A floating garbage cleaning control method based on inland waterway water surface track learning

Abstract

The application discloses a kind of based on inland waterway water surface trajectory learning's floating garbage cleaning control method, it is related to data analysis technical field, including: obtaining historical inland floating garbage multi-source heterogeneous data, constructs garbage track prediction model, predicts future each environment under inland floating garbage trajectory trend vector, real-time updates inland floating garbage density distribution;Obtain the multi-source heterogeneous data of water surface garbage cleaning robot, establish multi-ship cooperative path planning model, generate inland floating garbage water surface garbage cleaning robot cooperative path planning;Based on inland floating garbage water surface garbage cleaning robot cooperative path planning and real-time updates inland floating garbage density distribution, constructs optimal floating garbage cleaning trajectory control function, generates inland floating garbage water surface garbage cleaning robot adaptive cleaning task.The application has the advantages that: improve the efficiency and coverage of cleaning operation, significantly reduce the cost of human supervision.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a method for controlling the removal of floating debris based on learning the trajectory of water surfaces in inland waterways. Background Technology

[0002] Existing floating debris cleanup control methods based on inland waterway surface trajectory learning rely excessively on offline learning from historical trajectory data, lacking online optimization capabilities for dynamic debris distribution and multi-agent collaboration. They treat debris merely as static obstacles or simple moving targets, failing to establish dynamic prediction models driven by water flow and wind direction, resulting in planned paths lagging behind environmental changes. Furthermore, their learning models are mostly based on single-ship decision-making, which cannot effectively resolve task allocation conflicts and path collisions in multi-ship collaborative scenarios, leading to low cleanup efficiency, overlapping coverage areas, and internal consumption of system resources. Summary of the Invention

[0003] To address the aforementioned technical problems, a floating debris cleanup and control method based on inland waterway surface trajectory learning is provided. This technical solution solves the problems mentioned above.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A floating debris cleanup and control method based on inland waterway surface trajectory learning includes: S1. Obtain multi-source heterogeneous data of floating garbage in historical inland waterways, initialize the density distribution of floating garbage in inland waterways and external environmental factors as influencing factors, construct a garbage trajectory prediction model, predict the trajectory trend vector of floating garbage in inland waterways under various environments in the future, and update the density distribution of floating garbage in inland waterways in real time. S2. Obtain multi-source heterogeneous data of the water surface garbage cleaning robot, initialize the floating garbage cleaning capability of the water surface garbage cleaning robot and the trajectory trend vector of floating garbage in inland waterways under various future environments, establish a multi-vehicle collaborative path planning model, and generate collaborative path planning for the water surface garbage cleaning robot for floating garbage in inland waterways. S3. Based on the collaborative path planning and real-time updating of floating garbage density distribution in inland waterways, the robot generates real-time updated floating garbage cleaning task priorities, constructs the optimal floating garbage cleaning trajectory control function, and generates adaptive cleaning tasks for the floating garbage cleaning robot in inland waterways.

[0005] Preferably, step S1 specifically includes: Acquire spatial data of inland waterways, randomly divide the waterways into several regions of the same size, and construct a discretized network diagram of inland waterways; Based on multi-source heterogeneous data of floating garbage in historical inland waterways, a dynamic spatiotemporal graph of the inland waterway discretization network is constructed, using historical garbage density, water flow vector data, and wind direction and speed as attributes of each grid node in the inland waterway discretization network graph. Using Euclidean distance, the matching degree between the attribute flow vectors of each adjacent grid node in the dynamic spatiotemporal graph of the inland waterway discretization network is calculated and normalized. Then, the attribute weights of each adjacent grid node are assigned to construct the dynamic weighted spatiotemporal graph of the inland waterway discretization network. Based on the dynamically weighted spatiotemporal graph of the discretized network of inland waterways, an STGCN graph convolutional spatiotemporal neural network is trained to construct a garbage trajectory prediction model. The time series of attributes of each grid node in the dynamically weighted spatiotemporal graph of the discretized network of inland waterways is used as input. The graph convolutional temporal layer captures the density trend dependency of floating garbage density in inland waterways over time, and substitutes it into the graph convolutional spatial layer to aggregate the neighbor information of the attribute time series of each grid node. Through multi-task output, a density convolutional prediction head and a trend convolutional prediction head are output, which output the future floating garbage density scalar value and the future floating garbage movement trend vector of each grid node in the dynamically weighted spatiotemporal graph of the discretized network of inland waterways.

[0006] Preferably, step S1 further includes: According to multiple linear regression, the attribute of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways is used as the input of independent variables, and the actual floating garbage density value of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways is used as the output of dependent variables. The least squares method is used to fit the contribution weight of each grid node attribute in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways to the actual floating garbage density value of each grid node. The contribution weight of each grid node attribute to the actual floating debris density value of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network is weighted and fused with the future floating debris density scalar value and the future floating debris movement trend vector of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network to obtain the trajectory trend vector of floating debris in the inland waterway under various future environments in the dynamic weighted spatiotemporal graph of the inland waterway discretized network. Using a sliding window, with the actual unit as the observation window, and taking the trajectory trend vector of floating garbage in the inland waterway under various future environments in the dynamic weighted spatiotemporal graph of the inland waterway discretized network as the observation object, a real-time updated distribution of floating garbage density in the inland waterway is established.

[0007] Preferably, step S2 specifically includes: Based on multi-source heterogeneous data from the surface garbage cleaning robot, the data is standardized and spliced ​​according to the GPS / IMU module, remaining battery capacity and garbage bin weight of the surface garbage cleaning robot to obtain the initial floating garbage cleaning capability of the surface garbage cleaning robot. Based on the trajectory trend vector of floating garbage in inland waterways under various future environments and the real-time updated density distribution of floating garbage in inland waterways in the dynamic weighted spatiotemporal map of the discretized network of inland waterways, an advection diffusion equation is established and solved using the finite element method to obtain the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal map of the discretized network of inland waterways. Using cosine similarity, the matching degree between the thermal value of floating garbage distribution of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways and the floating garbage cleaning capacity of the initial floating garbage cleaning robot is calculated to obtain the initial screening set of floating garbage cleaning robots for inland waterways.

[0008] Preferably, step S2 further includes: Based on the partially observable Markov decision process of POMDP, a multi-vessel collaborative path planning model is established by using the initial screening set of floating garbage cleaning robots in inland waterways as the decision-making device set, the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network as the state space, the selection of decision-making device set by each grid node as the joint action space, the multi-source heterogeneous data of the decision-making device set as the actual state, the maximum observation range of the decision-making device set as the local situation map, and the joint observation range of the decision-making device set as the interaction information of the floating garbage cleaning process between the low-resolution global graph and the real-time decision-making device set. Based on the multi-ship collaborative path planning model, the belief update scheme uses the joint observation range low-resolution global map of the decision-making equipment set as the low-resolution global map observation belief, updates the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal map of the inland waterway discretized network as the real-time belief of the state space, and obtains the joint observation range of the updated decision-making equipment set as the low-resolution global map. Using the joint observation range of the updated decision-making device set as a low-resolution global graph, and employing a tree strategy, the joint action space to be decided in the state space is used as the node, and the specified number of times the decision-making device is selected in the joint action space to be decided in each state space is used as the leaf node to generate the joint action sequence of the water surface garbage cleaning robot for floating garbage in inland waterways, thus obtaining the collaborative path planning of the water surface garbage cleaning robot for floating garbage in inland waterways.

[0009] Preferably, step S3 specifically includes: Based on MARL multi-agent reinforcement learning, this study proposes a method to minimize the cleaning time of floating debris in inland waterways by using collaborative path planning for floating debris cleaning robots and real-time updates of floating debris density distribution. It also addresses the uniqueness of overlapping areas in the collaborative path planning regions and maximizes the battery capacity of the collaborative path planning to satisfy the real-time updates of floating debris density distribution in the corresponding regions. Finally, it establishes an optimal competitive reward function for floating debris cleaning and generates a real-time updated priority for floating debris cleaning tasks. Based on the MADDPG framework, and according to the real-time updated priority of floating garbage cleaning tasks in inland waterways, the collaborative path planning of floating garbage cleaning robots in inland waterways is verified to execute the interaction information of floating garbage cleaning process between the real-time updated priority of floating garbage cleaning tasks and the set of real-time decision-making devices. Dynamic weights are dynamically assigned to the collaborative path planning of floating garbage cleaning robots in inland waterways using an attention-based mechanism.

[0010] Preferably, step S3 further includes: Based on MPC model predictive control, the system uses the real-time update of floating garbage density distribution in inland waterways as the task state of the set of decision-making equipment, the optimal competitive reward function for floating garbage cleaning as the task decision desire, and the dynamic weight of the collaborative path planning of the floating garbage cleaning robot in inland waterways as the task decision coordination factor. The system constructs the optimal floating garbage cleaning trajectory control function and generates an adaptive cleaning task for the floating garbage cleaning robot in inland waterways.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a floating debris cleaning control scheme based on inland waterway trajectory learning. By constructing a fully autonomous collaborative decision-making system integrating perception, prediction, planning, and control, it achieves a leap from passive response to proactive intelligent management of floating debris cleaning in inland waterways. The beneficial effects are as follows: the system can deeply integrate multi-source heterogeneous data, accurately predict the dynamic spatiotemporal distribution of debris, and generate collaboratively optimal and dynamically adaptive cleaning paths for multiple cleaning robots; it improves the efficiency and coverage of cleaning operations, avoids resource waste and repetitive labor, and significantly reduces the cost of human supervision by endowing robots with the ability to make autonomous decisions and coordinate execution in complex and changing environments. Attached Figure Description

[0012] Figure 1 This is a flowchart of a floating debris cleanup and control method based on inland waterway surface trajectory learning. Detailed Implementation

[0013] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0014] Reference Figure 1 As shown, a floating debris cleanup and control method based on inland waterway surface trajectory learning includes: S1. Obtain multi-source heterogeneous data of floating garbage in historical inland waterways, initialize the density distribution of floating garbage in inland waterways and external environmental factors as influencing factors, construct a garbage trajectory prediction model, predict the trajectory trend vector of floating garbage in inland waterways under various environments in the future, and update the density distribution of floating garbage in inland waterways in real time. Step S1 specifically includes: Acquire spatial data of inland waterways, randomly divide the waterways into several regions of the same size, and construct a discretized network diagram of inland waterways; Based on multi-source heterogeneous data of floating garbage in historical inland waterways, a dynamic spatiotemporal graph of the inland waterway discretization network is constructed, using historical garbage density, water flow vector data, and wind direction and speed as attributes of each grid node in the inland waterway discretization network graph. Using Euclidean distance, the matching degree between the attribute flow vectors of each adjacent grid node in the dynamic spatiotemporal graph of the inland waterway discretization network is calculated and normalized. Then, the attribute weights of each adjacent grid node are assigned to construct the dynamic weighted spatiotemporal graph of the inland waterway discretization network. Based on the dynamically weighted spatiotemporal graph of the discretized network of inland waterways, an STGCN graph convolutional spatiotemporal neural network is trained to construct a garbage trajectory prediction model. The time series of attributes of each grid node in the dynamically weighted spatiotemporal graph of the discretized network of inland waterways is used as input. The graph convolutional temporal layer captures the density trend dependency of floating garbage density in inland waterways over time, and substitutes it into the graph convolutional spatial layer to aggregate the neighbor information of the attribute time series of each grid node. Through multi-task output, a density convolutional prediction head and a trend convolutional prediction head are output, which output the future floating garbage density scalar value and the future floating garbage movement trend vector of each grid node in the dynamically weighted spatiotemporal graph of the discretized network of inland waterways.

[0015] Step S1 also includes: According to multiple linear regression, the attribute of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways is used as the input of independent variables, and the actual floating garbage density value of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways is used as the output of dependent variables. The least squares method is used to fit the contribution weight of each grid node attribute in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways to the actual floating garbage density value of each grid node. The contribution weight of each grid node attribute to the actual floating debris density value of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network is weighted and fused with the future floating debris density scalar value and the future floating debris movement trend vector of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network to obtain the trajectory trend vector of floating debris in the inland waterway under various future environments in the dynamic weighted spatiotemporal graph of the inland waterway discretized network. Using a sliding window, with the actual unit as the observation window, and taking the trajectory trend vector of floating garbage in the inland waterway under various future environments in the dynamic weighted spatiotemporal graph of the inland waterway discretized network as the observation object, a real-time updated distribution of floating garbage density in the inland waterway is established.

[0016] When using it, please refer to the steps outlined above: As a further development, by discretizing the river channel into a grid structure and integrating multi-source data such as historical waste density, water flow, and wind direction to construct a dynamic spatiotemporal map, a graph convolutional neural network is used to accurately capture the spatiotemporal evolution of waste diffusion with water flow. Innovatively, Euclidean distance is used to quantify the hydraulic correlation between grids, combined with multiple linear regression to correct environmental factor weights, achieving a highly efficient fusion of physical mechanisms and data-driven approaches. This significantly improves the accuracy and real-time performance of waste distribution prediction, providing cleaning robots with refined trajectory guidance with clear physical meaning, effectively enhancing the dynamic response capability and system robustness of comprehensive cleaning operations.

[0017] The following implementation example is proposed for step S1 above: Obtain geographic information data of inland waterways and divide the waterway area into a uniform grid of 20 meters by 20 meters. Construct a discrete network diagram of the waterways, where nodes correspond to the waterway grid.

[0018] Historical observation data from the past 24 hours is loaded, including garbage density records for each grid, water flow vectors measured by ultrasonic flow meters, and wind direction and speed provided by weather stations. This data is used as attributes of the corresponding grid nodes to construct a dynamic spatiotemporal map of the river channel.

[0019] Calculate the Euclidean distance between the flow vectors of adjacent grid nodes, normalize the distance values ​​to a weight between zero and one, and use this weight to construct the adjacency matrix of the dynamic weighted spatiotemporal graph of the river channel.

[0020] Initialize an STGCN graph convolutional spatiotemporal neural network, training it with a dynamically weighted spatiotemporal graph sequence from the past twelve hours. The network input consists of the time-series attributes of each grid node. The graph convolutional temporal layer learns the pattern of garbage density change over time, while the graph convolutional spatial layer aggregates the spatial information of each node's neighbors. The network output layer contains two parallel convolutional heads: a density prediction head that outputs a scalar value of garbage density for each grid in the next hour, and a trend prediction head that outputs a garbage movement trend vector for each grid in the next hour. The trained network is the garbage trajectory prediction model.

[0021] Multiple linear regression analysis was performed, with the independent variables being the grid node attributes of water flow vector, wind direction and wind speed, and the dependent variable being the actual waste density value of the grid node. The least squares method was used to calculate the contribution weight of each environmental attribute to waste density.

[0022] The future waste density scalar value output by the STGCN model is weighted and fused with the future waste movement trend vector. The weights are derived from the contribution weights obtained from multiple linear regression analysis. The fusion result generates the floating waste trajectory trend vector for each grid in the future environment.

[0023] A 30-minute sliding observation window is set, sliding every five minutes. The latest real-time observation data is input into the prediction process, and the trajectory trend vector calculation step is repeated. This process achieves real-time updates of the density distribution of floating debris in the river channel.

[0024] S2. Obtain multi-source heterogeneous data of the water surface garbage cleaning robot, initialize the floating garbage cleaning capability of the water surface garbage cleaning robot and the trajectory trend vector of floating garbage in inland waterways under various future environments, establish a multi-vehicle collaborative path planning model, and generate collaborative path planning for the water surface garbage cleaning robot for floating garbage in inland waterways. Step S2 specifically includes: Based on multi-source heterogeneous data from the surface garbage cleaning robot, the data is standardized and spliced ​​according to the GPS / IMU module, remaining battery capacity and garbage bin weight of the surface garbage cleaning robot to obtain the initial floating garbage cleaning capability of the surface garbage cleaning robot. Based on the trajectory trend vector of floating garbage in inland waterways under various future environments and the real-time updated density distribution of floating garbage in inland waterways in the dynamic weighted spatiotemporal map of the discretized network of inland waterways, an advection diffusion equation is established and solved using the finite element method to obtain the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal map of the discretized network of inland waterways. Using cosine similarity, the matching degree between the thermal value of floating garbage distribution of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways and the floating garbage cleaning capacity of the initial floating garbage cleaning robot is calculated to obtain the initial screening set of floating garbage cleaning robots for inland waterways.

[0025] Step S2 also includes: Based on the partially observable Markov decision process of POMDP, a multi-vessel collaborative path planning model is established by using the initial screening set of floating garbage cleaning robots in inland waterways as the decision-making device set, the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network as the state space, the selection of decision-making device set by each grid node as the joint action space, the multi-source heterogeneous data of the decision-making device set as the actual state, the maximum observation range of the decision-making device set as the local situation map, and the joint observation range of the decision-making device set as the interaction information of the floating garbage cleaning process between the low-resolution global graph and the real-time decision-making device set. Based on the multi-ship collaborative path planning model, the belief update scheme uses the joint observation range low-resolution global map of the decision-making equipment set as the low-resolution global map observation belief, updates the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal map of the inland waterway discretized network as the real-time belief of the state space, and obtains the joint observation range of the updated decision-making equipment set as the low-resolution global map. Using the joint observation range of the updated decision-making device set as a low-resolution global graph, and employing a tree strategy, the joint action space to be decided in the state space is used as the node, and the specified number of times the decision-making device is selected in the joint action space to be decided in each state space is used as the leaf node to generate the joint action sequence of the water surface garbage cleaning robot for floating garbage in inland waterways, thus obtaining the collaborative path planning of the water surface garbage cleaning robot for floating garbage in inland waterways.

[0026] When using it, please refer to the steps outlined above: As a further development, the study achieves initial task screening by quantifying the matching degree between robot cleaning capabilities and waste distribution heatmaps, and accurately predicts waste evolution trends based on advection diffusion equations. The POMDP framework is used to model collaborative path planning as a sequential decision-making problem under partially observable conditions. Belief state updates are integrated with local observation information from multiple robots, and a tree search strategy is then used to generate the long-term optimal joint action sequence. This demonstrates the foresight and adaptability of multi-vessel surface cleaning robot collaboration, achieving efficient matching of cleaning resources and dynamic tasks in uncertain environments, effectively avoiding repetitive operations and ensuring full-area coverage efficiency.

[0027] The following implementation example is proposed for step S2 above: This process collects GPS / IMU coordinates, remaining battery capacity, and debris bin weight data from the water surface debris cleaning robot. Z-score normalization is performed, and the normalized data vectors are concatenated to output the robot's initial floating debris cleaning capability.

[0028] The advection diffusion equation is solved using the finite element method (FEM) as the input parameters: the trajectory trend vector of floating debris in inland waterways under future conditions and the real-time updated debris density distribution. The output is the thermodynamic value of floating debris distribution for each grid node in the dynamically weighted spatiotemporal graph of the discretized network of the inland waterway.

[0029] Calculate the cosine similarity. Vector A represents the thermal values ​​of floating debris distribution at the grid nodes. Vector B represents the robot's initial floating debris cleaning capability. Based on the similarity threshold, divide the robot set to generate an initial screening set of robots for cleaning floating debris in inland waterways.

[0030] A POMDP model is established, where the decision-making device set is defined as the set of robots initially selected. The state space is set as a heatmap of floating debris distribution on a grid of nodes. The joint action space is defined as the set of decision-making devices selected by each grid node. The actual state originates from multi-source heterogeneous data of the decision-making device set, and the boundary of the local situation map is defined by the maximum observation range of a single robot. The low-resolution global map is composed of the joint observation range of the decision-making device set.

[0031] The belief update is performed, taking the observed beliefs from a low-resolution global graph as input, and using Bayesian filtering as the update rule. The output is the updated real-time beliefs in the state space, i.e., the corrected heatmap of floating debris distribution.

[0032] The tree strategy search is run, and the joint action space to be decided is set as the nodes of the tree structure. The leaf nodes are defined as the number of times the specified decision-making device makes a choice in a given state. The search algorithm outputs the robot's joint action sequence, which is the collaborative path planning of the water surface garbage cleaning robot for floating garbage in inland waterways.

[0033] S3. Based on the collaborative path planning and real-time updating of floating garbage density distribution in inland waterways, the robot generates real-time updated floating garbage cleaning task priorities, constructs the optimal floating garbage cleaning trajectory control function, and generates adaptive cleaning tasks for floating garbage in inland waterways. Step S3 specifically includes: Based on MARL multi-agent reinforcement learning, this study proposes a method to minimize the cleaning time of floating debris in inland waterways by using collaborative path planning for floating debris cleaning robots and real-time updates of floating debris density distribution. It also addresses the uniqueness of overlapping areas in the collaborative path planning regions and maximizes the battery capacity of the collaborative path planning to satisfy the real-time updates of floating debris density distribution in the corresponding regions. Finally, it establishes an optimal competitive reward function for floating debris cleaning and generates a real-time updated priority for floating debris cleaning tasks. Based on the MADDPG framework, and according to the real-time updated priority of floating garbage cleaning tasks in inland waterways, the collaborative path planning of floating garbage cleaning robots in inland waterways is verified to execute the interaction information of floating garbage cleaning process between the real-time updated priority of floating garbage cleaning tasks and the set of real-time decision-making devices. Dynamic weights are dynamically assigned to the collaborative path planning of floating garbage cleaning robots in inland waterways using an attention-based mechanism.

[0034] Step S3 also includes: Based on MPC model predictive control, the system uses the real-time update of floating garbage density distribution in inland waterways as the task state of the set of decision-making equipment, the optimal competitive reward function for floating garbage cleaning as the task decision desire, and the dynamic weight of the collaborative path planning of the floating garbage cleaning robot in inland waterways as the task decision coordination factor. The system constructs the optimal floating garbage cleaning trajectory control function and generates an adaptive cleaning task for the floating garbage cleaning robot in inland waterways.

[0035] When using it, please refer to the steps outlined above: As a further development, a three-layer architecture of competitive reward, attention coordination, and model predictive control is utilized. A competitive reward function that integrates cleaning efficiency, regional exclusivity, and power constraints is constructed through multi-agent reinforcement learning to drive the dynamic generation of task priorities. The MADDPG framework based on the attention mechanism quantifies the collaborative utility among agents to achieve adaptive allocation of decision weights. Model predictive control transforms high-level instructions into rolling optimization trajectories that satisfy dynamic constraints, realizing multi-objective collaborative optimization and dynamic environmental adaptation, improving cleaning efficiency while ensuring real-time response to floating garbage in inland waterways.

[0036] The following implementation example is proposed for step S3 above: Initialize the MARL environment. Define the state space as the real-time updated density distribution of floating debris in inland waterways. Define the action space as the path planning sequence for the water surface debris cleaning robot. Set the optimization objective: minimize the cleaning time of the cooperative path planning. Apply constraints: ensure the uniqueness and non-overlapping of the robot's planned area. Apply further constraints: ensure the robot's battery capacity supports the cleaning task in the corresponding area. Design a competitive reward function, with the function value correlated with the degree to which the above objective and constraints are satisfied. Train the MARL network and output the real-time updated priority of the inland waterway floating debris cleaning task.

[0037] The MADDPG algorithm framework is deployed, with real-time updated task priorities as input parameters. The cumulative benefit of prioritizing task execution under collaborative robot path planning is evaluated, and real-time interaction data on the floating debris cleaning process among the decision-making devices is collected. An attention mechanism is introduced, and the calculation mechanism assigns weight coefficients based on the policy relevance between robots. The output is a dynamic weight set for the collaborative path planning of the water surface debris cleaning robots.

[0038] An MPC controller is constructed, with the system state variable defined as the real-time updated density distribution of floating debris in the inland waterway. The objective function embeds a competitive reward function for optimal floating debris cleaning. Coordination constraints are introduced with dynamic weights for robot cooperative path planning. A finite-time optimal control problem is solved to generate the optimal floating debris cleaning trajectory control function. The output of the executed control function completes the adaptive cleaning task of the floating debris cleaning robot in the inland waterway.

[0039] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for controlling floating debris cleanup based on inland waterway surface trajectory learning, characterized in that, include: S1. Obtain multi-source heterogeneous data of floating garbage in historical inland waterways, initialize the density distribution of floating garbage in inland waterways and external environmental factors as influencing factors, construct a garbage trajectory prediction model, predict the trajectory trend vector of floating garbage in inland waterways under various environments in the future, and update the density distribution of floating garbage in inland waterways in real time. S2. Obtain multi-source heterogeneous data of the water surface garbage cleaning robot, initialize the floating garbage cleaning capability of the water surface garbage cleaning robot and the trajectory trend vector of floating garbage in inland waterways under various future environments, establish a multi-vehicle collaborative path planning model, and generate collaborative path planning for the water surface garbage cleaning robot for floating garbage in inland waterways. S3. Based on the collaborative path planning and real-time updating of floating garbage density distribution in inland waterways, the robot generates real-time updated floating garbage cleaning task priorities, constructs the optimal floating garbage cleaning trajectory control function, and generates adaptive cleaning tasks for the floating garbage cleaning robot in inland waterways.

2. The floating debris cleaning and control method based on inland waterway surface trajectory learning according to claim 1, characterized in that, Step S1 specifically includes: Acquire spatial data of inland waterways, randomly divide the waterways into several regions of the same size, and construct a discretized network diagram of inland waterways; Based on multi-source heterogeneous data of floating garbage in historical inland waterways, a dynamic spatiotemporal graph of the inland waterway discretization network is constructed, using historical garbage density, water flow vector data, and wind direction and speed as attributes of each grid node in the inland waterway discretization network graph. Using Euclidean distance, the matching degree between the attribute flow vectors of each adjacent grid node in the dynamic spatiotemporal graph of the inland waterway discretization network is calculated and normalized. Then, the attribute weights of each adjacent grid node are assigned to construct the dynamic weighted spatiotemporal graph of the inland waterway discretization network. Based on the dynamically weighted spatiotemporal graph of the discretized network of inland waterways, an STGCN graph convolutional spatiotemporal neural network is trained to construct a garbage trajectory prediction model. The time series of attributes of each grid node in the dynamically weighted spatiotemporal graph of the discretized network of inland waterways is used as input. The graph convolutional temporal layer captures the density trend dependency of floating garbage density in inland waterways over time, and substitutes it into the graph convolutional spatial layer to aggregate the neighbor information of the attribute time series of each grid node. Through multi-task output, a density convolutional prediction head and a trend convolutional prediction head are output, which output the future floating garbage density scalar value and the future floating garbage movement trend vector of each grid node in the dynamically weighted spatiotemporal graph of the discretized network of inland waterways.

3. The floating debris cleaning and control method based on inland waterway surface trajectory learning according to claim 2, characterized in that, Step S1 also includes: According to multiple linear regression, the attribute of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways is used as the input of independent variables, and the actual floating garbage density value of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways is used as the output of dependent variables. The least squares method is used to fit the contribution weight of each grid node attribute in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways to the actual floating garbage density value of each grid node. The contribution weight of each grid node attribute to the actual floating debris density value of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network is weighted and fused with the future floating debris density scalar value and the future floating debris movement trend vector of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network to obtain the trajectory trend vector of floating debris in the inland waterway under various future environments in the dynamic weighted spatiotemporal graph of the inland waterway discretized network. Using a sliding window and taking a unit time as the observation window, the trend vector of floating garbage trajectory in the inland waterway under various future environments in the dynamic weighted spatiotemporal graph of the inland waterway discretized network is used as the observation object to establish a real-time updated distribution of floating garbage density in the inland waterway.

4. The floating debris cleaning and control method based on inland waterway surface trajectory learning according to claim 3, characterized in that, Step S2 specifically includes: Based on multi-source heterogeneous data from the surface garbage cleaning robot, the data is standardized and spliced ​​according to the GPS / IMU module, remaining battery capacity and garbage bin weight of the surface garbage cleaning robot to obtain the initial floating garbage cleaning capability of the surface garbage cleaning robot. Based on the trajectory trend vector of floating garbage in inland waterways under various future environments and the real-time updated density distribution of floating garbage in inland waterways in the dynamic weighted spatiotemporal map of the discretized network of inland waterways, an advection diffusion equation is established and solved using the finite element method to obtain the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal map of the discretized network of inland waterways. Using cosine similarity, the matching degree between the thermal value of floating garbage distribution of each grid node in the dynamic weighted spatiotemporal graph of the discretized network of inland waterways and the floating garbage cleaning capacity of the initial floating garbage cleaning robot is calculated to obtain the initial screening set of floating garbage cleaning robots for inland waterways.

5. The floating debris cleaning and control method based on inland waterway surface trajectory learning according to claim 4, characterized in that, Step S2 also includes: Based on the partially observable Markov decision process of POMDP, a multi-vessel collaborative path planning model is established by using the initial screening set of floating garbage cleaning robots in inland waterways as the decision-making device set, the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal graph of the inland waterway discretized network as the state space, the selection of decision-making device set by each grid node as the joint action space, the multi-source heterogeneous data of the decision-making device set as the actual state, the maximum observation range of the decision-making device set as the local situation map, and the joint observation range of the decision-making device set as the interaction information of the floating garbage cleaning process between the low-resolution global graph and the real-time decision-making device set. Based on the multi-ship collaborative path planning model, the belief update scheme uses the joint observation range low-resolution global map of the decision-making equipment set as the low-resolution global map observation belief, updates the floating garbage distribution heat map of each grid node in the dynamic weighted spatiotemporal map of the inland waterway discretized network as the real-time belief of the state space, and obtains the joint observation range of the updated decision-making equipment set as the low-resolution global map. Using the joint observation range of the updated decision-making device set as a low-resolution global graph, and employing a tree strategy, the joint action space to be decided in the state space is used as the node, and the specified number of times the decision-making device is selected in the joint action space to be decided in each state space is used as the leaf node to generate the joint action sequence of the water surface garbage cleaning robot for floating garbage in inland waterways, thus obtaining the collaborative path planning of the water surface garbage cleaning robot for floating garbage in inland waterways.

6. The floating debris cleaning and control method based on inland waterway surface trajectory learning according to claim 5, characterized in that, Step S3 specifically includes: Based on MARL multi-agent reinforcement learning, this study proposes a method to minimize the cleaning time of floating debris in inland waterways by using collaborative path planning for floating debris cleaning robots and real-time updates of floating debris density distribution. It also addresses the uniqueness of overlapping areas in the collaborative path planning regions and maximizes the battery capacity of the collaborative path planning to satisfy the real-time updates of floating debris density distribution in the corresponding regions. Finally, it establishes an optimal competitive reward function for floating debris cleaning and generates a real-time updated priority for floating debris cleaning tasks. Based on the MADDPG framework, and according to the real-time updated priority of floating garbage cleaning tasks in inland waterways, the collaborative path planning of floating garbage cleaning robots in inland waterways is verified to execute the interaction information of floating garbage cleaning process between the real-time updated priority of floating garbage cleaning tasks and the set of real-time decision-making devices. Dynamic weights are dynamically assigned to the collaborative path planning of floating garbage cleaning robots in inland waterways using an attention-based mechanism.

7. The floating debris cleaning and control method based on inland waterway surface trajectory learning according to claim 6, characterized in that, Step S3 also includes: Based on MPC model predictive control, the system uses the real-time update of floating garbage density distribution in inland waterways as the task state of the set of decision-making equipment, the optimal competitive reward function for floating garbage cleaning as the task decision desire, and the dynamic weight of the collaborative path planning of the floating garbage cleaning robot in inland waterways as the task decision coordination factor. The system constructs the optimal floating garbage cleaning trajectory control function and generates an adaptive cleaning task for the floating garbage cleaning robot in inland waterways.

Images