Unmanned aerial vehicle stain cleaning cooperative control system

Abstract

The invention discloses an unmanned aerial vehicle stain cleaning cooperative control system. According to the invention, real-time linkage of virtual simulation and actual operation is realized through the digital twin collaborative cleaning sand table module, and the collaborative efficiency of a multi-unmanned aerial vehicle cluster is significantly improved. The digital twin sand table can dynamically map unmanned aerial vehicle states, environmental data and stain distribution, rehearses task allocation and path planning in a virtual space, discovers and eliminates path conflicts among unmanned aerial vehicles in advance, and avoids the problem of repeated cleaning or missing cleaning caused by information lag in traditional operation. And meanwhile, the sand table adjusts sub-task allocation based on a real-time state, and preferentially allocates a dense dirty area to the unmanned aerial vehicle with sufficient electric quantity, thereby reducing air flight time and invalid energy consumption, and ensuring that the whole operation process is more coherent and the cleaning coverage is more complete. Intelligent and self-adaptive adjustment of the cleaning process is achieved through deep cooperation of multiple modules, and the cleaning effect and the resource utilization rate in a complex scene are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of unmanned aerial vehicle cleaning, and particularly relates to a collaborative control system for cleaning stains of unmanned aerial vehicles. BACKGROUND

[0002] The unmanned aerial vehicle stain cleaning system is an intelligent cleaning device based on unmanned aerial vehicle technology and equipped with high-pressure spraying, brushing or dry ice cleaning devices, and is specially used for cleaning stains on the surfaces of high-altitude or dangerous areas such as building outer walls, glass curtain walls, photovoltaic panels and viaducts that are difficult to access manually. The system realizes accurate positioning and efficient cleaning through autonomous flight, path planning and visual recognition technology, greatly improves the operation safety and reduces the labor cost, and meets various cleaning needs in complex environments. The unmanned aerial vehicle stain cleaning system has the characteristics of flexible operation, wide coverage and high cleaning efficiency, and has a wide application prospect in the fields of urban construction, energy facility maintenance and industrial equipment maintenance, and is an important technical support for the development of modern smart cities and intelligent manufacturing.

[0003] However, the prior art lacks a real-time linkage mechanism between virtual simulation and actual operation when multiple unmanned aerial vehicles are used for collaborative cleaning, and it is difficult to dynamically adjust task allocation and path planning, which may cause path conflicts between unmanned aerial vehicles or uneven resource allocation, resulting in repeated cleaning, missed cleaning or increased energy consumption of empty flight, and thus the overall operation efficiency and cleaning effect stability are insufficient. SUMMARY

[0004] The purpose of the present application is to solve the above-mentioned problems, and to provide a collaborative control system for cleaning stains of unmanned aerial vehicles.

[0005] The technical scheme adopted by the present application is as follows: a collaborative control system for cleaning stains of unmanned aerial vehicles, comprising: a multi-source heterogeneous sensor perception module, a stain feature analysis and cleaning strategy generation module, a digital twin collaborative cleaning sand table module, a distributed task allocation and path planning module, a multi-machine communication and collaborative control module, and a cleaning execution and feedback module.

[0006] The digital twin collaborative cleaning sand table module is internally provided with a physical world mapping unit, a virtual simulation engine and a multi-agent collaborative decision-making model.

[0007] The multi-source heterogeneous sensor perception module is connected to the stain feature analysis and cleaning strategy generation module and the digital twin collaborative cleaning sand table module.

[0008] The stain feature analysis and cleaning strategy generation module is connected to the cleaning execution and feedback module and the input end of the digital twin collaborative cleaning sand table module.

[0009] The digital twin collaborative cleaning sand table module is connected to the distributed task allocation and path planning module and receives the input end of the cleaning execution and feedback module data.

[0010] The distributed task allocation and path planning module is connected to the input of the multi-machine communication and collaborative control module;

[0011] The multi-machine communication and collaborative control module is connected to the input terminals of the cleaning execution and feedback module and the digital twin collaborative cleaning sandbox module;

[0012] The cleaning execution and feedback module is connected to the input terminals of the stain feature analysis and cleaning strategy generation module and the digital twin collaborative cleaning sandbox module.

[0013] In a preferred embodiment, the multi-source heterogeneous sensor sensing module internally includes: a multimodal sensor unit, a data preprocessing unit, and an edge computing node;

[0014] The multimodal sensor unit consists of an RGB-D camera, an infrared thermal imager, an ultrasonic sensor, and a water quality sensor. Its functions include acquiring visual features of stains, distinguishing temperature-sensitive stains, detecting stain thickness and adhesion, and monitoring residual ion concentration after cleaning. The data preprocessing unit consists of filtering and image segmentation algorithms. Its functions include removing noise from sensor data and extracting feature parameters such as stain contours and grayscale values. The edge computing node consists of an embedded chip. Its function is to process sensor data in real time on the drone, reducing cloud transmission latency and controlling data processing latency to within 50ms.

[0015] In a preferred embodiment, the stain feature analysis and cleaning strategy generation module uses a CNN-LSTM hybrid network to transform the input "stain feature vector" into customized cleaning parameters: water pressure, brush head speed, and cleaning agent concentration, and dynamically optimizes them in combination with historical cleaning results;

[0016] The input layer consists of: The input is the preprocessed stain feature vector: X = [x1, x2, ..., x...]. n ], where x1~x n The physical characteristic parameters of the stain are: x1 = stain area (㎡), x2 = gray value, x3 = hardness (N), x4 = adhesion (Pa), x5 = ambient temperature (°C), etc., n = 8~12. To eliminate dimensional differences, each feature is mapped to the range [0,1] through standardization.

[0017] To incorporate historical cleaned data, the spatial feature vector F output by the CNN is... CNN Concatenate with the historical time series: H = [h1, h2, h3], h k The parameter-effect pair for the k-th historical cleaning, including water pressure, rotation speed, concentration, and corresponding residual rate, is concatenated and input into an LSTM network. The LSTM learns temporal dependencies through a gating mechanism and outputs a 32-dimensional vector F that integrates spatiotemporal features.LSTM .

[0018] F LSTM Input 2-layer fully connected network, hidden layer neuron number 64→32, activation function is ReLU, the final output initial cleaning parameters:

[0019] Water pressure P0 (unit: MPa), brush head speed R0 (unit: rpm), cleaning agent concentration C0 (unit: %), the three constitute the initial parameter vector θ0=[P0, R0, C0].

[0020] The dynamic optimizer of the stain feature analysis and cleaning strategy generation module introduces a reinforcement learning mechanism, takes the historical cleaning residual rate r (r=0 represents complete cleaning, and r=1 represents no cleaning) as a reward signal, and corrects the initial parameters. Specifically, when the average residual rate of the historical same type of stains is , the parameters are adjusted by the policy gradient:

[0021] If is higher (insufficient cleaning), the water pressure or concentration is increased; if is lower (excessive cleaning wastes resources), the parameters are reduced. Finally, the corrected cleaning parameters θ=[P, R, C] are output as the execution layer instructions.

[0022] In a preferred embodiment, the physical world mapping unit comprises a UAV state synchronization module, an environment data acquisition module, a stain distribution mapping module, and a distributed Kalman filtering algorithm component.

[0023] Function: Real-time synchronization of UAV state (position, power, cleaning progress), environmental data (wind speed, obstacle coordinates), and stain distribution heat map to the virtual sand table; correct sensor errors through distributed Kalman filtering to ensure the mapping accuracy of the physical world and the virtual space (position error <0.5m).

[0024] In a preferred embodiment, the virtual simulation engine is internally provided with: a Unity3D three-dimensional scene construction tool, a multi-body physics engine (simulating water flow injection / brush head friction), and a parallel simulation calculation module.

[0025] Function: Build a three-dimensional virtual scene that is proportional to the physical world, support parallel motion simulation of 100+ UAV clusters; simulate the water flow impact force and the friction between the brush head and the stain during the cleaning process through the physics engine to predict the stain removal effect under different cleaning parameters (error <3%).

[0026] In a preferred embodiment, the multi-agent collaborative decision-making model realizes intelligent matching of sub-tasks and UAVs through an improved ant colony algorithm, and is internally provided with five components:

[0027] The task meshing and feature extraction component is responsible for converting physical space into distributable units, and the area division unit thereof is based on a three-dimensional model of a region to be cleaned, automatically divides a grid sub-task according to a radius of operation of a UAV, and marks each sub-task with a unique ID; the sub-task feature quantization unit collects total stain amount, average stain hardness, stain concentration, and center coordinates, generates a sub-task feature vector, the total stain amount is calculated by integrating the gray value of a sensor scan, the average hardness is converted by ultrasonic echo intensity, and the stain concentration is a ratio of the total stain amount to the area of the grid.

[0028] The multi-source state perception component collects real-time decision basis, the UAV state collection unit obtains residual power, maximum load, current position, and priority through built-in sensors, generates a UAV state vector, the residual power is converted by a battery voltage curve, the maximum load is determined by the UAV model, and the current position is based on GPS and IMU fusion positioning; the environmental interference perception unit collects wind speed and obstacle distribution, and is used to correct sub-task difficulty, for example, the actual hardness of a sub-task in a strong wind environment is the original hardness multiplied by (1+0.1x wind speed).

[0029] The pheromone dynamic management component builds a historical preference mechanism, the pheromone storage unit maintains a sub-task pheromone concentration matrix, and the initial value is set as a constant; the pheromone evaporation and enhancement unit proportionally decays all sub-task pheromones after each round of distribution, avoids repeated selection, and improves the pheromone concentration of high-quality sub-tasks according to actual effects after completing the task.

[0030] The heuristic decision component fuses multi-dimensional adaptability, the heuristic factor calculation unit generates a distance factor, a load matching factor, and a difficulty adaptation factor, the distance factor is the inverse of the distance from the UAV to the sub-task, the load matching factor is the minimum value of the maximum load of the UAV and the total stain amount of the sub-task divided by the maximum load, and the difficulty adaptation factor is the reciprocal of the product of (1+average hardnessx concentration); the probability decision unit calculates the probability of the UAV selecting a sub-task based on an ant colony algorithm selection probability formula, and the pheromone weight in the formula is set as 1.2, and the heuristic factor weight is set as 0.8.

[0031] The task allocation execution component completes matching and instruction generation, the greedy selection unit selects the sub-task with the highest probability for each UAV, if the sub-task is selected by multiple UAVs, the UAV with a higher priority is preferentially allocated; the allocation result verification unit checks whether there is an unallocated sub-task or a UAV with an empty load, and triggers secondary allocation; and the instruction generation unit issues execution instructions such as sub-task ID, target coordinates, and expected cleaning duration to the UAV.

[0032] In a preferred embodiment, the multi-agent collaborative decision-making model realizes real-time adjustment of path conflicts through virtual-real interaction of a digital twin sand table, and the core includes four components:

[0033] The path preview component preforms the complete path in the virtual space, in which the path planning engine generates a path point sequence based on the improved A* algorithm, inputting the UAV state and subtask coordinates; the time axis mapping unit calculates the arrival time of each path point according to the UAV flight speed, and constructs the space-time path.

[0034] The conflict detection component identifies the intersection risk, in which the space-time distance calculation unit traverses the space-time paths of any two UAVs to calculate the spatial distance at the same time; the conflict determination unit determines that there is a path conflict if the distance is less than 2 meters and the duration exceeds 0.5 seconds, and records the conflict time, location and involved UAV ID.

[0035] The dynamic adjustment component corrects the conflict path, in which the priority sorting unit sorts the conflict UAVs according to the remaining power and task urgency, and preferentially adjusts the UAVs with low priority; the path offset unit offsets the path by 1-2 meters in the vertical direction or reallocates the subtasks through the path offset algorithm in the virtual sand table; the speed adjustment unit adjusts the flight speed when path offset is not feasible to stagger the conflict time.

[0036] The response optimization component ensures real-time performance, in which the parallel computing unit uses GPU to accelerate path preview and conflict detection, supporting parallel path calculation for 100 UAVs; the lightweight model unit simplifies the environment model, ignores the influence of low wind speed, and combines dense obstacles into regional blocks to reduce the calculation complexity.

[0037] The heuristic factor fusion formula realizes accurate matching between subtasks and UAVs by nonlinearly fusing the distance load difficulty and concentration, and the expression is as follows:

[0038] F i,m = D i,m M i,m A i,m exp(-kC m );

[0039] In the formula:

[0040] F i,m is the comprehensive attraction of subtask m to UAV i, and the higher the dimensionless value, the stronger the adaptability;

[0041] D i,m is the distance factor, which is dimensionless, and the value is the reciprocal of the distance between the current coordinates of the UAV and the center coordinates of the subtask; the closer the distance, the greater the value;

[0042] M i,m is the load matching factor, which is dimensionless, and the value is the minimum value of the maximum load of the UAV divided by the total amount of the subtask stain, and the higher the matching degree, the closer the value to 1;

[0043] A i,mThe difficulty adaptation factor is dimensionless, and its value is 1 divided by the product of the average hardness and the concentration. The lower the difficulty, the closer the value is to 1;

[0044] K is a concentration penalty coefficient, which is dimensionless and takes a value of 0.5. The attraction of high-concentration sub-tasks is inhibited by an exponential term to avoid overloading a single unmanned aerial vehicle.

[0045] C m The sub-task stain concentration is dimensionless, and its value is the ratio of the total amount of stains to the grid area, indicating the concentration of stain distribution.

[0046] In a preferred embodiment, the internal structure of the distributed task allocation and path planning module comprises a task decomposition unit, a path optimization algorithm, and a dynamic reallocation trigger.

[0047] The task decomposition unit is composed of an energy consumption evaluation matrix (flight distance, load weight weight), and its function is to divide the grid sub-tasks according to the operation radius of the unmanned aerial vehicle (5-8 m), and to preferentially allocate the dense stain area (concentration > 0.5 kg / m2) to the unmanned aerial vehicle with more than 60% power; the path optimization algorithm is composed of an improved A* algorithm (fusing energy consumption weight factor: flight energy consumption 40%, cleaning energy consumption 60%), and its function is to generate a low-energy path, and the planning time is less than 100 ms per vehicle; the dynamic reallocation trigger is composed of a low power threshold (15%) and a fault detection mechanism (current fluctuation > 20%), and its function is to transfer the remaining tasks to the unmanned aerial vehicle with less than 30 m distance and more than 50% power when the unmanned aerial vehicle is insufficient or fails.

[0048] In a preferred embodiment, the internal structure of the multi-vehicle communication and cooperative control module comprises a Mesh self-organizing network unit, a distributed control protocol, and an anti-interference mechanism.

[0049] The Mesh self-organizing network unit is composed of 5G+LoRa dual-mode communication technology, and its function is to support 100 nodes concurrent connection within 10 km (5G transmission rate > 10 Mbps, LoRa backup communication distance > 5 km); the distributed control protocol is composed of the Time-Sensitive Network (TSN) standard, and its function is to realize real-time synchronization of task instructions and state data, with transmission delay < 1 ms and time synchronization accuracy ± 0.1 μs; the anti-interference mechanism is composed of adaptive frequency hopping technology (2.4 GHz ISM band, 50 frequency hopping channels) and AES-256 encryption, and its function is to avoid electromagnetic interference and protect critical instructions (such as task allocation, emergency return) from tampering.

[0050] In a preferred embodiment, the internal structure of the cleaning execution and feedback module comprises an execution mechanism controller, an effect detection unit, and a fault diagnosis sub-module.

[0051] The composition of the actuator controller is a PID closed-loop control component, and the function is to adjust the water pump water pressure (0.5-2 MPa), the brush head motor speed (1500-3000 rpm) and the cleaning agent pump flow (0.1-1 L / min), and the control accuracy is ±0.1 MPa (water pressure), 50 rpm (speed); the composition of the effect detection unit is an RGB-D camera and a gray value analysis algorithm, and the function is to shoot images before and after cleaning, and calculate the stain residual rate through the gray value mean square difference (formula: residual rate = gray value after cleaning / gray value before cleaning x 100%); the composition of the fault diagnosis submodule is a current sensor and a fluctuation monitoring algorithm, and the function is to monitor the motor current (normal range 5-15 A) and water pressure fluctuation (threshold ±0.2 MPa) in real time, and trigger an audible and light alarm and upload the fault code when abnormal.

[0052] In summary, due to the adoption of the technical scheme, the present application has the following advantages:

[0053] 1. In the present application, the digital twin cooperative cleaning sand table module realizes real-time linkage between virtual simulation and actual operation, significantly improving the cooperative efficiency of multiple unmanned aerial vehicle clusters. The digital twin sand table can dynamically map the unmanned aerial vehicle state, environmental data and stain distribution, pre-visualize task allocation and path planning in the virtual space, and discover and resolve path conflicts between unmanned aerial vehicles in advance, avoiding the problem of repeated cleaning or missed cleaning caused by information lag in traditional operations. At the same time, the sand table adjusts subtask allocation based on real-time state, preferentially allocates dense stain areas to unmanned aerial vehicles with sufficient power, reduces empty flight time and invalid energy consumption, and ensures that the overall operation process is more coherent and the cleaning coverage is more complete.

[0054] 2. In the present application, the intelligentization and adaptive adjustment of the cleaning process are realized through multi-module deep cooperation, improving the cleaning effect and resource utilization rate in complex scenarios. Multi-source heterogeneous sensors accurately collect stain characteristics, generate customized cleaning parameters based on the stain characteristic library and AI decision model, and ensure that different types of stains can match the optimal water pressure, speed and cleaning agent concentration. The digital twin sand table further optimizes these parameters, predicts the cleaning effect through virtual simulation and dynamically corrects the strategy, avoids resource waste caused by excessive cleaning or residual problems left by insufficient cleaning. In addition, the system supports self-learning of new stains, continuously updates the decision model through trial-and-error cleaning and effect feedback, adapts to diversified cleaning demands, and reduces manual intervention. BRIEF DESCRIPTION OF DRAWINGS

[0055] Fig. 1 The overall system block diagram of the present application is shown in the figure;

[0056] Fig. 2 The digital twin cooperative cleaning sand table module system block diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and not to limit the present application.

[0058] Embodiments:

[0059] Referring to Figs. 1-2 An unmanned aerial vehicle stain cleaning cooperative control system comprises a multi-source heterogeneous sensor perception module, a stain feature analysis and cleaning strategy generation module, a digital twin cooperative cleaning sand table module, a distributed task allocation and path planning module, a multi-machine communication and cooperative control module, and a cleaning execution and feedback module.

[0060] The digital twin cooperative cleaning sand table module is internally provided with a physical world mapping unit, a virtual simulation engine, and a multi-agent cooperative decision-making model.

[0061] The multi-source heterogeneous sensor perception module is connected to the stain feature analysis and cleaning strategy generation module and the digital twin cooperative cleaning sand table module.

[0062] The stain feature analysis and cleaning strategy generation module is connected to the input end of the cleaning execution and feedback module and the digital twin cooperative cleaning sand table module.

[0063] The digital twin cooperative cleaning sand table module is connected to the distributed task allocation and path planning module and receives the input end of the cleaning execution and feedback module data.

[0064] The distributed task allocation and path planning module is connected to the input end of the multi-machine communication and cooperative control module.

[0065] The multi-machine communication and cooperative control module is connected to the input end of the cleaning execution and feedback module and the digital twin cooperative cleaning sand table module.

[0066] The cleaning execution and feedback module is connected to the input end of the stain feature analysis and cleaning strategy generation module and the digital twin cooperative cleaning sand table module.

[0067] The multi-source heterogeneous sensor perception module is internally provided with a multi-modal sensor unit, a data preprocessing unit, and an edge computing node.

[0068] The multimodal sensor unit consists of an RGB-D camera, an infrared thermal imager, an ultrasonic sensor, and a water quality sensor. Its functions include acquiring visual features of stains, distinguishing temperature-sensitive stains, detecting stain thickness and adhesion, and monitoring residual ion concentration after cleaning. The data preprocessing unit consists of filtering and image segmentation algorithms. Its functions include removing noise from sensor data and extracting feature parameters such as stain contours and grayscale values. The edge computing node consists of an embedded chip. Its function is to process sensor data in real time on the drone, reducing cloud transmission latency and controlling data processing latency to within 50ms.

[0069] The stain feature analysis and cleaning strategy generation module uses a CNN-LSTM hybrid network to transform the input "stain feature vector" into customized cleaning parameters: water pressure, brush head speed, and cleaning agent concentration, and dynamically optimizes them by combining historical cleaning results.

[0070] The input layer consists of: The input is the preprocessed stain feature vector: X = [x1, x2, ..., x...]. n ], where x1~x n The physical characteristic parameters of the stain are: x1 = stain area (㎡), x2 = gray value, x3 = hardness (N), x4 = adhesion (Pa), x5 = ambient temperature (°C), etc., n = 8~12. To eliminate dimensional differences, each feature is mapped to the range [0,1] through standardization.

[0071] To incorporate historical cleaned data, the spatial feature vector F output by the CNN is... CNN Concatenate with the historical time series: H = [h1, h2, h3], h k The parameter-effect pair for the k-th historical cleaning, including water pressure, rotation speed, concentration, and corresponding residual rate, is concatenated and input into an LSTM network. The LSTM learns temporal dependencies through a gating mechanism and outputs a 32-dimensional vector F that integrates spatiotemporal features. LSTM .

[0072] F LSTM Input a 2-layer fully connected network, with the number of hidden layer neurons reduced from 64 to 32, and both layers using ReLU activation function. The final output is the initial cleaning parameters.

[0073] Water pressure P0 (unit: MPa), brush head rotation speed R0 (unit: rpm), and cleaning agent concentration C0 (unit: %) constitute the initial parameter vector θ0 = [P0, R0, C0].

[0074] The dynamic optimizer of the stain feature analysis and cleaning strategy generation module incorporates a reinforcement learning mechanism, using the historical cleaning residue rate r (r=0 indicates complete cleaning, r=1 indicates no cleaning) as a reward signal to adjust the initial parameters. Specifically, when the average residue rate of similar historical stains... At this time, adjust the parameters by policy gradient:

[0075] If higher (insufficient cleaning), increase water pressure or concentration; if lower (excessive cleaning waste resources), reduce parameters. The final output is the corrected cleaning parameter θ = [P, R, C], which is used as an execution layer instruction.

[0076] The physical world mapping unit includes a UAV state synchronization module, an environment data acquisition module, a stain distribution mapping module, and a distributed Kalman filter algorithm component.

[0077] Function: Real-time synchronization of UAV state (position, power, cleaning progress), environmental data (wind speed, obstacle coordinates), and stain distribution heat map to virtual sand table; correct sensor errors through distributed Kalman filtering to ensure the mapping accuracy of the physical world and virtual space (position error <0.5m);

[0078] The virtual simulation engine is internally provided with: Unity3D three-dimensional scene construction tool, multi-body physics engine (simulating water flow injection / brush head friction), and parallel simulation calculation module.

[0079] Function: Build a three-dimensional virtual scene proportional to the physical world, support parallel motion simulation of 100+ UAV clusters; simulate water flow impact force and brush head friction with stains during the cleaning process through the physics engine to predict stain removal effects under different cleaning parameters (error <3%).

[0080] The multi-agent collaborative decision-making model realizes intelligent matching of sub-tasks and UAVs through an improved ant colony algorithm, and is internally provided with five components:

[0081] The task gridding and feature extraction component is responsible for converting the physical space into assignable units, and its area division unit automatically divides the grid sub-tasks based on the three-dimensional model of the area to be cleaned according to the UAV operation radius, and each sub-task is marked with a unique ID; the sub-task feature quantization unit collects the total amount of stains, average stain hardness, stain concentration, and center coordinates, generates a sub-task feature vector, the total amount of stains is calculated by integrating the gray value scanned by the sensor, the average hardness is converted by ultrasonic echo intensity, and the stain concentration is the ratio of the total amount of stains to the grid area.

[0082] The multi-source state perception component collects decision basis in real time. The unmanned aerial vehicle state acquisition unit obtains the residual power, maximum load, current position and priority through the built-in sensor, generates an unmanned aerial vehicle state vector, the residual power is converted through the battery voltage curve, the maximum load is determined by the unmanned aerial vehicle model, and the current position is based on the GPS and IMU fusion positioning; the environmental interference perception unit collects the wind speed and obstacle distribution, which is used to correct the sub-task difficulty, for example, the actual hardness of the sub-task in a strong wind environment is the original hardness multiplied by (1+0.1×wind speed).

[0083] The pheromone dynamic management component constructs a historical preference mechanism. The pheromone storage unit maintains a sub-task pheromone concentration matrix, and the initial value is set as a constant; the pheromone volatilization and enhancement unit attenuates all sub-task pheromones by a certain proportion after each round of allocation to avoid repeated selection, and improves the pheromone concentration of high-quality sub-tasks according to the actual effect after completing the task.

[0084] The heuristic decision component fuses multi-dimensional adaptability. The heuristic factor calculation unit generates distance factor, load matching factor and difficulty adaptation factor. The distance factor is the inverse of the distance from the unmanned aerial vehicle to the sub-task, the load matching factor is the minimum value of the maximum load of the unmanned aerial vehicle and the total amount of stains of the sub-task divided by the maximum load, and the difficulty adaptation factor is the product of 1 divided by (1+average hardness×concentration). The probability decision unit calculates the probability of the unmanned aerial vehicle selecting the sub-task based on the ant colony algorithm selection probability formula, and the pheromone weight in the formula is set to 1.2 and the heuristic factor weight is set to 0.8.

[0085] The task allocation execution component completes matching and instruction generation. The greedy selection unit selects the sub-task with the highest probability for each unmanned aerial vehicle, and if the sub-task is selected by multiple unmanned aerial vehicles, it is preferentially allocated to the unmanned aerial vehicle with higher priority. The allocation result verification unit checks whether there is an unallocated sub-task or an unmanned aerial vehicle with empty load, and triggers secondary allocation. The instruction generation unit issues execution instructions such as sub-task ID, target coordinates and expected cleaning duration to the unmanned aerial vehicle.

[0086] The multi-agent collaborative decision-making model realizes real-time adjustment of path conflicts through virtual-real interaction of the digital twin sand table, and the core includes four components.

[0087] The path preview component previews the complete path in the virtual space. The path planning engine in it is based on the improved A* algorithm, inputs the unmanned aerial vehicle state and sub-task coordinates, and generates a path point sequence; the time axis mapping unit calculates the arrival time of each path point according to the flight speed of the unmanned aerial vehicle, and constructs the space-time path.

[0088] The conflict detection component identifies intersection risks. The space-time distance calculation unit in it traverses the space-time paths of any two unmanned aerial vehicles to calculate the spatial distance at the same time; the conflict judgment unit determines that there is a path conflict if the distance is less than 2 meters and the duration is more than 0.5 seconds, and records the conflict time, position and involved unmanned aerial vehicle ID.

[0089] The dynamic adjustment component corrects conflict paths. The priority sorting unit sorts conflicting drones based on remaining battery power and task urgency, prioritizing the adjustment of drones with lower priority. The path offset unit uses a path offset algorithm in the virtual sandbox to offset the drones vertically by 1 to 2 meters or to reassign sub-tasks. The speed adjustment unit adjusts the flight speed when path offset is not feasible to avoid conflict moments.

[0090] The response optimization component ensures real-time performance. Its parallel computing unit uses GPU-accelerated path pre-simulation and conflict detection, supporting parallel path calculation for 100 drones. The lightweight model unit simplifies the environment model, ignores the impact of low wind speed, and merges dense obstacles into region blocks to reduce computational complexity.

[0091] The heuristic factor fusion formula achieves precise matching between subtasks and UAVs by nonlinearly fusing distance load difficulty and concentration. The expression is as follows:

[0092] F i,m =D i,m M i,m A i,m exp(-kC m );

[0093] In the formula:

[0094] F i,m The higher the dimensionless value of the overall attractiveness of subtask m to UAV i, the stronger the adaptability.

[0095] D i,m The distance factor is dimensionless and its value is the reciprocal of the distance between the current coordinates of the UAV and the coordinates of the sub-task center. The closer the distance, the larger the value.

[0096] M i,m The load matching factor is dimensionless and its value is the minimum value of the maximum load of the drone and the total amount of dirt in the sub-task divided by the maximum load. The higher the matching degree, the closer the value is to 1.

[0097] A i,m The difficulty adaptation factor is dimensionless and its value is 1 divided by 1 plus the product of average hardness and concentration. The lower the difficulty, the closer the value is to 1.

[0098] k is a dimensionless concentration penalty coefficient with a value of 0.5. The exponential term suppresses the attraction of high-concentration sub-tasks to avoid overloading a single drone.

[0099] C m The stain concentration of the subtask is dimensionless and its value is the ratio of the total stain amount to the grid area, representing the density of stain distribution.

[0100] The internal settings of the distributed task allocation and path planning module include a task decomposition unit, a path optimization algorithm, and a dynamic reallocation trigger.

[0101] The task decomposition unit is composed of an energy consumption evaluation matrix (flight distance, load weight weight), and its function is to divide the grid task according to the unmanned aerial vehicle operation radius (5-8 m), and preferentially allocate the dense stain area (concentration > 0.5 kg / ㎡) to the unmanned aerial vehicle with more than 60% power; the path optimization algorithm is composed of an improved A* algorithm (fusing energy consumption weight factor: flight energy consumption 40%, cleaning energy consumption 60%), and its function is to generate a low energy consumption path, and the planning time is less than 100 ms per frame; the dynamic reallocation trigger is composed of a low power threshold (15%) and a fault detection mechanism (current fluctuation > 20%), and its function is to transfer the remaining task to the unmanned aerial vehicle with a distance less than 30 m and a power greater than 50% when the unmanned aerial vehicle is insufficient or fails.

[0102] The internal settings of the multi-machine communication and cooperative control module include a Mesh self-organizing network unit, a distributed control protocol, and an anti-interference mechanism.

[0103] The Mesh self-organizing network unit is composed of 5G+LoRa dual-mode communication technology, and its function is to support 100 nodes concurrent connection within 10 km (5G transmission rate > 10 Mbps, LoRa backup communication distance > 5 km); the distributed control protocol is composed of the Time Sensitive Network (TSN) standard, and its function is to realize real-time synchronization of task instructions and state data, transmission delay < 1 ms, and time synchronization accuracy ± 0.1 μs; the anti-interference mechanism is composed of adaptive frequency hopping technology (2.4 GHz ISM frequency band, 50 frequency hopping channels) and AES-256 encryption, and its function is to avoid electromagnetic interference and protect key instructions (such as task allocation, emergency return) from tampering.

[0104] The internal settings of the cleaning execution and feedback module include an actuator controller, an effect detection unit, and a fault diagnosis sub-module.

[0105] The actuator controller is composed of a PID closed-loop control component, and its function is to adjust the water pump water pressure (0.5-2 MPa), the brush head motor speed (1500-3000 rpm), and the cleaning agent pump flow (0.1-1 L / min), and the control accuracy is ± 0.1 MPa (water pressure) and 50 rpm (speed); the effect detection unit is composed of an RGB-D camera and a gray value analysis algorithm, and its function is to shoot images before and after cleaning, and calculate the stain residual rate by gray value mean square deviation (formula: residual rate = post-cleaning gray value / pre-cleaning gray value x 100%); the fault diagnosis sub-module is composed of a current sensor and a fluctuation monitoring algorithm, and its function is to monitor the motor current (normal range 5-15 A) and water pressure fluctuation (threshold ± 0.2 MPa) in real time, and trigger audible and light alarms and upload fault codes when abnormal.

[0106] From the above, it can be seen that:

[0107] In the present application, the real-time linkage of virtual simulation and actual operation is realized through the digital twin cooperative cleaning sand table module, significantly improving the cooperative efficiency of multi-unmanned aerial vehicle cluster. The digital twin sand table can dynamically map the unmanned aerial vehicle state, environmental data and stain distribution, pre-act task allocation and path planning in the virtual space, and discover and eliminate path conflicts between unmanned aerial vehicles in advance, avoiding the problem of repeated cleaning or missed cleaning caused by information lag in traditional operation. At the same time, the sand table adjusts the sub-task allocation based on real-time state, preferentially allocates dense stain areas to unmanned aerial vehicles with sufficient power, reduces the time of empty flying and invalid energy consumption, and ensures that the overall operation process is more coherent and the cleaning coverage is more complete.

[0108] In the present application, the intelligentization and adaptive adjustment of the cleaning process are realized through multi-module deep cooperation, improving the cleaning effect and resource utilization rate in complex scenarios. Multi-source heterogeneous sensors accurately collect stain characteristics, generate customized cleaning parameters combined with the stain characteristic library and AI decision model, and ensure that different types of stains can match the optimal water pressure, speed and cleaning agent concentration. The digital twin sand table further optimizes these parameters, predicts the cleaning effect through virtual simulation and dynamically corrects the strategy to avoid resource waste caused by excessive cleaning or residual problems left by insufficient cleaning. In addition, the system supports self-learning of new stains, continuously updates the decision model through trial-and-error cleaning and effect feedback, adapts to diversified cleaning needs, and reduces manual intervention.

[0109] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or equipment including the element.

[0110] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A UAV stain cleaning cooperative control system, characterized in that: The application relates to a multi-source heterogeneous sensor perception module, a stain feature analysis and cleaning strategy generation module, a digital twin cooperative cleaning sand table module, a distributed task allocation and path planning module, a multi-machine communication and cooperative control module and a cleaning execution and feedback module. The digital twin cooperative cleaning sand table module is internally provided with a physical world mapping unit, a virtual simulation engine and a multi-agent cooperative decision-making model. The multi-source heterogeneous sensor perception module is connected with the stain feature analysis and cleaning strategy generation module and the digital twin cooperative cleaning sand table module. The stain feature analysis and cleaning strategy generation module is connected with the input end of the cleaning execution and feedback module and the digital twin cooperative cleaning sand table module. The digital twin cooperative cleaning sand table module is connected with the distributed task allocation and path planning module and receives the input end of the cleaning execution and feedback module data. The distributed task allocation and path planning module is connected with the input end of the multi-machine communication and cooperative control module. The multi-machine communication and cooperative control module is connected with the input end of the cleaning execution and feedback module and the digital twin cooperative cleaning sand table module. The cleaning execution and feedback module is connected with the input end of the stain feature analysis and cleaning strategy generation module and the digital twin cooperative cleaning sand table module. The multi-source heterogeneous sensor perception module is internally provided with a multi-modal sensor unit, a data preprocessing unit and an edge computing node. 2.The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: The multi-modal sensor unit comprises an RGB-D camera, an infrared thermal imager, an ultrasonic sensor and a water quality sensor. The dynamic optimizer of the stain feature analysis and cleaning strategy generation module corrects initial parameters by taking historical cleaning residual rate r as a reward signal through a reinforcement learning mechanism. 3.The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: The input layer of the stain feature analysis and cleaning strategy generation module includes: the input is the preprocessed stain feature vector: X = [x1, x2, ..., x n ], where x1~x n The physical characteristic parameters of the stain are: x1 = stain area, x2 = gray value, x3 = hardness, x4 = adhesion, x5 = ambient temperature, n = 8~12; The physical world mapping unit comprises a UAV state synchronization module, an environment data acquisition module, a stain distribution mapping module and a distributed Kalman filtering algorithm component. 4.The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: The virtual simulation engine is internally provided with a Unity3D three-dimensional scene construction tool, a multi-body physics engine and a parallel simulation calculation module. 5.The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: The multi-agent cooperative decision-making model is internally provided with five components: a task gridding and feature extraction component, a multi-source state perception component, a pheromone dynamic management component, a heuristic decision-making component and a task allocation and execution component.

6. The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: The task gridding and feature extraction component converts a physical space into assignable units. The multi-source state perception component collects decision-making basis in real time. The pheromone dynamic management component constructs a historical preference mechanism. The heuristic decision-making component fuses multi-dimensional adaptability. The task allocation and execution component completes matching and instruction generation. The multi-agent cooperative decision-making model realizes real-time adjustment of path conflicts through virtual-real linkage of the digital twin sand table, and the core comprises four components: 7.The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: The path pre-performance component pre-performs a complete path in a virtual space, wherein a path planning engine in the path pre-performance component is based on an improved A* algorithm, inputs UAV states and subtask coordinates and generates a path point sequence; The conflict detection component identifies intersection risks, wherein a space-time distance calculation unit in the conflict detection component traverses space-time paths of any two UAVs and calculates spatial distances at the same time. ​ The dynamic adjustment component corrects the conflict path, and a priority sorting unit in the dynamic adjustment component sorts the unmanned aerial vehicles according to the residual power and the task urgency, and preferentially adjusts the unmanned aerial vehicle with low priority; The response optimization component ensures real-time performance, and a parallel computing unit in the response optimization component adopts GPU to accelerate path rehearsal and conflict detection, and supports parallel path calculation of 100 unmanned aerial vehicles; The heuristic factor fusion formula of the multi-agent collaborative decision-making model realizes accurate matching between subtasks and unmanned aerial vehicles by nonlinearly fusing distance load difficulty and concentration, and has the following expression: F i,m = D i,m M i,m A i,m exp(-kC m ); In the formula, k is a concentration penalty coefficient, and a dimensionless value of 0.5 is taken, an exponential term is used to inhibit the attraction of a high-concentration subtask, and overloading of a single unmanned aerial vehicle is avoided; and F i,m The higher the dimensionless value of the comprehensive attractiveness of the subtask m to the UAV i is, the stronger the adaptability is. D i,m The distance factor is dimensionless, and its value is the reciprocal of the distance between the current coordinates of the UAV and the coordinates of the sub-task center. The closer the distance, the greater the value. M i,m The dimensionless value of the load matching factor is the minimum value of the total amount of stains of the sub-tasks divided by the maximum load matching degree of the maximum load of the UAV. The higher the maximum load matching degree, the closer the value is to 1. A i,m Difficulty adaptation factor is dimensionless, its value is 1 divided by 1 plus the product of average hardness and concentration. The lower the difficulty, the closer the value is to 1; The distributed task allocation and path planning module is internally provided with a task decomposition unit, a path optimization algorithm and a dynamic reallocation trigger. C m The non-dimensional value of the subtask stain concentration is the ratio of the total amount of stains to the area of the grid, which represents the degree of stain distribution concentration. 8.The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: The multi-machine communication and collaborative control module is internally provided with a Mesh self-organizing network unit, a distributed control protocol and an anti-interference mechanism module. 9.The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: The cleaning execution and feedback module is internally provided with an execution mechanism controller, an effect detection unit and a fault diagnosis sub-module. 10.The unmanned aerial vehicle stain cleaning cooperative control system of claim 1, wherein: ​

Images