Basin water environment intelligent biological treatment robot system and customized feeding method thereof

Abstract

The present application belongs to the field of automatic control technology, and particularly relates to a basin water environment intelligent biological treatment robot system and a customized adding method thereof, comprising the following steps: step S1: a cloud control platform fuses space-time water quality data collected by fixed monitoring stations and robots to form a water quality field topology mutation event set; step S2: the cloud control platform calculates water quality risk scores of each water area unit according to the water quality field topology mutation event and static water quality indexes, and sets a robot target density to obtain a converged robot density distribution and a water quality treatment priority level; step S3: the cloud control platform generates sampling, adding and inspection tasks according to the converged robot density distribution and the water quality treatment priority level, and determines a task basic value, and distributes the generated tasks to the winning robots for execution. The present application can timely capture key topology characteristics when the pollution structure changes, so that the robot colony can quickly concentrate in the high-risk area, and realize spatial optimization of treatment resources.

Description

Technical Field

[0001] This invention belongs to the field of automatic control technology, specifically relating to an intelligent biological treatment robot system for watershed water environment and its customized dosing method. Background Technology

[0002] In the field of watershed water environment management, the comprehensive treatment of non-point source pollution, resuspended pollution caused by sediment disturbance, and sudden high-intensity pollution has always relied on a combination of various monitoring and treatment methods. Traditional monitoring systems mainly consist of fixed water quality monitoring stations, manual sampling, and a small number of mobile monitoring platforms, providing basic water quality indicators such as ammonia nitrogen, total phosphorus, dissolved oxygen, turbidity, and temperature. However, due to the limited number and sparse spatial distribution of fixed monitoring stations, they cannot accurately reflect the transient diffusion behavior of pollution at the watershed scale. In scenarios where pollution distribution exhibits significant spatial heterogeneity and rapid temporal changes, traditional monitoring methods struggle to capture phenomena such as the formation, movement, and fragmentation of pollution masses, leading to a lag in the assessment of pollution trends.

[0003] In identifying pollution diffusion patterns, existing technologies typically employ concentration threshold determination methods based on regular grids. These methods set multiple concentration thresholds, classifying areas exceeding the thresholds as polluted. However, such methods often fail to robustly identify the topological structure of pollution clumps in complex watersheds, particularly when clumps are rapidly splitting, merging, or undergoing morphological changes, leading to potential false positives or false negatives in monitoring systems. Furthermore, traditional methods often rely on single thresholds or simple connectivity analysis, failing to adequately reflect the continuity of pollution clumps across different intensity ranges and accurately describe the lifecycle of pollution structures, resulting in a lack of targeted remediation strategies. Summary of the Invention

[0004] Therefore, the main objective of this invention is to provide an intelligent biological treatment robot system for watershed water environment and its customized dosing method. This system constructs a pollution intensity field by integrating spatiotemporal water quality data, identifies topological abrupt changes such as the generation and splitting of pollution clumps using topological data analysis, sets a target robot density based on water quality risk, and achieves automatic convergence of robot density through mean-field game theory. Furthermore, it combines market auction methods with constraints on remaining power and matching biological agents to complete the allocation of sampling, dosing, and inspection tasks, forming an integrated closed-loop system for dynamic monitoring, intelligent planning, and precise operation in watershed water environment treatment. This system can promptly capture key topological features when the pollution structure changes, enabling the robot swarm to quickly concentrate in high-risk areas, achieving spatial optimization of treatment resources. Through a task-value-driven competition mechanism, it significantly improves the efficiency and accuracy of task execution, thereby achieving superior monitoring accuracy, response speed, and treatment effects compared to traditional methods in complex watershed scenarios.

[0005] The technical solution adopted in this invention is as follows:

[0006] A watershed water environment intelligent biological treatment robot system includes: several fixed water quality monitoring stations and mobile monitoring units mounted on the robots for collecting water quality data and its temporal and spatial locations within the watershed; several robots for moving within the watershed to perform sampling, dosing, and inspection tasks; and a cloud control platform including a memory and a processor. The memory stores a program that can run on the processor, which is configured to: construct a watershed grid pollution intensity field based on the spatiotemporal water quality data collected by the fixed water quality monitoring stations and the robots using distance attenuation weighted interpolation; extract topological features through multi-level threshold scanning and active grid connectivity analysis; identify pollution clusters based on the lifecycle of the topological features within continuous threshold intervals; and determine the pollution intensity field based on the spatial overlap of pollution clusters in adjacent time layers. The generation and splitting of pollutant clusters form a set of topological mutation events in the water quality field. Based on these events and static water quality indicators, the water quality risk score for each water unit is calculated, and a target robot density is set. A single-unit revenue function is constructed, incorporating water quality risk, density deviation penalty, and time and energy consumption terms. Multiple rounds of mean-field game iteration are performed, and the density penalty coefficient is adaptively adjusted to converge the actual robot density to the target density, resulting in a converged robot density distribution and a water quality treatment priority level. Based on this converged robot density distribution and water quality treatment priority level, sampling, dosing, and inspection tasks are generated, and their basic value is determined. Each robot is guided to calculate its bidding score under the constraints of remaining power and matching biological agents. A market auction is conducted according to the principle of maximizing the bidding score, and the generated tasks are assigned to the winning robot for execution.

[0007] A smart biological treatment robot system for watershed water environment and its customized dosing method include the following steps:

[0008] Step S1: The cloud control platform integrates spatiotemporal water quality data collected by fixed monitoring stations and robots, constructs a watershed grid pollution intensity field using distance attenuation weighted interpolation, extracts topological features through multi-level threshold scanning and active grid connectivity analysis, identifies pollution clumps based on the life cycle of the topological features in continuous threshold intervals, and determines pollution clumping generation events and splitting events according to the spatial overlap relationship of pollution clumps in adjacent time layers, forming a set of water quality field topological mutation events;

[0009] Step S2: The cloud control platform calculates the water quality risk score of each water unit based on the water quality field topological mutation event and static water quality indicators, sets the target robot density, constructs a single-unit benefit function that includes water quality risk, density deviation penalty, and time and energy consumption terms, executes multiple rounds of mean field game iteration and adaptively adjusts the density penalty coefficient to make the actual robot density converge to the robot target density, and obtains the converged robot density distribution and water quality treatment priority level.

[0010] Step S3: The cloud control platform generates sampling, dosing, and inspection tasks based on the converged robot density distribution and water quality treatment priority level, and determines the basic value of the tasks. Each robot calculates its bidding score under the constraints of remaining power and matching biological agents. The cloud control platform conducts a market auction according to the principle of maximizing the bidding score and assigns the generated tasks to the winning robot for execution.

[0011] Furthermore, the construction of the watershed grid pollution intensity field in step S1 includes: setting up several fixed water quality monitoring stations along the main river channel and tributaries at all levels, and configuring mobile monitoring units on robots to collect data on ammonia nitrogen, total phosphorus, dissolved oxygen, turbidity, and water temperature, as well as their temporal and spatial locations, at preset sampling intervals; after two-dimensional projection of the watershed in the cloud control platform, dividing it into multiple watershed grid units according to a given resolution; summarizing all original water quality samples within the time window in each time layer; searching for a set of neighboring samples within a limited distance for each watershed grid unit; calculating the inverse proportional weights based on Euclidean distance and standardizing them; and performing a weighted average of the pollution indices of the samples to obtain the grid pollution intensity value; the pollution index is the weighted sum of ammonia nitrogen concentration and total phosphorus concentration.

[0012] Furthermore, the topological feature extraction and lifecycle determination in step S1 include: finding the minimum and maximum pollution intensity values ​​of all watershed grid cells in the same time layer, and dividing the numerical range into multiple pollution intensity thresholds; marking grid cells with pollution intensity not lower than the current threshold as active grids under each threshold, and dividing connected regions based on the four-adjacency relationship of shared common boundaries; using the connected regions divided by the lowest threshold as the initial topological feature, and associating them under subsequent thresholds according to the grid overlap ratio between the current connected region and the region corresponding to the topological feature of the previous threshold. When the overlap ratio is not less than a preset ratio threshold, it is considered that the same topological feature exists continuously; otherwise, a new topological feature is generated, and a stopping threshold is recorded when the overlap ratio between the current connected region and all connected regions is lower than the preset ratio threshold. The lifecycle length is determined by the difference between the birth threshold and the stopping threshold.

[0013] Furthermore, step S1, which involves identifying contaminant clumps and generating a set of water quality field topological mutation events, includes: marking topological features whose lifecycle length reaches a preset length threshold and whose number of active grids is not less than a preset number threshold under any threshold as contaminant clumps, and using the coverage area of ​​the active grids under the corresponding threshold at the midpoint of the lifecycle as the spatial range of the contaminant clump; calculating the intersection-union ratio (IUR) of the spatial ranges of contaminant clumps in adjacent time layers on multiple consecutive time layers; when the IUR of a certain contaminant clump in the current time layer with all contaminant clumps in the previous time layer is lower than a preset lower limit, recording the location of the current contaminant clump as a contaminant clump generation event; when the IUR of a certain contaminant clump in the previous time layer with at least two contaminant clumps in the current time layer is not lower than a preset upper limit, recording the contaminant clump corresponding to the previous time layer as a contaminant clump splitting event and associating it with the spatial range of subsequent contaminant clumps, thereby forming a set of water quality field topological mutation events.

[0014] Furthermore, step S2, which involves calculating the water quality risk score and setting the robot target density, includes: dividing the main river channel and tributaries of the basin into multiple water area units along the river centerline, with each water area unit corresponding to a section of river length and associated with a set of corresponding covered basin grid units; calculating the average pollution intensity, pollution cluster coverage ratio, and the number of pollution cluster generation events and splitting events within a preset time window for each water area unit, and obtaining the water quality risk score by weighting and summing the average pollution intensity, pollution cluster coverage ratio, and event statistics; and then obtaining the robot target density by superimposing the basic density constant through a linear relationship based on the water quality risk score, with the robot target density measured in units of robots per square kilometer.

[0015] Furthermore, the initialization of the mean-field game density iteration algorithm in step S2 includes: collecting the current location information and remaining power of all robots; assigning each robot to the corresponding water area unit according to its location on the river centerline; counting the number of robots in the water area unit and dividing by the water area estimated based on the river length and average river width to obtain the initial actual robot density; and performing mean-field game iteration in a preset number of rounds. In each round, a candidate water area unit set consisting of its current water area unit and adjacent upstream and downstream water areas is generated for each robot, and the estimated travel time and energy consumption are calculated based on the centerline distance between the center points of the water area units and the preset cruising speed.

[0016] Furthermore, step S2, migration probability update and density adjustment, includes: for each robot and its candidate water area, calculating the individual benefit, which is the sum of the water quality risk score of the water area minus the density penalty term and the travel time and energy consumption term. The density penalty term is obtained by multiplying the difference between the current actual robot density and the robot's target density by the density penalty coefficient; the individual benefits of the same robot in the candidate water area are translated and normalized to obtain the migration probability vector, and the expected number of robots and the expected robot density of each water area are calculated accordingly, and the actual robot density is updated with the expected robot density; at the same time, the density penalty coefficient is adjusted in stages according to the density error range between the actual robot density and the robot's target density of each water area until the preset number of iterations is completed, and the final actual robot density is used as the converged robot density distribution, and each water area is divided into at least three water quality treatment priority levels according to the water quality risk score threshold.

[0017] Furthermore, step S3, task generation and basic value determination, includes: the cloud control platform generating a task set for each water area unit based on the converged robot density distribution table. The task set includes at least three types of tasks: sampling tasks, dosing tasks, and inspection tasks. Sampling tasks consist of several sampling points, each recording the sampling location coordinates and sampling time window; dosing tasks consist of several dosing points, each recording the dosing location coordinates, dosing time window, and biological agent type number; inspection tasks consist of several inspection paths, each represented by a series of continuous coordinate points and an inspection time window; and basic values ​​are set for tasks according to the water quality treatment priority level and task type of the water area unit. In high-level water quality treatment units, dosing tasks are assigned a higher basic value than sampling and inspection tasks.

[0018] Furthermore, step S3, the calculation and auction allocation of robot bidding scores, includes: the cloud control platform calculating travel time, time cost, and energy cost for each robot and each task combination within the current water unit based on the distance between the robot's current position and the task target position, and setting a resource coefficient based on the robot's remaining battery percentage; for delivery tasks, determining whether the type of biological agent carried by the robot matches the task requirements, amplifying the task's basic value by the first coefficient if they match, and reducing the task's basic value by the second coefficient if they do not match; for sampling and inspection tasks, the basic value after matching is equal to the task's basic value; based on this, the basic value after matching is multiplied by the resource coefficient and the time is subtracted. Cost and energy costs are used to obtain bidding scores, and bid scores less than zero are truncated to zero. The cloud control platform sorts tasks according to their basic value within the same water unit, and selects the top-ranked tasks as the current auction tasks. Among the robots that have not yet won the bid and whose bid scores are greater than zero, the robot with the highest bid score is selected as the winning robot. When there are multiple robots with the same bid score, the one with the smallest estimated travel time is selected as the winning robot. The current task is then bound to the winning robot until all tasks are assigned or all robots have acquired tasks. A task allocation result table for the entire watershed is obtained, which is used by the robot's local path planning module to generate task execution paths and complete sampling, dosing, and inspection operations.

[0019] By adopting the above technical solutions, this invention achieves the following beneficial effects: By systematically combining water quality field topological mutation identification, robot target density setting based on water quality risk, mean field game density convergence, and task allocation based on market auction methods, this invention significantly improves watershed water environment management in terms of monitoring accuracy, information response speed, robot swarm collaboration efficiency, and the precision of biological agent application. Compared to the traditional model relying on fixed monitoring stations and manual sampling, this invention constructs a pollution intensity field with higher spatiotemporal resolution through distance attenuation weighted interpolation and uses topological data analysis methods to identify the generation and splitting of pollution clumps. This allows the system to proactively grasp the continuous changing structure of pollution at different threshold scales, thereby forming a stable and dynamically updated pollution structure representation at the watershed scale. Based on this representation, this invention refines the differentiation of water area units through water quality risk scores, making the robot target density more consistent with the dynamic distribution characteristics of pollution and avoiding the problems of excessive robot concentration or prolonged absence. In terms of robot spatial configuration, this invention employs a mean-field game-theoretic method, enabling robots to autonomously adjust their movement strategies. Through density penalties and energy consumption constraints, overall density convergence is gradually achieved, reducing the computational burden of global scheduling and improving the response speed of the robot swarm to sudden pollution spread. In the task execution phase, this invention sets basic task values ​​for sampling, application, and inspection tasks according to water quality treatment priority and task type. The most suitable robot is selected for task execution through a market auction, ensuring that the execution plan prioritizes high-value tasks while considering remaining robot power, biological agent matching, and path costs, achieving efficient allocation of treatment resources. The closed loop formed by this invention—monitoring and identification, density planning, task allocation, and execution—ensures that each round of robot decision-making is based on the latest pollution topology and density change trends, significantly improving the system's adaptability to complex watershed pollution behavior and ultimately achieving a comprehensive improvement in treatment efficiency and effectiveness. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the spatial distribution of pollution intensity field in a watershed grid provided in an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram illustrating the distance attenuation weighted interpolation principle provided in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of active grid connectivity analysis under multi-level threshold scanning provided in an embodiment of the present invention. Detailed Implementation

[0023] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0024] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0025] A smart biological treatment robot system for watershed water environment and its customized dosing method include the following steps:

[0026] Step S1: The cloud control platform integrates spatiotemporal water quality data collected by fixed monitoring stations and robots, constructs a watershed grid pollution intensity field using distance attenuation weighted interpolation, extracts topological features through multi-level threshold scanning and active grid connectivity analysis, identifies pollution clumps based on the life cycle of the topological features in continuous threshold intervals, and determines pollution clumping generation events and splitting events according to the spatial overlap relationship of pollution clumps in adjacent time layers, forming a set of water quality field topological mutation events.

[0027] In one specific implementation, fixed water quality monitoring stations are deployed along the main channel and tributaries of various levels in the target watershed. For example, in a typical medium-scale watershed, no fewer than 30 fixed water quality monitoring stations can be deployed. Each fixed water quality monitoring station is fixed to the riverbank or bridge structure, records its planar coordinates in a unified projected coordinate system, and continuously measures ammonia nitrogen concentration, total phosphorus concentration, dissolved oxygen, turbidity, and water temperature at 5-minute sampling intervals, while also recording the sampling time. Multiple robots move and cruise within the watershed, each equipped with a mobile water quality monitoring unit. This mobile monitoring unit collects the same types of indicators as the fixed water quality monitoring stations, with a sampling interval that can be set to 2 minutes to provide higher resolution on the time scale. All measurement data is uploaded to a cloud control platform via wireless network or wired transmission and stored according to timestamps and spatial coordinates.

[0028] Before constructing the watershed grid pollution intensity field, the cloud-based control platform first projects the watershed in two dimensions under a unified coordinate system and determines the grid resolution. In a preferred embodiment, the watershed area can be divided into several watershed grid units with a resolution of 50 meters by 50 meters. Each watershed grid unit records the planar coordinates of its center point. This grid size is sufficient to refine local variations within the river channel width while avoiding an excessively large number of grids that would lead to excessive computational load for subsequent topology analysis. Those skilled in the art can adjust the grid resolution to different values, such as 25 meters by 25 meters or 100 meters by 100 meters, depending on the watershed area, river channel width, and computational resources, while maintaining a consistent overall processing flow.

[0029] like Figure 1The figure shows a schematic diagram of the spatial distribution of the pollution intensity field in a watershed grid. This diagram illustrates how the cloud-based control platform in this invention establishes a gridded pollution intensity field in the target watershed and achieves the collection and fusion of spatiotemporal water quality data through fixed monitoring stations and mobile robots.

[0030] In this embodiment, the target watershed is divided into several watershed grid cells, with a grid resolution of 50 meters by 50 meters. In the figure, the horizontal axis represents the X-coordinate in meters, ranging from 0 to 1500 meters; the vertical axis represents the Y-coordinate in meters, ranging from 0 to 1000 meters. The entire watershed coverage area is divided into 600 watershed grid cells (30 columns, 20 rows), with each grid cell recording the planar coordinates of its center point.

[0031] In the figure, different shades of gray represent different pollution intensity values. The darker the grid color, the higher the pollution intensity value of that grid cell. As shown in the legend, the pollution intensity values ​​range from 0 to 200, with grid cells of pollution intensity 200 appearing as black, 150 as dark gray, 100 as medium gray, 50 as light gray, and 0 as white. This gray-scale gradient visually reflects the spatial distribution characteristics of pollution intensity within the watershed.

[0032] The pollution distribution along the river's morphology can be observed in the figure. Pollution intensity values ​​are relatively high along the river's centerline, forming a distinct high-value zone, while the pollution intensity values ​​gradually decrease in areas farther from the river channel. Furthermore, three distinct high-value core areas of pollution clusters have formed within the basin, located near coordinates (X≈750m, Y≈750m), (X≈400m, Y≈400m), and (X≈1100m, Y≈200m), respectively. The pollution intensity values ​​in these areas reach between 180 and 200, indicating the presence of localized high-pollution sources or pollutant accumulation.

[0033] The location of seven fixed water quality monitoring stations is marked on the map using solid black dots within square frames. These stations are deployed along the main channel and tributaries of the target watershed, for example, at coordinates (X≈350m, Y≈850m), (X≈750m, Y≈700m), (X≈1000m, Y≈500m), (X≈500m, Y≈300m), (X≈1250m, Y≈100m), (X≈250m, Y≈600m), and (X≈900m, Y≈200m). Each station is fixed to the riverbank or bridge structure and continuously measures ammonia nitrogen concentration, total phosphorus concentration, dissolved oxygen, turbidity, and water temperature at 5-minute sampling intervals, recording the sampling time. The spatial layout of the fixed monitoring stations is optimized to cover both high-pollution areas and background areas, providing stable and reliable basic data for constructing a grid-based pollution intensity field.

[0034] The map also marks the locations of five mobile robots, indicated by large solid black dots. These robots are located at coordinates (X≈600m, Y≈650m), (X≈400m, Y≈450m), (X≈1100m, Y≈250m), (X≈900m, Y≈750m), and (X≈300m, Y≈350m). Each robot is equipped with a mobile water quality monitoring unit, which collects the same types of indicators as the fixed water quality monitoring stations, with a sampling interval of 2 minutes, thus providing higher resolution water quality data on a time scale. The robots can roam freely within the watershed, flexibly reaching areas that fixed monitoring stations cannot cover, compensating for the spatial coverage limitations of fixed stations.

[0035] For any target time layer, the cloud-based control platform considers data from fixed water quality monitoring stations and mobile robotic monitoring units whose timestamps fall within a two-minute range before and after that time layer as the original water quality sample set for that time layer. For each watershed grid unit, the cloud-based control platform calculates the pollution intensity value of that grid unit using a distance-attenuation weighted interpolation method. Specifically, for any watershed grid unit in the diagram, the cloud-based control platform calculates the straight-line distance between the planar coordinates of each original water quality sample record and the coordinates of the grid unit's center point. Original water quality samples within a distance of no more than 300 meters are selected to form a neighboring sample set. Weights are constructed based on the distance between the samples and the grid center; samples closer to the center have higher weights, while samples farther away have significantly lower weights. Finally, the pollution indices of all samples within the neighboring sample set are weighted and summed according to their respective standardized weights to obtain the pollution intensity value of that grid unit in the current time layer.

[0036] In this way, the cloud-based control platform converts water quality data collected from discrete spatial locations by fixed monitoring stations and mobile robots into a continuous pollution intensity distribution covering the entire watershed at the current time level. Figure 1 The grayscale distribution shown represents the watershed grid pollution intensity field obtained after distance attenuation weighted interpolation at a certain time level. This gridded representation provides the basic data structure for subsequent multi-level threshold scanning, active grid connectivity analysis, topological feature extraction, and pollution cluster identification, and is a key step in realizing the identification of topological abrupt events in the water quality field.

[0037] For any given target time frame, the cloud-based control platform considers data from fixed water quality monitoring stations and mobile robotic monitoring units whose timestamps fall within a two-minute range before and after that time frame as the original water quality sample set for that time frame. This time window stitching method can tolerate clock shifts of tens of seconds while ensuring that all samples reflect essentially the same water quality state.

[0038] For each watershed grid cell, the cloud-based control platform calculates the pollution intensity value of that grid cell using distance-attenuation weighted interpolation as follows: First, it calculates the straight-line distance between the planar coordinates of each original water quality sample record and the coordinates of the current watershed grid cell center point. In one embodiment, a search radius can be defined, for example, 300 meters, and original water quality samples within this radius from the current grid center point are selected to form a neighboring sample set. If the number of neighboring sample records is greater than 10, only the 10 samples with the smallest distance can be retained as the neighboring sample set, thereby controlling computational complexity and avoiding excessive influence of excessively distant samples on the local interpolation results. If the number of neighboring sample records is less than 3, the search radius can be temporarily expanded to 500 meters for a second search to reduce the probability of an infinite number of values ​​for the grid cell.

[0039] refer to Figure 2 ,like Figure 2 The diagram illustrates the principle of distance attenuation weighted interpolation. It demonstrates how the cloud-based control platform in this invention calculates the pollution intensity value of a single watershed grid cell using water quality data from neighboring sampling points through the distance attenuation weighted interpolation method.

[0040] A target grid cell, indicated by a solid black line, is marked in the center of the diagram. The side length of this grid cell is set to match the grid resolution in step S1, which is 50 meters by 50 meters in this embodiment. The text "Target Grid" is labeled inside the target grid cell, indicating that this is the watershed grid cell for which the pollution intensity value needs to be calculated. The coordinates of the center point of this target grid cell have been pre-recorded in a unified projected coordinate system and serve as a reference point for subsequent distance calculations.

[0041] Eight sampling points are distributed around the target grid cell. These sampling points come from the original water quality sample set collected by fixed water quality monitoring stations and robotic mobile monitoring units at the current time level. Each sampling point is represented by a circular symbol, and the size of the circle is proportional to the weight value corresponding to that sampling point. That is, the greater the weight of the sampling point, the larger the radius of its circle symbol. This visual design can intuitively reflect the contribution of different sampling points to the calculation of the pollution intensity of the target grid.

[0042] In the figure, each sampling point is connected to the center of the target grid cell by a dashed line, which is represented by a short black dash. These connecting lines represent the straight-line distance from the center of the target grid to each sampling point. The distance value is marked near the midpoint of each connecting line, such as "d1=XXXm", "d2=XXXm", etc., where d1 to d8 represent the distances from the eight sampling points to the center of the target grid, in meters.

[0043] Specifically, the first sampling point is located in the upper left of the target grid cell, approximately 180 meters from the center of the target grid, and is labeled d1. Above this sampling point is a concentration value "C=85", indicating that the pollution index measured at this sampling point is 85. Below this sampling point is a weight value "w=0.25", indicating that after standardization, the weight coefficient of this sampling point in calculating the pollution intensity of the target grid is 0.25. Since the weight value of this sampling point is the largest among all sampling points, the radius of its circular symbol is also the largest.

[0044] The second sampling point is located slightly above the left of the target grid cell, approximately 100 meters from the center of the target grid, and is labeled d2. The concentration value of this sampling point is labeled "C=92", and the weight value is labeled "w=0.18". Although this sampling point is relatively close to the target grid, its weight is 0.18 after standardization because the influence of other closer sampling points needs to be considered.

[0045] The third sampling point is located at the upper right of the target grid cell, approximately 80 meters from the center of the target grid, and is labeled d3. The concentration value of this sampling point is labeled "C=78", and the weight value is labeled "w=0.15".

[0046] The fourth sampling point is located to the right of the target grid cell, approximately 160 meters from the center of the target grid, and is labeled d4. The concentration value of this sampling point is labeled "C=88", and the weight value is labeled "w=0.12".

[0047] The fifth sampling point is located at the lower right of the target grid cell, approximately 150 meters from the center of the target grid, and is labeled d5. The concentration value of this sampling point is labeled "C=70", and the weight value is labeled "w=0.10".

[0048] The sixth sampling point is located slightly to the right below the target grid cell, approximately 40 meters from the center of the target grid, and is labeled d6. The concentration value of this sampling point is labeled "C=65", and the weight value is labeled "w=0.08". This sampling point is one of the closest sampling points to the target grid, but due to its relatively low concentration value, its contribution to the pollution intensity of the target grid is mainly reflected in its high weight.

[0049] The seventh sampling point is located at the lower left of the target grid cell, approximately 120 meters from the center of the target grid, and is labeled d7. The concentration value of this sampling point is labeled "C=95", and the weight value is labeled "w=0.07".

[0050] The eighth sampling point is located to the left of the target grid cell, approximately 200 meters from the center of the target grid, and is labeled d8. The concentration value of this sampling point is labeled "C=82", and the weight value is labeled "w=0.05". Because this sampling point is farthest from the target grid, its weight value is the smallest among all sampling points, and the corresponding circle symbol radius is also the smallest. Figure 2 It can be seen that the weight value of a sampling point is inversely proportional to its distance from the center of the target grid. Sampling points that are closer in distance, such as d6 (40 meters) with a weight of 0.08 and d3 (80 meters) with a weight of 0.15, have a stronger influence than sampling points that are farther away, such as d8 (200 meters) with a weight of 0.05 and d1 (180 meters) with a weight of 0.25. However, it should also be noted that due to differences in the total number and spatial distribution of sampling points, distance is not the only determining factor. The calculation of standardized weights needs to comprehensively consider the sum of the intermediate weights of all neighboring sampling points.

[0051] When performing distance-attenuation weighted interpolation, the cloud control platform first calculates the planar straight-line distance d_i between each sampling point and the center of the target grid. Then, it constructs intermediate weights w_i based on this distance value. Specifically, the distance value is increased by a constant value of 1 to avoid the extreme case of zero distance, i.e., (d_i+1). The reciprocal of this result is then taken as the intermediate weight for that sampling point, i.e., w_i = 1 / (d_i+1). This results in a larger intermediate weight for closer sampling points and a significantly smaller intermediate weight for farther sampling points. Subsequently, the intermediate weights of all sampling points in the neighboring sampling point set are summed to obtain the total intermediate weight Σw_i. The intermediate weight of each sampling point is then divided by the total intermediate weight to obtain the standardized weight W_i = w_i / Σw_i for each sampling point. The sum of these standardized weights equals 1. Figure 2The weight values ​​w=0.25, w=0.18, w=0.15, w=0.12, w=0.10, w=0.08, w=0.07, and w=0.05, as labeled in the code, are the weight coefficients after the above standardization process, and their sum is 1.00. The cloud control platform weights and sums the pollution index C of each sampling point according to its respective standardized weight W_i, that is, it calculates 85×0.25+92×0.18+78×0.15+88×0.12+70×0.10+65×0.08+95×0.07+82×0.05, thus obtaining the pollution intensity value of the target grid cell in the current time layer. Compared with simple averaging, this distance attenuation weighted interpolation method can better preserve the local concentration gradient of the river channel, so that the high pollution values ​​near narrow tributaries or local discharge outlets are not diluted by low-value samples at a distance. Figure 2 This can be intuitively understood as "the closer the measurement point is to the center of the grid, the stronger its representativeness of the grid". This is consistent with the basic principle of spatial data interpolation, that is, samples closer to the grid are more representative of the true state of the target location than samples farther away.

[0052] For each sample in the neighboring sample set, the cloud control platform constructs a weight based on the distance between that sample and the grid center. Specifically, the distance value is first increased by a constant value of 1 to avoid the extreme case of a distance of 0. Then, the reciprocal of this result is used as the intermediate weight for that sample. Thus, samples closer to the grid have a larger intermediate weight, while samples farther away have a significantly smaller intermediate weight. This can be intuitively understood as "the closer the measurement point is to the grid center, the stronger its representativeness of that grid." Subsequently, the intermediate weights of all samples in the neighboring sample set are summed to obtain a total intermediate weight. Then, the intermediate weight of each sample is divided by the total intermediate weight to obtain the standardized weight of each sample. The sum of these weights equals 1. Compared to simple averaging, this distance-attenuation weighting method better preserves the local concentration gradient in the river channel, preventing high pollution values ​​near narrow tributaries or local discharge outlets from being diluted by low-value samples further away.

[0053] The cloud-based control platform calculates a pollution index for each raw water quality sample. The pollution index can be set as a weighted sum of ammonia nitrogen and total phosphorus concentrations. For example, the ammonia nitrogen concentration can be multiplied by 1, and the total phosphorus concentration by 0.5, then summed to obtain a single index. This setting is because, in most watershed water environment standards, ammonia nitrogen has a higher environmental risk sensitivity than total phosphorus. By artificially increasing the weight of ammonia nitrogen in the pollution index, the index can more sensitively reflect the risks posed by ammonia nitrogen pollution. In another implementation, historical monitoring data can be analyzed during the watershed survey phase, adjusting the weight ratio of ammonia nitrogen and total phosphorus to a value suitable for the characteristics of the watershed, such as 1:1 or 2:1.

[0054] For a given watershed grid cell, the cloud-based control platform performs a weighted summation of the pollution indices of all samples within its neighboring sample sets according to their respective standardized weights, thereby obtaining the pollution intensity value of that grid cell at the current time level. This process is repeated for all watershed grid cells across the entire watershed to obtain a pollution intensity distribution covering the entire watershed at the current time level. Watershed grid cells for which no neighboring samples are found can be marked as invalid grid cells and ignored in subsequent topology analyses. These grid cells are often located in areas far from the river channel or in marginal areas lacking monitoring coverage.

[0055] The cloud-based control platform statistically analyzes the pollution intensity values ​​of all watershed grid units within the same time frame to obtain the minimum and maximum pollution intensity values ​​for all grid units. Subsequently, this range is divided into several pollution intensity thresholds. In one embodiment, 20 pollution intensity thresholds can be set, divided in equal increments from the minimum to the maximum pollution intensity value, resulting in 20 multi-level pollution intensity thresholds numbered sequentially from low to high. This multi-level threshold division is equivalent to gradually raising the "water level" in the pollution intensity field, observing the extensive areas formed at low thresholds and the high-pollution core areas retained only at high thresholds. This facilitates subsequent identification of structures that remain connected across multiple thresholds, thereby characterizing stable pollution clumps.

[0056] like Figure 3 The diagram illustrates active grid connectivity analysis under multi-level threshold scanning. It demonstrates how the cloud-based control platform in this invention identifies active grid cells and extracts connected regions at each pollution intensity threshold, thus providing a foundation for lifecycle analysis of topological features.

[0057] Figure 3 It contains four subplots, each corresponding to a different pollution intensity threshold. Each subplot is labeled with its corresponding threshold number and value at the top, from left to right and top to bottom: the first subplot is labeled "Threshold 1 = 40" in the upper left; the second subplot is labeled "Threshold 2 = 80" in the upper right; the third subplot is labeled "Threshold 3 = 120" in the lower left; and the fourth subplot is labeled "Threshold 4 = 160" in the lower right. Each subplot uses the same grid division method, with a grid resolution of 25 columns by 15 rows, and the grid cell size is... Figure 1 Maintain consistency.

[0058] In the first subplot "Threshold 1=40", the cloud control platform marks watershed grid cells with pollution intensity values ​​of 40 or higher as active grid cells, indicated by black fill in the figure; and marks watershed grid cells with pollution intensity values ​​lower than 40 as inactive grid cells, indicated by white fill in the figure. It can be observed that at this lower threshold, a large number of grid cells are marked as active, and the active grids form two main continuous distribution areas in space.

[0059] Specifically, the first connected region is located in the middle left of the subgraph, roughly covering the area from row 1 to row 11 and column 3 to column 14. This connected region is marked with a black dashed box, which is a regular rectangle to emphasize its extent. Inside this region, active black grid cells are connected to each other through four-adjacency relationships, meaning that each active grid cell has at least one active neighboring cell in its four directions (up, down, left, and right), thus forming spatial connectivity. This connected region is labeled "Region A" in the figure, and the label is located near the center of the region.

[0060] The second connected region is located in the lower right part of the subgraph, roughly covering the area from row 5 to row 14 and column 14 to column 23. This connected region is also marked with a black dashed box. Within this region, active grid cells form another independent connected whole through four-adjacent relationships. This connected region is labeled "Region B" in the figure, and the label is located near the center of the connected region.

[0061] With a threshold of 1=40, due to the low threshold setting, most grid cells with a certain level of pollution are included in the active grid category. Therefore, the spatial range of the two connected regions is large, covering a wide area including the core pollution area and the surrounding pollution diffusion area.

[0062] In the second subplot, "Threshold 2 = 80", the cloud control platform raises the pollution intensity threshold to 80. Under this threshold, only watershed grid cells with a pollution intensity value of 80 or higher are marked as active grid cells. Compared to threshold 1, the number of active grid cells is significantly reduced, while the area of ​​inactive grids (white) is significantly increased.

[0063] At this threshold, the spatial extent of region A shrinks, with its connected regions roughly covering the area from row 2 to row 10 and column 4 to column 13, marked with a black dashed box. The number of active grid cells in this connected region decreases compared to threshold 1, but spatial connectivity is still maintained; that is, active grid cells within this connected region are still interconnected through four-adjacency relationships. The label for region A remains at the center of this connected region, indicating that although the spatial extent of the region shrinks, topological connectivity is preserved, and it can be identified as a continuation of the same topological feature at a higher threshold.

[0064] The spatial extent of region B is also reduced accordingly, with its connected regions roughly covering the area from row 6 to row 13 and column 15 to column 22, marked with a black dashed box. The number of active grid cells in this connected region is also reduced, but it still maintains connectivity. The label for region B is located at the center of this connected region.

[0065] In the third subplot, "Threshold 3=120", the pollution intensity threshold is further increased to 120. At this higher threshold, only watershed grid cells with a pollution intensity value of 120 or higher are marked as active grid cells. The number of active grids is further significantly reduced, mainly concentrated in the core high-value areas of the original pollution patches.

[0066] The connected region of region A has been further reduced, roughly covering the area from row 3 to row 8 and column 5 to column 12, marked with a black dashed box. This connected region has been reduced to the core area with high pollution intensity, and the surrounding areas with relatively low pollution intensity no longer meet the threshold condition and are therefore no longer included in the active grid. Despite the significant reduction in spatial extent, the active grid cells within region A still maintain four-adjacent connectivity, allowing this topological feature to continue to be tracked at threshold 3.

[0067] The connected regions of region B have also shrunk significantly, roughly covering the area from row 8 to row 12 and column 16 to column 21, marked with a black dashed box. This region has shrunk to its core contamination area, and the number of active grids is significantly less than in the previous two threshold cases.

[0068] In the fourth subplot "Threshold 4=160", the pollution intensity threshold is set to the highest value of 160. At this extremely high threshold, only watershed grid cells with a pollution intensity value of 160 or higher are marked as active grid cells. The number of active grid cells is reduced to a minimum, retaining only the core area with the highest pollution intensity.

[0069] Region A's connected region is reduced to a very small area, roughly covering rows 4 to 7 and columns 6 to 10, marked with a black dashed box. This connected region contains only a few grid cells near the peak pollution intensity, but these grid cells still maintain spatial connectivity, indicating that the core structure of this topological feature still exists at high thresholds. Region B's connected region is also reduced to a minimum core, roughly covering rows 9 to 11 and columns 17 to 20, marked with a black dashed box. This region retains only the highest intensity core of the pollution clumps, with the fewest active grid cells among the four thresholds.

[0070] By comparing the four sub-graphs, it is clear that as the pollution intensity threshold gradually increases from 40 to 80, 120, and 160, the number of active grid cells gradually decreases, and the spatial range of connected regions gradually shrinks. However, regions A and B maintain spatial connectivity at all thresholds. These regions that maintain connectivity across multiple threshold intervals are the topological features defined in this invention. By tracking the existence of these topological features across consecutive threshold intervals, the cloud-based control platform can calculate their lifecycle length and thus identify stable pollution clusters.

[0071] Figure 3 The bottom is labeled with an illustration. Black-filled squares are labeled "Active Grids (Pollution Intensity ≥ Threshold)," representing grid cells whose pollution intensity values ​​meet the condition at the current threshold. White-filled squares are labeled "Inactive Grids (Pollution Intensity < Threshold)," representing grid cells whose pollution intensity values ​​do not meet the condition at the current threshold. Black dashed boxes are labeled "Connected Region Boundaries," representing the spatial extent of connected regions identified through four-adjacency relationships. In actual calculations, for each pollution intensity threshold, the cloud control platform uses a row-by-row scanning combined with queue expansion. Starting from any unvisited active grid cell, it groups all four adjacent active grid cells into the same connected region and assigns them a unified region identifier number, until no new adjacent active grid cells can be added. Then, it restarts this process from the next unvisited active grid cell until all active grid cells are classified into a connected region.

[0072] For each pollution intensity threshold, the cloud-based control platform marks watershed grid cells with pollution intensity values ​​not lower than the current threshold as active grid cells, and marks the remaining watershed grid cells as inactive grid cells. Then, a four-adjacency relationship is established between active grid cells; that is, when two watershed grid cells share a common edge on the two-dimensional plane, these two grid cells are considered spatially adjacent and belong to the same connected region. The cloud-based control platform can use a row-by-row scanning combined with queue expansion to start from any unvisited active grid cell, grouping all four adjacent active grid cells into the same connected region and assigning them a unified region identifier number, until no new adjacent active grid cells can be added. This process is then restarted from the next unvisited active grid cell until all active cells are classified into a connected region. In this way, a set of connected regions can be obtained for each pollution intensity threshold, with each connected region corresponding to a continuous high-value region in the pollution intensity field.

[0073] Each connected region obtained under the lowest pollution intensity threshold is considered an initial topological feature. The cloud control platform records the birth threshold number for each initial topological feature, for example, as 1, and also records the set of all watershed grid cells contained in that topological feature. When processing the second pollution intensity threshold, the cloud control platform similarly divides connected regions and records the set of grid cells for each connected region. For each connected region under the current threshold, the number of intersection grids and the number of union grids between its grid cell set and the corresponding region sets of all topological features under the previous threshold are calculated, thus obtaining the intersection-union ratio (IU). The IU reflects the degree of spatial coverage consistency between two regions. If the IU of a current connected region with a topological feature under a previous threshold is not less than a preset ratio threshold, for example, 0.5, then the topological feature is considered to continue to exist under the current threshold, and the grid cell set of the current connected region is merged into the spatial range record of the topological feature; if the IU of a current connected region with all topological features under the previous threshold is lower than the preset ratio threshold, then a new topological feature is generated, and the current threshold number is used as the birth threshold of the topological feature.

[0074] As the cloud-based control platform continues processing each subsequent pollution intensity threshold, it repeats the aforementioned correlation process. When an existing topological feature cannot find any currently connected region with an intersection-union ratio (IU) not less than a certain threshold at a certain threshold, it is considered that the topological feature has ended its existence at the previous threshold. The previous threshold number is recorded as the stopping threshold for this topological feature, and the lifecycle length is calculated as the stopping threshold number minus the birth threshold number. The larger the lifecycle length, the more spatially connected the topological feature is, indicating that it maintains a certain level of spatial connectivity across multiple pollution intensity thresholds. Such topological structures typically correspond to stable polluted blocks in actual water bodies, rather than random noise. Through this process, the evolution trajectories of all topological features in the current time layer, along with their lifecycle lengths and spatial extents, can be obtained.

[0075] In a preferred embodiment, the cloud control platform considers topological features with a lifecycle length of 3 or more as stable candidates. Simultaneously, it requires that the number of active grid cells for each topological feature at a certain pollution intensity threshold be no less than 10. This excludes small, speckled structures composed of only a few grid cells, thus correlating topological features with engineering-significant pollution areas. For topological features meeting these conditions, the cloud control platform marks them as pollution clumps, and uses the coverage area of ​​active grid cells at the pollution intensity threshold corresponding to the midpoint of their lifecycle as the spatial extent of the pollution clump. Using the spatial extent corresponding to the midpoint of the lifecycle offers a trade-off: the same topological feature may have a larger extent at the lowest threshold, while at the highest threshold, only the core area may remain. The spatial extent at the midpoint of the lifecycle typically covers the main pollution area without extending to a large number of low-value peripheral areas, making it more suitable for subsequent event identification and remediation area delineation.

[0076] To identify contaminant cluster generation and fragmentation events, the cloud-based control platform repeats the aforementioned grid-based contaminant intensity field construction and topological feature extraction process across multiple consecutive time layers, thereby obtaining a set of contaminant clusters at each time layer. For two adjacent time layers, the cloud-based control platform sequentially traverses each contaminant cluster in the current time layer, calculating the intersection-union ratio (IUR) between the spatial extent of the contaminant cluster and the spatial extent of all contaminant clusters in the previous time layer. When the IUR of a contaminant cluster in the current time layer and all contaminant clusters in the previous time layer is lower than a preset lower limit, such as 0.1, it can be considered that the contaminant cluster did not have a corresponding predecessor in the previous time layer and is a newly added contaminant cluster in the current time layer. The cloud-based control platform records the spatial location and time of the contaminant cluster as a contaminant cluster generation event. The consideration for setting this lower limit is that extremely small spatial overlap is more likely to come from numerical errors or edge contamination, while a consistently lower IUR usually indicates a truly new contaminant source in space and time or an aggregation area caused by local hydrodynamic changes.

[0077] For each contaminated cluster in the previous time layer, the cloud control platform also calculates the intersection-union ratio (IUR) between its spatial extent and the spatial extent of all contaminated clusters in the current time layer. When it is found that the IUR of a contaminated cluster in the previous time layer with at least two contaminated clusters in the current time layer is not lower than a preset upper limit, such as 0.3, and the sum of these IURs is not lower than 0.6, it can be considered that the contaminated cluster in the previous time layer has undergone significant spatial splitting, evolving into multiple independent contaminated clusters. The cloud control platform marks the contaminated cluster in the previous time layer as a contaminated cluster splitting event and associates and records the spatial extent of all contaminated clusters in the current time layer with which it has a sufficient IUR. This method of judging splitting by the upper limit of the IUR can filter out cases of only minor deformation or slight shearing, highlighting those true splitting behaviors that require differentiated treatment in governance, which is beneficial for subsequently issuing governance tasks separately for different branches of contaminated clusters.

[0078] Through the above processing, the cloud-based control platform generates a set of water quality field topological mutation events, including pollution cluster generation and fragmentation events, over a continuous time series. Each water quality field topological mutation event records at least the event type, occurrence time, spatial extent of the corresponding pollution cluster, average pollution intensity of the pollution cluster, and lifespan length. Subsequently, when performing robot density distribution calculations and task allocation, resources can be concentrated to address newly emerging high-risk areas based on these events, or different robot groups can be deployed at different tributaries and bifurcations for fragmented clusters, thereby improving the overall response efficiency of intelligent biological governance of the watershed's water environment.

[0079] Step S2: The cloud control platform calculates the water quality risk score of each water unit based on the water quality field topological mutation event and static water quality indicators, sets the target robot density, constructs a single-unit benefit function that includes water quality risk, density deviation penalty, and time and energy consumption terms, executes multiple rounds of mean field game iteration and adaptively adjusts the density penalty coefficient to make the actual robot density converge to the robot target density, and obtains the converged robot density distribution and water quality treatment priority level.

[0080] In one specific implementation, the cloud control platform first discretizes the main channel and tributaries of the watershed based on the set of water quality field topological mutation events and the grid pollution intensity field obtained in step S1. The channel centerline is divided into several water area units according to length, with each unit's length potentially set to 500 meters. The platform then records the corresponding set of watershed grid units and the water surface area for each unit. In a medium-sized watershed scenario, the main channel is 20 kilometers long, and the total length of the tributaries is 30 kilometers, resulting in approximately 100 water area units. The area of ​​each unit is estimated by multiplying the river length by the average river width, for example, set to 50 meters. This area estimation provides a uniform scale for subsequently converting the number of robots into density, ensuring comparability in robot density comparisons across river segments of different lengths.

[0081] The cloud-based control platform constructs dynamic water quality indicators for each watershed unit based on the pollution intensity values ​​of each watershed grid unit and the set of abrupt topological changes in the water quality field over a recent period. Specifically, for each watershed unit, the cloud-based control platform reads the pollution intensity values ​​of all watershed grid units covered by that watershed unit at the most recent time level and obtains the average pollution intensity of that watershed unit through arithmetic averaging. This avoids overestimation of risk due to a single high-value grid, ensuring that the water quality risk score simultaneously reflects both high-value areas and the overall background level. For the same watershed unit, the cloud-based control platform counts the number of watershed grid units belonging to any pollution cluster and divides this number by the total number of watershed grid units within that watershed unit to obtain the pollution cluster coverage ratio. The pollution cluster coverage ratio reflects the proportion of areas identified as stable pollution areas within that watershed unit. When the pollution cluster coverage ratio is high, even if the average pollution intensity is not particularly prominent temporarily, it indicates the existence of a long-term pollution structure in the area, requiring more robots for remediation. The cloud-based control platform further counts the number of pollution cluster generation events and pollution cluster splitting events that fall into the watershed unit within a preset time window, such as the most recent 30 minutes. The number of pollution clump formation events represents the frequency of new highly polluted clumps being generated, while the number of pollution clump splitting events represents the spatial expansion trend of pollution patches. Both of these indicate that the local pollution morphology is in an unstable state and may deteriorate rapidly in the future, so they need to be reflected in the risk assessment.

[0082] Static water quality indicators can be derived from watershed functional zoning and ecological sensitivity surveys. For each water body, its water function zone type can be pre-labeled, such as a drinking water source protection zone, a landscape functional zone, or a general agricultural irrigation zone. The cloud-based control platform can assign static sensitivity levels to different water function zones; for example, a drinking water source protection zone can be assigned level 3, a landscape functional zone level 2, and a general agricultural irrigation zone level 1. Static water quality indicators are then weighted when calculating water quality risk scores. In this way, when two water bodies have similar dynamic pollution characteristics, the water body located in an ecologically sensitive zone or a drinking water source protection zone will receive a higher water quality risk score, which is beneficial for prioritizing the allocation of governance resources to sensitive areas.

[0083] Based on the above, the cloud-based control platform calculates a water quality risk score for each water area unit. A linear weighting method can be used, combining average pollution intensity, pollution cluster coverage ratio, number of topological abrupt events in the water quality field, and static water quality indicators onto the same scoring scale. In one implementation, the cloud-based control platform can multiply the average pollution intensity by 20, the pollution cluster coverage ratio by 50, the number of pollution cluster generation events by 10, the number of pollution cluster splitting events by 15, and the static water quality sensitivity level by 30, then sum these values ​​to obtain the water quality risk score for that water area unit. This numerical selection allows the average pollution intensity to differentiate between different water areas, the pollution cluster coverage ratio to exert a greater influence on local long-term pollution structures, and pollution cluster generation and splitting events to amplify the risk of unstable areas with smaller but sensitive weights. Static water quality sensitivity is then weighted appropriately to shift the overall risk level of sensitive areas upwards. In another implementation, those skilled in the art can adjust the above weights using historical environmental incident and remediation effect data. For example, the weight of average pollution intensity can be reduced to 10, and the weight of pollution patch coverage ratio can be increased to 80, in order to further emphasize the importance of spatial structure information. As long as the overall calculation order remains the same, this is a reasonable adjustment to this implementation.

[0084] After obtaining the water quality risk score for each water area, the cloud control platform determines the target robot density based on the score. In one embodiment, the target robot density can be set using a simple linear relationship superimposed with a basic density constant. Specifically, the water quality risk score can be divided by 100, and then a constant of 2 can be added to obtain the number of robots per square kilometer as the target robot density. Thus, when the water quality risk score for a water area is 200, the target robot density is 4 robots per square kilometer; when the water quality risk score is 50, the target robot density is 2.5 robots per square kilometer. This setup is based on the consideration that even in low-risk areas, a certain number of robots are still needed for routine inspections and data collection. Therefore, the basic density constant ensures that the target robot density is not too low, while the higher the water quality risk score, the greater the increase in the target robot density, thereby achieving a more intensive monitoring and deployment capability in high-risk areas.

[0085] Next, the cloud control platform collects the current location and remaining battery information of all robots, assigning each robot to a specific water area unit based on its position in the watershed coordinate system. For each water area unit, the number of robots assigned to that unit is counted, and the number of robots is divided by the area of ​​that water area unit to obtain the initial actual robot density, also in units of robots per square kilometer. In most scenarios, the initial distribution of robots within the watershed may be based on the results of the previous round of task execution and may not match the current water quality risk distribution. Therefore, it is necessary to gradually adjust the distribution through subsequent mean-field game iterations.

[0086] To gradually bring the actual robot density closer to the target robot density, the cloud-based control platform constructs an individual payoff function based on mean-field game theory. This ensures that each robot, when choosing its movement direction, is both inclined to go to water units with higher water quality risk scores and constrained by density deviation penalties, travel time, and energy consumption. The intention behind this design is that if only water quality risk is considered, high-risk water units may attract too many robots, leading to local overcrowding while other areas are neglected. Including density deviation penalties gradually reduces the attractiveness of water units that are already close to or exceed the target density, promoting balanced governance across the entire watershed. Simultaneously, incorporating travel time and energy consumption into the payoff function encourages robots to choose shorter or lower-energy-consumption movement strategies while ensuring governance effectiveness, mitigating the impact of power limitations on task completion rates.

[0087] In one implementation, the cloud control platform sets the number of iterations for the mean-field game, for example, 10 rounds. In each iteration, for each robot, a candidate water area set is constructed based on its current water area unit. The candidate water area set may include the current water area unit, its upstream neighboring water area units, and its downstream neighboring water area units. Typically, each robot corresponds to 2 to 3 candidate water area units. This setting allows the robot to consider only movement choices within a local range in each iteration, thereby reducing computational complexity and conforming to the reality that the robot can only move a limited distance in a finite amount of time. Those skilled in the art can also increase the number of candidate water area units, for example, by simultaneously considering the current water area unit, two upstream neighboring water area units, and two downstream neighboring water area units, giving the robot a larger decision space, but the convergence speed of the mean-field game may be slower.

[0088] For each robot and each combination of its candidate water units, the cloud control platform calculates an estimated travel time based on the distance to the river centerline and the preset cruising speed. This estimated travel time is then converted into an estimated energy consumption based on the robot's power consumption characteristics. For example, if the cruising speed is 0.5 meters per second, the estimated travel time for a water unit 250 meters away is approximately 500 seconds. If the robot's average power consumption during cruising is 100 watts, the estimated energy consumption for this movement can be set to approximately 50,000 joules. This energy consumption estimate can be further converted into a percentage of battery power consumption. This percentage is used as an energy consumption term in the individual revenue function and can also be used to limit the robot's movement to avoid traveling to distant water units when its remaining battery power is low.

[0089] When calculating unit-level revenue, the cloud-based control platform uses the water quality risk score of a water area unit as the basis for the revenue component, and deducts density deviation penalty, travel time, and energy consumption as costs. The density deviation penalty is determined by the difference between the actual robot density and the target robot density, along with a density penalty coefficient. Specifically, when the actual robot density of a water area unit is significantly higher than the target robot density (i.e., the density error is positive and large), the cloud-based control platform sets the density penalty coefficient to a larger value, such as 20, giving the density deviation penalty a greater weight in unit-level revenue. This way, even if the water quality risk score of that water area unit is high, robots will tend to choose other water areas with similarly high risk but insufficient robot density, thus avoiding excessive robot concentration. When the actual robot density of a water area unit is close to the target robot density, the density penalty coefficient can be set to a medium value, such as 10, so that the density deviation has only a moderate impact. When the actual robot density of a water area unit is lower than the target robot density, the density penalty coefficient can be set to a smaller value, such as 5, so that the water area unit still maintains high attractiveness in unit-level revenue, encouraging more robots to fill the gap. The travel time and energy consumption items can be directly deducted from the revenue by multiplying the estimated travel time and energy consumption by a uniform cost coefficient. This ensures that when the water quality risk scores of multiple candidate water areas are similar, the robot is more inclined to choose the destination with shorter travel time and lower energy consumption.

[0090] For the same robot, the cloud control platform can shift and normalize the individual payoff values ​​of each candidate water unit to form the robot's migration probability distribution. Specifically, the process is as follows: First, find the minimum payoff value among all candidate water units. Subtract this minimum value from all individual payoff values ​​to obtain a set of non-negative values. If all individual payoff values ​​are the same, the robot can be considered to have no preference among all candidate water units, and can be assigned the same migration probability to each water unit. If differences exist, sum all non-negative values ​​as a normalization factor, and then divide each non-negative value by the normalization factor to obtain the robot's migration probability for each candidate water unit. This shifting and normalization method ensures that the migration probability of the water unit with the lowest individual payoff is 0, and the migration probabilities of other water units are proportional to their relative payoffs. Mathematically, this is similar to a linearly weighted random selection, which is smoother than a simple greedy selection and facilitates gradual adjustment of the entire robot swarm during iteration rather than violent oscillations.

[0091] After obtaining the migration probabilities of all robots, the cloud control platform updates the expected number of robots for each water area unit based on these probabilities. For each water area unit, the cloud control platform sums the migration probabilities of all robots for that water area unit to obtain the expected number of robots for that water area unit, and divides this by the area of ​​the water area unit to obtain the expected robot density. The expected robot density can be regarded as the predicted value of the robot density after the current iteration round, so the cloud control platform can directly overwrite the expected robot density with the new actual robot density, providing a basis for the next round of iteration. At the same time, the cloud control platform calculates the new density error for each water area unit by subtracting the actual robot density from the target robot density. Based on the numerical range of the density error, the cloud control platform updates the density penalty coefficient used for that water area unit in the next round of iteration. For example, when the density error is greater than 2 robots per square kilometer, the density penalty coefficient is increased to 20; when the density error is between 0 and 2 robots per square kilometer, the density penalty coefficient is set to 15; and when the density error is less than 0, the density penalty coefficient is set to 5. The larger the density error, the higher the density penalty coefficient, which further weakens the individual benefits of the water area unit in the next iteration. This adaptive adjustment method can accelerate the "cooling down" of high-density areas, encourage robots to migrate to low-density areas, and reduce the density error as a whole.

[0092] The above process is repeated in each iteration. After 10 iterations, the cloud control platform recalculates the density error for each water area unit. When the density error of most water area units has narrowed to a small range, such as no more than one robot per square kilometer, the actual robot density can be considered to be close to the target robot density. The actual robot density distribution at the end of that iteration is then used as the converged robot density distribution. In another implementation, instead of fixing the number of iterations in advance, the maximum value of the density error across the entire watershed is calculated after each iteration. When this maximum value is lower than a preset convergence threshold, the iteration is terminated early, and the actual robot density at this point is also used as the converged robot density distribution. This adaptive termination method can reduce unnecessary iterations in simple scenarios and maintain a sufficient number of iterations in complex scenarios to achieve better convergence.

[0093] After obtaining the converged robot density distribution and the water quality risk score for each water area, the cloud control platform classifies water quality management priorities based on the water quality risk score. Several water quality risk score thresholds can be set; for example, a water area with a water quality risk score greater than or equal to 150 is classified as Level 3, a score between 80 and 150 as Level 2, and a score below 80 as Level 1. Level 3 indicates areas requiring priority management and significant resource investment, Level 2 indicates areas requiring routine management and close monitoring, and Level 1 indicates areas requiring basic inspections. This classification method can be directly used in subsequent task generation and market auction allocation, allowing application and sampling tasks to obtain higher basic task value in high-priority water quality management areas, thereby ensuring that the entire system automatically adjusts robot configuration and workload based on the water quality risk distribution.

[0094] Step S3: The cloud control platform generates sampling, dosing, and inspection tasks based on the converged robot density distribution and water quality treatment priority level, and determines the basic value of the tasks. Each robot calculates its bidding score under the constraints of remaining power and matching biological agents. The cloud control platform conducts a market auction according to the principle of maximizing the bidding score and assigns the generated tasks to the winning robot for execution.

[0095] In step S3, the cloud control platform, having obtained the converged robot density distribution and water quality treatment priority levels, generates sampling, dosing, and inspection tasks for each water area. A basic task value is assigned to each task. Then, each robot, considering its remaining power and biological agent loading, calculates a bidding score for each achievable task. The cloud control platform executes a market auction based on the principle of maximizing the bidding score, assigning each task to the most suitable winning robot. The winning robot then performs sampling, dosing, and inspection operations according to the task requirements. This entire process utilizes the results of the previous density convergence stage and, through a market-like competition mechanism, allocates limited robot resources to tasks that contribute more to overall water quality treatment.

[0096] In one specific implementation, the converged robot density distribution and water quality treatment priority level have been determined in step S2. Each water area unit has a water quality risk score and a water quality treatment priority level, for example, it can be divided into three levels: Level 1, Level 2, and Level 3, where Level 3 represents a high-risk area requiring priority treatment, and Level 1 represents an area with relatively low risk. Based on this information, the cloud control platform generates a task set for each water area unit. Introducing the converged robot density distribution when generating the task set serves two purposes: firstly, it avoids deploying too many tasks in areas where the robot density is already low, causing tasks to remain idle for extended periods; secondly, it increases the number of tasks in areas where the robot density has reached the target density, making full use of the robots already gathered in that area and improving robot utilization.

[0097] The sampling task is primarily used to acquire water quality data with higher spatiotemporal resolution to verify the judgment and application effect of the water quality field topological abrupt change event set. In one embodiment, for a water area with a water quality treatment priority level of 3, sampling points can be set at equal intervals along the centerline of the river within the water area of ​​that water area, for example, one sampling point every 200 meters. If the river length of the water area is 500 meters, then three sampling points are generated. Each sampling point records a planar coordinate as the sampling location and a sampling time window, which can be set to a 30-minute interval extending backward from the current time. This ensures that the robot has enough time to move from its current location to the sampling point and can acquire new data in a short time to quickly report on the pollution diffusion situation. For water areas with a water quality treatment priority level of 2, the sampling point spacing can be widened to one sampling point every 400 meters, while for water areas with a water quality treatment priority level of 1, only one sampling point can be set near the center of the water area to maintain basic monitoring coverage.

[0098] The application task is used to deliver customized biological agents at key locations to reduce pollution. In one embodiment, the cloud control platform selects the geometric center of the pollution cluster's range as the application point within a water unit with a water quality treatment priority level of 3, based on the spatial range of the pollution cluster recorded in the set of topological mutation events in the water quality field. If multiple pollution clusters exist within the same water unit, a separate application task can be generated for each pollution cluster. Each application task records the coordinates of the application point, the application time window, and the biological agent type number. The application time window can be set to a 60-minute interval extending backward from the current time, ensuring that the robot has time to complete its navigation and that intervention can be carried out before the pollution cluster spreads further. The biological agent type number can be pre-selected based on the experimental results of the biological agent reaction device. For example, biological agent number 1 is used for pollution clusters dominated by ammonia nitrogen, and biological agent number 2 is used for pollution clusters dominated by total phosphorus. For water bodies with a water quality treatment priority level of 2, application tasks can be generated only for polluted clumps whose area and average pollution intensity both exceed the threshold. For water bodies with a priority level of 1, application tasks are usually not generated directly unless the water body is marked as a key functional area.

[0099] The inspection task is used to perform continuous status checks and safety patrols along the river channel. In one embodiment, the cloud control platform can generate one or more inspection paths for each water area unit based on water quality management priority and river segment length. The inspection path consists of a series of continuous coordinate points, typically laid out along a trajectory slightly offset to one side of the river channel centerline to reduce the risk of impacts to riverbanks and bridge piers. For water area units with a water quality management priority of 3, the inspection path length can cover the entire river segment of the water area unit, with multiple round trips set within 60 minutes after the current time. For water area units with a water quality management priority of 1, only a one-way inspection path can be set, and the inspection time window can be set to a longer period, such as one inspection within 120 minutes after the current time.

[0100] After generating the task set, the cloud control platform assigns a base value to each task based on the water quality management priority and task type. The base value reflects the task's contribution to the overall watershed management goals and serves as a crucial scoring criterion in subsequent market auctions. In one embodiment, the following numerical allocation method can be used: In water units with a water quality management priority of 3, the base value for the delivery task is set to 100, the base value for the sampling task is set to 80, and the base value for the inspection task is set to 60; in water units with a water quality management priority of 2, the base values ​​for the delivery, sampling, and inspection tasks are set to 80, 60, and 40 respectively; in water units with a water quality management priority of 1, the base values ​​for the delivery, sampling, and inspection tasks are set to 60, 40, and 20 respectively. This design ensures that tasks in high-risk areas have significantly higher values ​​than those in low-risk areas, especially delivery tasks, which have the highest value. Therefore, in subsequent auctions, robots are more likely to prioritize delivery tasks in high-risk areas, allowing limited resources to be concentrated on critical tasks.

[0101] With the basic value of the task already determined, each robot calculates its bidding score based on its own status and task information. A higher bidding score indicates a greater overall benefit for the robot performing the task. In one embodiment, the cloud control platform independently executes a market auction within each water area unit to reduce computational scale. For a target water area unit, all tasks within that unit and the set of robots currently located in that unit and its adjacent units are considered. For each robot and each task combination within that water area unit, the travel distance is first approximated along the river centerline based on the coordinate difference between the robot's current position and the task target position. The travel distance is divided by a preset cruising speed constant to obtain an estimated travel time. For example, the cruising speed constant can be set to 0.5 meters per second; when the distance between the robot and the task position on the river centerline is 250 meters, the estimated travel time is approximately 500 seconds.

[0102] The cloud-based control platform converts the estimated travel time into an estimated energy consumption based on the robot's energy consumption characteristics. For example, assuming the robot's average power during cruising is 100 watts, a 500-second cruise would require approximately 50,000 joules of energy. After converting the battery's nominal capacity into a percentage, the approximate percentage of electricity consumed during this trip can be obtained. The cloud-based control platform then sets a resource coefficient based on the robot's current remaining battery percentage. When the robot's remaining battery percentage is greater than or equal to 50%, the resource coefficient is 1, indicating that the robot can complete the current task without significantly affecting subsequent tasks. When the robot's remaining battery percentage is between 30% and 50%, the resource coefficient is 0.8, indicating that the robot can perform the task, but its value assessment needs to be moderately reduced compared to a fully charged state. When the robot's remaining battery percentage is less than 30%, the resource coefficient is 0.5, indicating that the robot's power is low and high-load tasks should be reduced. This reduction naturally decreases its competitiveness for long-distance tasks during the auction process.

[0103] For dosing tasks, it's also necessary to consider whether the type of biological agent carried by the robot matches the task requirements. If the biological agent type number recorded in the task log matches the biological agent type number actually carried by the robot, it can be assumed that the robot can directly perform the dosing without returning to the resupply point or going through a complex reloading process. This matching is crucial for response speed and treatment effectiveness. Therefore, in one implementation, when the biological agent type matches, the task's base value can be multiplied by 1.2 to obtain the matched base value; when the biological agent type does not match, the task's base value can be multiplied by 0.5 to obtain the matched base value. Robots with this matching will have a significant advantage in the dosing task auction, thus encouraging the system to prioritize the use of robots already carrying the correct biological agent, reducing unnecessary empty trips and waiting time. For sampling and inspection tasks, since they do not depend on the biological agent type, the matched base value can be directly equal to the task's base value.

[0104] When calculating the bid score, the cloud control platform multiplies the matched base value by the resource coefficient to obtain a comprehensive value that combines the task's importance with the robot's remaining battery power. Then, it subtracts the time cost corresponding to the estimated travel time and the energy cost corresponding to the estimated energy consumption from this comprehensive value. For example, the estimated travel time can be directly used as the time cost, and the estimated travel time multiplied by 0.5 can be used as the energy cost. Thus, when a task is far from a robot, even if the task's base value is high, the travel time cost and energy cost will significantly reduce the bid score for that combination, preventing the system from assigning the robot a task that is clearly mismatched with its location. When the sum of the time cost and energy cost exceeds the comprehensive value, the bid score may become negative. For ease of subsequent comparison, in one implementation, bid scores less than 0 can be truncated to 0, indicating that the robot is not competitive for that task. This truncation prevents extremely low values ​​from affecting numerical stability and also facilitates considering only combinations with bid scores greater than 0 during implementation.

[0105] After calculating the bid scores for all robots and task combinations within a specific water area, the cloud control platform executes the market auction process. In one implementation, all tasks within the water area are first sorted from highest to lowest based on their basic task value. This sorting ensures that the auction prioritizes tasks with the greatest impact on overall governance, allowing these tasks to select the most suitable robots. At the start of the auction, the first unassigned task in the sorted list is selected as the current auction task. All unsuccessful robots are iterated through, searching for a set of robots with bid scores greater than 0 for this task. If this set is not empty, the robot with the highest bid score is selected as the winning robot. When multiple robots submit the same maximum bid score for the task, their estimated travel times are compared, and the robot with the lowest estimated travel time is selected as the winning robot. This prioritizes robots with faster response times when the returns are equal, shortening the time interval between task assignment and completion. After determining the winning robot, the current task is bound to that robot, and the robot is removed from the set of robots eligible to participate in this round of auction to prevent it from undertaking multiple tasks simultaneously in this round.

[0106] The cloud control platform repeats the auction steps described above, performing the same selection process for the next unassigned task in the task list. For a given task, if all unsuccessful robots have a bid score of 0 for that task, it indicates that the current robot set is unsuitable for executing the task due to limitations in power, location, and biological agent matching. In this case, the task can be marked as a task awaiting replanning and retained for re-evaluation in the next round of global planning. The auction process continues until all tasks within the watershed unit have been assigned, or all robots have won bids for tasks. Afterward, the cloud control platform moves to the next watershed unit and executes the auction in the same manner. Finally, the auction results for each watershed unit can be aggregated to form a comprehensive watershed task allocation result table, recording the task type, target location, time window, and type of biological agent required for each robot.

[0107] After the task allocation result table is generated, each robot generates its actual execution path based on its assigned task. In a simplified implementation, each robot can be restricted to receiving only one task in each round of task allocation. This way, the robot only needs to plan a path from its current position to the single task target location. The path can be generated along the river centerline, and the robot adjusts its posture to the specified water depth and lateral position before approaching the target. For scenarios that require processing multiple task sequences, some robots can be allowed to receive multiple tasks at once. In this case, the robots can sort the tasks according to their time windows from earliest to latest, or adopt a proximity-first strategy, placing the task closest to the current position first. Regardless of the method used, when the robot reaches the sampling point of the sampling task, it samples the water quality according to the task time window and uploads the data to the cloud control platform; when it reaches the dosing point of the dosing task, it controls the onboard biological agent dosing device to complete the dosing according to the dosing time window and biological agent type number recorded in the task record; while navigating along the inspection path specified by the inspection task, it collects water quality data and operational status data at preset intervals. Through this market auction strategy centered on maximizing bid scores, the system can dynamically allocate limited robot resources to the tasks most valuable for improving the overall water environment quality, based on the converged robot density distribution and water quality treatment priority levels, within each time period.

[0108] In another alternative implementation, the market auction process can be extended to a joint auction across multiple water units. In this case, the cloud control platform merges tasks from multiple adjacent water units, sorts them uniformly according to their basic value, and allows robots to participate in the auction across water units. This approach is advantageous when the number of robots is limited and they are concentrated in high-risk areas, breaking the limitations of a single water unit and allowing more robots to be dispatched from low-risk areas to high-risk areas to perform tasks. Those skilled in the art can also incorporate the influence of the remaining time window length of the task into the bidding score calculation based on the urgency of the task. For example, a certain bonus can be added to tasks with shorter remaining time windows, allowing the system to prioritize tasks that are prone to being missed. These adjustments can all be implemented based on existing processes.

[0109] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A basin water environment intelligent biological treatment robot system, characterized in that, The system includes: several fixed water quality monitoring stations and mobile monitoring units mounted on robots, used to collect water quality data and its temporal and spatial locations within the watershed; several robots used to move within the watershed to perform sampling, dosing, and inspection tasks; and a cloud control platform, which includes a memory and a processor. The memory stores programs that can run on the processor, which is configured to: construct a watershed grid pollution intensity field using distance attenuation weighted interpolation based on the spatiotemporal water quality data collected by the fixed water quality monitoring stations and robots; extract topological features through multi-level threshold scanning and active grid connectivity analysis; identify pollution clusters based on the lifecycle of the topological features within continuous threshold intervals; and determine pollution cluster generation events and their spatial overlap relationships between adjacent time-layer pollution clusters. The splitting event forms a set of water quality field topological mutation events. Based on the water quality field topological mutation events and static water quality indicators, the water quality risk score of each water unit is calculated, and the target robot density is set. A single-unit revenue function is constructed, which includes water quality risk, density deviation penalty, and time and energy consumption terms. Multiple rounds of mean field game iteration are performed, and the density penalty coefficient is adaptively adjusted to make the actual robot density converge to the robot target density, thus obtaining the converged robot density distribution and water quality treatment priority level. Based on the converged robot density distribution and water quality treatment priority level, sampling tasks, dosing tasks, and inspection tasks are generated, and the basic value of the tasks is determined. Each robot is guided to calculate the bidding score under the constraint of remaining power and matching biological agents. The market auction is conducted according to the principle of maximizing the bidding score, and the generated tasks are assigned to the winning robot for execution.

2. A customized dosing method for the watershed water environment intelligent biological treatment robot system of claim 1, characterized in that, Includes the following steps: Step S1: The cloud control platform integrates spatiotemporal water quality data collected by fixed monitoring stations and robots, constructs a watershed grid pollution intensity field using distance attenuation weighted interpolation, extracts topological features through multi-level threshold scanning and active grid connectivity analysis, identifies pollution clumps based on the life cycle of the topological features in continuous threshold intervals, and determines pollution clumping generation events and splitting events according to the spatial overlap relationship of pollution clumps in adjacent time layers, forming a set of water quality field topological mutation events; Step S2: The cloud control platform calculates the water quality risk score of each water unit based on the water quality field topological mutation event and static water quality indicators, sets the target robot density, constructs a single-unit benefit function that includes water quality risk, density deviation penalty, and time and energy consumption terms, executes multiple rounds of mean field game iteration and adaptively adjusts the density penalty coefficient to make the actual robot density converge to the robot target density, and obtains the converged robot density distribution and water quality treatment priority level. Step S3: The cloud control platform generates sampling, dosing, and inspection tasks based on the converged robot density distribution and water quality treatment priority level, and determines the basic value of the tasks. Each robot calculates its bidding score under the constraints of remaining power and matching biological agents. The cloud control platform conducts a market auction according to the principle of maximizing the bidding score and assigns the generated tasks to the winning robot for execution.

3. The method of claim 2, wherein, Step S1 involves constructing a watershed grid pollution intensity field, which includes: setting up several fixed water quality monitoring stations along the main river channel and tributaries at all levels, and equipping a robot with a mobile monitoring unit to collect data on ammonia nitrogen, total phosphorus, dissolved oxygen, turbidity, and water temperature at preset sampling intervals, along with their temporal and spatial locations; projecting the watershed in two dimensions on a cloud control platform and dividing it into multiple watershed grid units at a given resolution; summarizing all original water quality samples within a time window at each time layer; searching for a set of neighboring samples within a defined distance for each watershed grid unit; calculating and standardizing the weighted average of the pollution indices of the samples based on Euclidean distance; and obtaining the grid pollution intensity value by weighting the pollution index of the samples. The pollution index is the weighted sum of ammonia nitrogen concentration and total phosphorus concentration.

4. The method according to claim 3, characterized in that, Step S1, topological feature extraction and lifecycle determination, includes: finding the minimum and maximum pollution intensity values ​​of all watershed grid cells in the same time layer, dividing the numerical range into multiple pollution intensity thresholds; marking grid cells with pollution intensity not lower than the current threshold as active grids under each threshold, and dividing connected regions based on the four-adjacency relationship of shared common boundaries; using the connected regions divided by the lowest threshold as initial topological features, and associating them under subsequent thresholds based on the grid overlap ratio between the current connected region and the region corresponding to the topological feature of the previous threshold; when the overlap ratio is not less than a preset ratio threshold, the same topological feature is considered to exist continuously; otherwise, a new topological feature is generated, and a stopping threshold is recorded when the overlap ratio between the current connected region and all connected regions is lower than the preset ratio threshold; the lifecycle length is determined by the difference between the birth threshold and the stopping threshold.

5. The method according to claim 4, characterized in that, Step S1, which involves identifying contaminant clumps and generating a set of water quality field topological mutation events, includes: marking topological features whose lifecycle length reaches a preset length threshold and whose number of active grids is not less than a preset number threshold under any threshold as contaminant clumps, and using the coverage area of ​​the active grids under the corresponding threshold at the midpoint of the lifecycle as the spatial range of the contaminant clump; calculating the intersection-union ratio (IUR) of the spatial ranges of contaminant clumps in adjacent time layers on multiple consecutive time layers; when the IUR of a certain contaminant clump in the current time layer with all contaminant clumps in the previous time layer is lower than a preset lower limit, recording the location of the current contaminant clump as a contaminant clump generation event; when the IUR of a certain contaminant clump in the previous time layer with at least two contaminant clumps in the current time layer is not lower than a preset upper limit, recording the contaminant clump corresponding to the previous time layer as a contaminant clump splitting event and associating it with the spatial range of subsequent contaminant clumps, thereby forming a set of water quality field topological mutation events.

6. The method according to claim 5, characterized in that, Step S2, which involves calculating the water quality risk score and setting the robot target density, includes: dividing the main river channel and tributaries of the basin into multiple water area units along the river centerline, with each water area unit corresponding to a section of river length and associated with a set of corresponding covered watershed grid units; calculating the average pollution intensity, pollution cluster coverage ratio, and the number of pollution cluster generation events and splitting events within a preset time window for each water area unit, and obtaining the water quality risk score by weighting and summing the average pollution intensity, pollution cluster coverage ratio, and event statistics; and then obtaining the robot target density by superimposing the basic density constant through a linear relationship based on the water quality risk score, with the robot target density measured in units of robots per square kilometer.

7. The method according to claim 6, characterized in that, The initialization of the mean-field game density iteration algorithm in step S2 includes: collecting the current location information and remaining power of all robots; assigning each robot to the corresponding water area unit according to its location on the river centerline; counting the number of robots in the water area unit and dividing by the water area estimated based on the river length and average river width to obtain the initial actual robot density; and performing mean-field game iteration in a preset number of rounds. In each round, a candidate water area unit set consisting of its current water area unit and adjacent upstream and downstream water areas is generated for each robot, and the estimated travel time and energy consumption are calculated based on the centerline distance between the center points of the water area units and the preset cruising speed.

8. The method according to claim 7, characterized in that, Step S2 involves performing multiple rounds of mean-field game iterations and adaptively adjusting the density penalty coefficient. This includes: calculating the individual reward for each robot and its candidate water area units. The individual reward is the sum of the water quality risk score of the water area unit minus the density penalty term and the travel time and energy consumption term. The density penalty term is obtained by multiplying the difference between the current actual robot density and the robot's target density by the density penalty coefficient. The individual rewards of the same robot in the candidate water area units are translated and normalized to obtain the migration probability vector. Based on this, the expected number of robots and the expected robot density of each water area unit are calculated, and the actual robot density is updated using the expected robot density. At the same time, the density penalty coefficient is adjusted in stages according to the density error range between the actual robot density and the robot's target density in each water area unit until the preset number of iterations is completed. The final actual robot density is used as the converged robot density distribution, and each water area unit is divided into at least three levels of water quality treatment priority according to the water quality risk score threshold.

9. The method according to claim 8, characterized in that, Step S3, task generation and basic value determination, includes: the cloud control platform generating a task set for each water area unit based on the converged robot density distribution table. The task set includes at least three types of tasks: sampling tasks, dosing tasks, and inspection tasks. Sampling tasks consist of several sampling points, each recording the sampling location coordinates and sampling time window; dosing tasks consist of several dosing points, each recording the dosing location coordinates, dosing time window, and biological agent type number; inspection tasks consist of several inspection paths, each represented by a series of continuous coordinate points and an inspection time window; and basic values ​​are set for tasks according to the water quality treatment priority level and task type of the water area unit. In high-level water quality treatment units, dosing tasks are assigned a higher basic value than sampling and inspection tasks.

10. The method according to claim 9, characterized in that, Step S3 involves calculating the bidding score and conducting a market auction based on the principle of maximizing the bidding score to allocate tasks to the winning robots. This includes: the cloud control platform calculating the travel time, time cost, and energy cost for each robot and each task combination within the current water unit based on the distance between the robot's current position and the task target position, and setting a resource coefficient based on the robot's remaining battery percentage; for delivery tasks, determining whether the type of biological agent carried by the robot matches the task requirements, amplifying the task's basic value by the first coefficient if they match, and reducing the task's basic value by the second coefficient if they do not match; for sampling and inspection tasks, the basic value after matching is equal to the task's basic value; based on this, the basic value after matching is multiplied by... The resource coefficient is subtracted from the time and energy costs to obtain the bidding score, and the bidding scores less than zero are truncated to zero. The cloud control platform sorts the tasks in the same water unit according to their basic value, and selects the tasks with the highest ranking as the current auction task. Among the robots that have not yet won the bid and whose bidding scores are greater than zero, the robot with the highest bidding score is selected as the winning robot. When there are multiple robots with the same bidding score, the one with the smallest estimated travel time is selected as the winning robot. The current task is bound to the winning robot until all tasks are allocated or all robots have obtained tasks. The task allocation result table of the whole watershed is obtained, which is used by the robot's local path planning module to generate task execution paths and complete sampling, dosing and inspection operations.

Images