In the formula, For the first The distance traveled by the ship to point k in the sub-water area; For the first The ship's remaining battery power; The first normalization standard; This is the second normalization standard; The third normalization standard; The first optimized calibration coefficient; This is the second optimized calibration coefficient; The third optimization calibration coefficient; This is the fourth optimized calibration coefficient; The fifth optimized calibration coefficient; The anomaly confidence level of sub-water area k; The time-induced warming factor; Let be the prediction uncertainty of sub-water area k.

[0013] Furthermore, the solution assignment scheme includes: In the process of solving the problem with the goal of maximizing the sum of bids for the combination of all unmanned vessels and all sub-water areas, the Pareto solution archive of candidate assignment schemes is maintained with coverage, average response time and total energy consumption as optimization objectives. Select the corresponding weight vector according to the current working mode, which includes daily inspection mode and emergency response mode; The Pareto solution file is projected using the weight vector to determine the assignment scheme for the current round from the Pareto solution file.

[0014] Furthermore, the preset constraint specifically includes: Each unmanned surface vessel (USV) can serve at most one sub-water area. Each sub-water area can be served by at most one USV in the daily patrol mode and by at most m USVs in the emergency response mode, where m is an integer greater than 1.

[0015] Furthermore, the step of performing the monitoring task according to the reinforcement learning strategy includes: A state set containing positional deviation and pollutant concentration is constructed, and a reward function is set. The reward function provides positive incentives for the behavior of moving towards the high concentration area and negative penalties for energy consumption. The source tracing strategy is trained based on the state set and reward function. The task type in the assignment scheme is used as the condition input of the source tracing strategy to drive the unmanned vessel to perform the survey, source tracing or sampling tasks corresponding to the task type.

[0016] Furthermore, the triggering condition for the online re-clustering is:

[0017] In the formula, Let be the centroid of the sub-water area k at time t; For the last partitioning time of sub-water area k The center of mass; JS divergence for the spatial distribution of outlier confidence scores; The first threshold; The second threshold; The anomaly confidence level of the current sub-water area at time t; For the current sub-water area in The confidence level of the anomalies at any given time.

[0018] Secondly, this application provides a collaborative monitoring task allocation system for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, including: The water area clustering and partitioning module is used to cluster and divide the target water area into multiple sub-water areas, and to build an independently trained distributed prediction model for each sub-water area; each of the distributed prediction models outputs the predicted concentration and prediction uncertainty of the corresponding sub-water area in real time. The anomaly confidence assessment module is used to calculate the anomaly confidence of each sub-water area based on the predicted concentration, the predicted uncertainty, and the rate of change obtained from the predicted concentration; and to calculate the time warming factor of each sub-water area based on the interval since the last service. The bid generation module is used to calculate the bid for each combination of unmanned surface vessel and each sub-water area based on the anomaly confidence level, prediction uncertainty, and time warming factor of the sub-water area, combined with the travel distance of the unmanned surface vessel to the sub-water area and the remaining battery power of the unmanned surface vessel. The allocation decision module is used to solve the assignment scheme under preset constraints with the goal of maximizing the sum of bids of all unmanned vessels and all sub-water areas, and to resolve multi-vessel conflicts in the same sub-water area according to the suboptimal assignment rules, so as to complete the task allocation of the unmanned vessel swarm to all current sub-water areas. The single-vessel execution module enables each unmanned vessel to obtain the assigned target sub-water area and task type according to the assignment scheme, and to perform monitoring tasks according to the reinforcement learning strategy. The re-partitioning trigger module is used to monitor the centroid drift of sub-water areas and the divergence of the spatial distribution of abnormal confidence in real time. When the centroid drift of the sub-water area or the divergence of the spatial distribution of abnormal confidence is greater than the corresponding threshold, online re-clustering is performed to update the water area division of the target watershed, and the distributed prediction model is returned to re-output the predicted concentration and prediction uncertainty until all sub-water areas have been served and no new abnormalities trigger re-clustering.

[0019] Compared with the prior art, the beneficial effects of the technical solution of this application are: This application divides water quality into multiple sub-water areas through clustering and independently constructs distributed prediction models, making the water quality predictions for each sub-water area targeted and spatially differentiated. It introduces anomaly confidence and time-heating factors to quantify the urgency of water quality anomalies, prediction uncertainty, and patrol coverage timeliness into comparable bid values, which are then integrated with the unmanned surface vessel's (USV) travel distance and remaining battery power for calculation. The application solves the assignment scheme with the objective of maximizing the total bid, enabling task allocation decisions to simultaneously consider anomaly priority response and monitoring blind spot elimination, improving the response timeliness and resource utilization efficiency of USV swarm collaborative monitoring. Online re-clustering achieves adaptive updates of water area partitions, ensuring that the partitions remain consistent with the current spatial distribution of water quality anomalies and guaranteeing continuous matching between the prediction model and environmental characteristics. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a collaborative monitoring task allocation method for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of a collaborative monitoring task allocation system for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives and advantages of this application clearer, the application will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining this application and are not intended to limit this application.

[0022] The prefixes such as "first" and "second" used in this application embodiment are merely for distinguishing different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. The use of ordinal numbers and other prefixes used to distinguish descriptive objects in this application embodiment does not constitute a restriction on the described objects. The description of the described objects is given in the context of the embodiments, and the use of such prefixes should not constitute unnecessary restrictions.

[0023] Preferred embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of this application and are not intended to limit the scope of protection of this application.

[0024] Please see Figure 1 This is a flowchart illustrating a method for allocating collaborative monitoring tasks for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, provided in an embodiment of this application. The method includes: Step S1: Cluster the target watershed to form multiple sub-water areas, and construct an independently trained distributed prediction model for each sub-water area; each distributed prediction model outputs the predicted concentration and prediction uncertainty of the corresponding sub-water area in real time. Step S2: Calculate the anomaly confidence level for each sub-water area based on the predicted concentration, the predicted uncertainty, and the rate of change obtained from the predicted concentration; and calculate the time warming factor for each sub-water area based on the interval since the last service. Step S3: For each combination of unmanned vessel and each sub-water area, calculate the bid for the combination based on the anomaly confidence level, prediction uncertainty and time warming factor of the sub-water area, and in combination with the travel distance of the unmanned vessel to the sub-water area and the remaining power of the unmanned vessel. Step S4: With the goal of maximizing the sum of bids of all unmanned vessels and all sub-water areas, solve the assignment scheme under preset constraints, and resolve multi-vessel conflicts in the same sub-water area according to the suboptimal assignment rule, and complete the task allocation of the unmanned vessel swarm to all current sub-water areas. Step S5: Each unmanned vessel obtains the assigned target sub-water area and task type according to the assignment scheme, and performs the monitoring task according to the reinforcement learning strategy. Step S6: Monitor the centroid drift of sub-water areas and the divergence of the spatial distribution of abnormal confidence in real time. When the centroid drift of the sub-water area or the divergence of the spatial distribution of abnormal confidence is greater than the corresponding threshold, perform online re-clustering to update the water area division of the target watershed, and return to the distributed prediction model to re-output the predicted concentration and prediction uncertainty until all sub-water areas have been served and no new anomalies trigger re-clustering.

[0025] In one embodiment, this embodiment provides a method for allocating collaborative monitoring tasks of unmanned vessel swarms based on water area clustering and anomaly confidence bidding, which is applied to the water environment monitoring scenario of a large lake, and deploys unmanned vessel swarms to perform daily patrol and emergency response tasks.

[0026] First, visual and non-visual samples from various monitoring points within the target lake were collected as historical data and then cleaned. Non-visual samples included water quality physicochemical indicators such as dissolved oxygen, permanganate index, ammonia nitrogen, total phosphorus, total nitrogen, pH, conductivity, turbidity, and water temperature, along with their corresponding sampling point coordinates and collection times. Visual samples were water surface images collected by visual sensors mounted on unmanned surface vessels, used to reflect visualized water quality characteristics such as floating debris, algal blooms, and abnormal water color.

[0027] For the cleaned historical data, the elbow method was used to analyze the relationship between clustering error and the number of clusters, determining the optimal number of clusters as N (N is an integer greater than 1). N initial cluster centers were determined using probability sampling. For each data sample in the cleaned historical data, the distance from the sample to all cluster centers was calculated, and soft clustering was performed according to the correlation weights between samples. The cluster centers were updated based on the contribution weights of each data sample. After iterating to a preset standard, the silhouette coefficient was used to evaluate and generate the final clustering result. This final clustering result divided the target lake into N sub-water areas, with the centroid coordinates of each sub-water area denoted as... , , …, .

[0028] Based on the zoning results, a distributed prediction model is constructed independently for each sub-water area. Each model is trained using only historical sample data falling within the corresponding sub-water area. In this embodiment, the distributed prediction model employs a Long Short-Term Memory (LSTM) network, taking the water quality monitoring sequence of the past 24 hours as input and outputting the predicted concentration for the next hour. With prediction uncertainty , where k = 1, 2, …, N.

[0029] Each distributed prediction model outputs the predicted concentration and prediction uncertainty for the corresponding sub-water area in real time. Taking ammonia nitrogen as an example, the water quality indicator evaluation limit... According to GB 3838 "Environmental Quality Standard for Surface Water" Class III water standard, the concentration is taken as 1.0 mg / L, and the rate of change is a reference value. Based on the empirical level of the lake's historical concentration change rate, it was taken as 0.05 mg / ( Calibration coefficient Take 1.0, Take 0.5. Based on the predicted concentration of each sub-water area at a certain moment. Prediction uncertainty and predicted concentration change rate The anomaly confidence level of each sub-water area is calculated using the following formula:

[0030] In the formula, This refers to the predicted concentration output by the distributed prediction model for this sub-water area. The prediction uncertainty is the output of the distributed prediction model for this sub-water area. These are the limits for water quality indicators. To predict the rate of change in concentration; This is a reference value for the rate of change; This is the first calibration coefficient; This is the second calibration coefficient.

[0031] The time-varying warming factor is calculated based on the interval since the last service for each sub-water area. Survey cycle. According to the specifications, a 24-hour heating period is used, and the heating rate coefficient μ is taken as 1.0. The calculation formula is as follows:

[0032] In the formula, For the last time that sub-water area k was served; The required inspection cycle is specified in the regulations; This is the heating rate coefficient; This refers to the current moment.

[0033] After calculating for each sub-water area separately, the anomaly confidence vector for each sub-water area is obtained. , , …, ] and time-heating factor vector[ , , …, ].

[0034] M unmanned vessels are designated as vessel 1 to vessel M, and the travel distance of each vessel to the representative point in each sub-water area is... Remaining battery power is obtained in real time by the ship's onboard positioning system. Reported by the battery management system. In this embodiment, the normalized reference value is: navigation distance reference. Taking a characteristic length of 10km for the watershed, the prediction uncertainty benchmark is... Taking the historical average uncertainty of 0.25 mg / L, based on the electricity consumption standard. Charge to 100%. Calibration coefficient. to Calibrated by simulation experiments.

[0035] For each unmanned ship ( =1, 2, …, M) and each sub-water area ( The combination consisting of (=1, 2, …, N) is used to determine the anomaly confidence level of the sub-water area. Prediction uncertainty and time-induced warming factor And the distance the unmanned vessel traveled to that sub-water area. and the remaining battery power of the unmanned vessel Calculate the bid for this combination using the following formula:

[0036] In the formula, For the first The distance traveled by the ship to point k in the sub-water area; For the first The ship's remaining battery power; The first normalization standard; This is the second normalization standard; The third normalization standard; The first optimized calibration coefficient; This is the second optimized calibration coefficient; The third optimization calibration coefficient; This is the fourth optimized calibration coefficient; The fifth optimized calibration coefficient; The anomaly confidence level of sub-water area k; The time-induced warming factor; Let be the prediction uncertainty of sub-water area k.

[0037] This results in an M x N bid matrix. The objective is to maximize the sum of all M x N bid combinations, and the assignment scheme is solved under preset constraints. The preset constraints are: each unmanned surface vessel (USV) serves at most one sub-water area; in routine patrol mode, each sub-water area is served by at most one USV; in emergency response mode, each sub-water area is served by at most m USVs, where m is an integer greater than 1 and m ≤ M.

[0038] When M≥N, one round of allocation can cover all sub-water areas, and surplus unmanned vessels enter standby mode; when M<N, a single round can only cover part of the sub-water areas, and the time warming factor of the unserved sub-water areas continues to increase with the duration of loss to follow-up. In subsequent rounds, the bid gradually increases until the winning bid is obtained, ensuring that any sub-water area will eventually be served and no monitoring blind spots are formed.

[0039] The assignment scheme is solved using either the Hungarian algorithm or a greedy sequential auction. When using a greedy sequential auction, the highest-bid combination is awarded round by round, and the already awarded unmanned vessels and sub-areas are removed from the candidate set until allocation is complete. When multiple vessels conflict within the same sub-area, the second-highest bid vessel is assigned to its globally second-best sub-area according to the suboptimal assignment rules, ensuring convergence within one round. The entire bidding process is documented and audited, creating replayable decision audit data.

[0040] During the solution process, a Pareto solution archive is maintained, with coverage, average response time, and total energy consumption as optimization objectives. When in routine survey mode, a weight vector prioritizing coverage is projected onto the Pareto solution archive to determine the assignment scheme for the current round; if switching to emergency response mode, a weight vector prioritizing response time is projected onto the solution archive to determine the allocation scheme.

[0041] Furthermore, Pareto solution archives are maintained with coverage, average response time, and total energy consumption as optimization objectives, so that the allocation scheme can achieve a balance among multiple conflicting objectives; corresponding weight projections are selected according to daily inspection or emergency response modes, so that the same assignment framework can adapt to different task modes and improve the applicability of the system in multiple scenarios.

[0042] The target sub-water areas and mission types assigned to each unmanned surface vessel (USV) in the assignment plan are distributed to the corresponding USVs. USVs assigned to sub-water areas with high anomaly confidence receive source tracing missions, while USVs assigned to sub-water areas with low anomaly confidence but high time warming factor receive survey missions. Sampling missions may also be distributed if necessary.

[0043] Each unmanned surface vessel (USV) employs a reinforcement learning strategy when performing monitoring tasks. A model is constructed that includes positional bias (Δx, Δy) and pollutant concentration. The set of states, and the reward function R is defined as:

[0044] In the formula, The weight of the first reward item; The weight of the second reward item; The weight of the third reward item; The weight of the fourth reward item; The power of the unmanned vessel; The travel time of the unmanned vessel; The Euclidean distance between the unmanned vessel and the pollution source; The pollutant concentration at the current location of the unmanned vessel; This is the conversion factor for the concentration benefit term; the reward function provides a positive incentive for the behavior of moving towards high concentration areas and a negative penalty for energy consumption and time consumption.

[0045] Based on the state set and reward function, a source tracing strategy is trained. The task type in the assigned scheme is used as the conditional input of the policy network, so that the same policy network can output the corresponding behavior strategy according to different task types such as surveying, source tracing or sampling, and drive each unmanned vessel to complete the designated task.

[0046] Furthermore, a state set containing positional deviation and pollutant concentration is constructed, and a reward function with positive incentive for moving towards high concentration areas and negative penalty for energy consumption drives the source tracing strategy, enabling unmanned vessels to efficiently approach and locate pollution sources under limited energy constraints. The task type is used as a conditional input to the strategy network, allowing the same strategy network to perform survey, source tracing, or sampling tasks according to different target conditions, thereby improving the flexibility and reusability of single-vessel control.

[0047] Real-time monitoring of the divergence between centroid drift and the spatial distribution of anomaly confidence in sub-water areas. Centroid drift threshold. Take 15% of the watershed characteristic length, JS divergence threshold Take a value of 0.3 (base 2). When the average centroid drift of each sub-water area exceeds the threshold:

[0048] Or the JS divergence of the abnormal confidence spatial distribution exceeds the threshold:

[0049] In the formula, Let be the centroid of the sub-water area k at time t; For the last partitioning time of sub-water area k The center of mass; JS divergence for the spatial distribution of outlier confidence scores; The first threshold; The second threshold; The anomaly confidence level of the current sub-water area at time t; For the current sub-water area in The confidence level of the anomalies at any given time.

[0050] During online re-clustering, the original prediction model is directly used for sub-water areas whose boundaries remain unchanged. For sub-water areas that have changed due to splitting or merging, the parameters of the original model with the greatest spatial overlap with the new sub-water area are selected as the initial values ​​for warm-up. Historical samples falling within the new sub-water area are used for fine-tuning training, rather than directly weighting and averaging the parameters of multiple independently trained models, to ensure that the learned water quality characterization is not destroyed. After re-clustering is completed, the distributed prediction model re-outputs the predicted concentration and prediction uncertainty, entering a new round of task allocation loop until all sub-water areas have been served and no new anomalies trigger re-clustering.

[0051] The system summarizes the mission completion status, actual navigation trajectory, collected water quality data, and energy consumption statistics of each vessel. It also writes the entire process of bidding details, assignment results, conflict resolution records, and re-clustering trigger reasons into the decision log, forming replayable audit data. This data can be visualized, historical data queried, and monitoring reports generated through remote terminals.

[0052] The upper-level bidding allocation operates on a minute-level cycle and is triggered immediately to supplement allocation when new anomalies occur, vessels malfunction or return to port due to low battery, or mission completion. The lower-level single-vessel control operates on a second-level cycle. The upper and lower levels are decoupled by time scale, allowing for a smooth expansion from a single vessel to a cluster of multiple unmanned vessels without altering the single-vessel control strategy.

[0053] Please see Figure 2 This is a schematic diagram of a collaborative monitoring task allocation system for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, provided in an embodiment of this application. This embodiment of the application provides a collaborative monitoring task allocation system for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, used to implement the method for collaborative monitoring task allocation for unmanned vessel swarms provided in the above embodiment, including: The water area clustering and partitioning module is used to cluster and divide the target water area into multiple sub-water areas, and to build an independently trained distributed prediction model for each sub-water area; each of the distributed prediction models outputs the predicted concentration and prediction uncertainty of the corresponding sub-water area in real time. The anomaly confidence assessment module is used to calculate the anomaly confidence of each sub-water area based on the predicted concentration, the predicted uncertainty, and the rate of change obtained from the predicted concentration; and to calculate the time warming factor of each sub-water area based on the interval since the last service. The bid generation module is used to calculate the bid for each combination of unmanned surface vessel and each sub-water area based on the anomaly confidence level, prediction uncertainty, and time warming factor of the sub-water area, combined with the travel distance of the unmanned surface vessel to the sub-water area and the remaining battery power of the unmanned surface vessel. The allocation decision module is used to solve the assignment scheme under preset constraints with the goal of maximizing the sum of bids of all unmanned vessels and all sub-water areas, and to resolve multi-vessel conflicts in the same sub-water area according to the suboptimal assignment rules, so as to complete the task allocation of the unmanned vessel swarm to all current sub-water areas. The single-vessel execution module enables each unmanned vessel to obtain the assigned target sub-water area and task type according to the assignment scheme, and to perform monitoring tasks according to the reinforcement learning strategy. The re-partitioning trigger module is used to monitor the centroid drift of sub-water areas and the divergence of the spatial distribution of abnormal confidence in real time. When the centroid drift of the sub-water area or the divergence of the spatial distribution of abnormal confidence is greater than the corresponding threshold, online re-clustering is performed to update the water area division of the target watershed, and the distributed prediction model is returned to re-output the predicted concentration and prediction uncertainty until all sub-water areas have been served and no new abnormalities trigger re-clustering.

[0054] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

[0055] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope 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.

Claims

1. A method for allocating collaborative monitoring tasks for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, characterized in that, include: The target watershed is clustered and divided into multiple sub-water areas, and an independently trained distributed prediction model is constructed for each sub-water area; each distributed prediction model outputs the predicted concentration and prediction uncertainty of the corresponding sub-water area in real time. Based on the predicted concentration, the predicted uncertainty, and the rate of change obtained from the predicted concentration, the anomaly confidence level of each sub-water area is calculated; and based on the interval since the last service of each sub-water area, the time warming factor of each sub-water area is calculated. For each combination of an unmanned surface vessel (USV) and each sub-water area, the bid for the combination is calculated based on the anomaly confidence level, prediction uncertainty, and time warming factor of the sub-water area, combined with the travel distance of the USV to the sub-water area and the remaining battery power of the USV. With the objective of maximizing the sum of bids for all unmanned vessels and all sub-water areas, the assignment scheme is solved under preset constraints, and conflicts between multiple vessels in the same sub-water area are resolved according to the suboptimal assignment rule, thus completing the task allocation of the unmanned vessel swarm to all current sub-water areas. Each unmanned vessel obtains its assigned target sub-water area and task type according to the assignment scheme, and performs monitoring tasks according to a reinforcement learning strategy. The centroid drift of sub-water areas and the divergence of the spatial distribution of anomaly confidence are monitored in real time. When the centroid drift of the sub-water area or the divergence of the spatial distribution of anomaly confidence exceeds the corresponding threshold, online re-clustering is performed to update the water area division of the target watershed, and the distributed prediction model is returned to re-output the predicted concentration and prediction uncertainty until all sub-water areas have been served and no new anomalies trigger re-clustering.

2. The method according to claim 1, characterized in that, The target watershed is clustered and divided into multiple sub-water areas, including: Collect historical data from all monitoring points within the target watershed and perform data cleaning. The historical data includes visual samples and non-visual samples. For the cleaned historical data, the elbow method is used to determine the optimal number of clusters, and the probability sampling method is used to determine the initial cluster centers. For each data sample in the cleaned historical data, calculate the distance from the data sample to all cluster centers, and perform soft clustering according to the correlation weight between samples to obtain the contribution weight of each data sample. The cluster centers are updated based on the contribution weight of each data sample. After iterating to a preset standard, the silhouette coefficient is used to evaluate and generate the final clustering result. The multiple sub-water areas are divided according to the final clustering result.

3. The method according to claim 1, characterized in that, The formula for calculating the anomaly confidence level is as follows: In the formula, This refers to the predicted concentration output by the distributed prediction model for this sub-water area. The prediction uncertainty is the output of the distributed prediction model for this sub-water area. These are the limits for water quality indicators. To predict the rate of change in concentration; This is a reference value for the rate of change; This is the first calibration coefficient; This is the second calibration coefficient.

4. The method according to claim 1, characterized in that, The formula for calculating the time-heating factor is: In the formula, For the last time that sub-water area k was served; The required inspection cycle is specified in the regulations; This is the heating rate coefficient; This refers to the current moment.

5. The method according to claim 1, characterized in that, The formula for calculating the bid for the combination is: In the formula, For the first The distance traveled by the ship to point k in the sub-water area; For the first The ship's remaining battery power; The first normalization standard; This is the second normalization standard; The third normalization standard; The first optimized calibration coefficient; This is the second optimized calibration coefficient; The third optimization calibration coefficient; This is the fourth optimized calibration coefficient; The fifth optimized calibration coefficient; The anomaly confidence level of sub-water area k; The time-induced warming factor; Let be the prediction uncertainty of sub-water area k.

6. The method according to claim 1, characterized in that, The solution assignment scheme includes: In the process of solving the problem with the goal of maximizing the sum of bids for the combination of all unmanned vessels and all sub-water areas, the Pareto solution archive of candidate assignment schemes is maintained with coverage, average response time and total energy consumption as optimization objectives. Select the corresponding weight vector according to the current working mode, which includes daily inspection mode and emergency response mode; The Pareto solution file is projected using the weight vector to determine the assignment scheme for the current round from the Pareto solution file.

7. The method according to claim 6, characterized in that, The preset constraints are specifically as follows: Each unmanned surface vessel (USV) can serve at most one sub-water area. Each sub-water area can be served by at most one USV in the daily patrol mode and by at most m USVs in the emergency response mode, where m is an integer greater than 1.

8. The method according to claim 1, characterized in that, The step of performing the monitoring task according to the reinforcement learning strategy includes: A state set containing positional deviation and pollutant concentration is constructed, and a reward function is set. The reward function provides positive incentives for the behavior of moving towards the high concentration area and negative penalties for energy consumption. The source tracing strategy is trained based on the state set and reward function. The task type in the assignment scheme is used as the condition input of the source tracing strategy to drive the unmanned vessel to perform the survey, source tracing or sampling tasks corresponding to the task type.

9. The method according to claim 1, characterized in that, The triggering condition for the online re-clustering is: In the formula, Let be the centroid of the sub-water area k at time t; For the last partitioning time of sub-water area k The center of mass; JS divergence for the spatial distribution of outlier confidence scores; The first threshold; The second threshold; The anomaly confidence level of the current sub-water area at time t; For the current sub-water area in The confidence level of the anomalies at any given time.

10. A collaborative monitoring task allocation system for unmanned vessel swarms based on water area clustering and anomaly confidence bidding, characterized in that, include: The water area clustering and partitioning module is used to cluster and divide the target water area into multiple sub-water areas, and to build an independently trained distributed prediction model for each sub-water area; each of the distributed prediction models outputs the predicted concentration and prediction uncertainty of the corresponding sub-water area in real time. The anomaly confidence assessment module is used to calculate the anomaly confidence of each sub-water area based on the predicted concentration, the predicted uncertainty, and the rate of change obtained from the predicted concentration; and to calculate the time warming factor of each sub-water area based on the interval since the last service. The bid generation module is used to calculate the bid for each combination of unmanned surface vessel and each sub-water area based on the anomaly confidence level, prediction uncertainty, and time warming factor of the sub-water area, combined with the travel distance of the unmanned surface vessel to the sub-water area and the remaining battery power of the unmanned surface vessel. The allocation decision module is used to solve the assignment scheme under preset constraints with the goal of maximizing the sum of bids of all unmanned vessels and all sub-water areas, and to resolve multi-vessel conflicts in the same sub-water area according to the suboptimal assignment rules, so as to complete the task allocation of the unmanned vessel swarm to all current sub-water areas. The single-vessel execution module enables each unmanned vessel to obtain the assigned target sub-water area and task type according to the assignment scheme, and to perform monitoring tasks according to the reinforcement learning strategy. The re-partitioning trigger module is used to monitor the centroid drift of sub-water areas and the divergence of the spatial distribution of abnormal confidence in real time. When the centroid drift of the sub-water area or the divergence of the spatial distribution of abnormal confidence is greater than the corresponding threshold, online re-clustering is performed to update the water area division of the target watershed, and the distributed prediction model is returned to re-output the predicted concentration and prediction uncertainty until all sub-water areas have been served and no new abnormalities trigger re-clustering.

Images