Dual-algorithm-driven multi-mode high-altitude operation system and intelligent control method

Abstract

The invention relates to the technical field of intelligent control of high-altitude operation robots, and discloses a dual-algorithm-driven multi-mode high-altitude operation system and an intelligent control method, and the method comprises the steps: obtaining environment sensing data and robot state parameters, and generating an operation task parameter set, a genetic algorithm optimization module is used for generating a global job task scheduling scheme through chromosome coding and fitness function evaluation, dynamically adjusting a job control strategy based on a reinforcement learning algorithm according to a state space, an action space and a reward function to generate a real-time control instruction, and realizing closed-loop cooperation of the genetic algorithm and the reinforcement learning algorithm; according to the invention, the intelligent control and decision-making problem of the aerial work robot is solved, and autonomous decision-making, path planning and multi-task efficient cooperation are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of high-altitude operation robots, and more particularly to a dual-algorithm-driven multi-modal high-altitude operation system and an intelligent control method. BACKGROUND

[0002] With the rapid development of urban construction, glass curtain wall buildings and high-rise buildings are increasing, and the demand for high-altitude glass cleaning, maintenance, and detection operations is growing. Traditional high-altitude operations rely on manual labor, and workers need to use tools such as ropes and scaffolding to work in high-altitude environments, which poses a high safety risk, with frequent falling accidents, high labor costs, and low operation efficiency, making it difficult to meet the demand for large-scale and normalized high-altitude operations.

[0003] To address the above problems, high-altitude operation robot systems have been gradually applied to high-altitude cleaning, detection, and maintenance scenarios. Existing high-altitude operation robot systems typically use preset path planning or manual remote control to perform operation tasks. In terms of task scheduling, traditional methods rely on manual experience to develop scheduling rules or use simple heuristic algorithms for path planning; in terms of operation control, they mainly rely on manual real-time remote control or execution of pre-programmed fixed action sequences.

[0004] However, the existing technology has the following defects: first, the traditional scheduling method relies on manual experience or simple rules, which easily falls into a local optimal solution and cannot fully consider the global optimization of multi-dimensional parameters such as operation area characteristics, task type, and robot state, resulting in low operation efficiency. Second, static scheduling methods develop scheduling schemes before the start of operation tasks, making it difficult to respond to changes in the high-altitude dynamic environment, such as sudden obstacles and weather changes, resulting in a lag in response and a decrease in operation safety and reliability. Third, single algorithms lack coordination capabilities and cannot simultaneously consider global optimization and dynamic adaptation, making it difficult to achieve task parallelism, process connection, and load balancing for multi-robot clusters, resulting in difficulty in multi-task coordination. SUMMARY

[0005] The present application provides a dual-algorithm-driven multi-modal high-altitude operation system and an intelligent control method, which solves the technical problems of insufficient intelligent control and decision-making capabilities of high-altitude operation robots in related technologies, and cannot achieve autonomous decision-making and efficient multi-task coordination.

[0006] The present application provides a dual-algorithm-driven multi-modal high-altitude operation system and an intelligent control method, comprising the following steps: Obtain high-altitude operation environment perception data and robot state parameters, generate operation task parameter sets, the environment perception data includes building surface three-dimensional model data, obstacle position data, wind speed data, the robot state parameters include robot position coordinates, battery remaining capacity, mechanical arm load state; input the set of job task parameters into a genetic algorithm optimization module, encode the job task scheduling scheme as a chromosome, evaluate the scheduling scheme using a fitness function, and perform selection, crossover, and mutation operations to generate a global job task scheduling scheme; Based on the global job task scheduling scheme and real-time environmental perception data, the optimal job control strategy is obtained by using a reinforcement learning algorithm, and real-time control instructions are generated.

[0007] Further, the processing flow of the genetic algorithm optimization module includes: Based on the set of job task parameters, an initial population is generated, and the job task scheduling scheme is encoded as a chromosome, each chromosome representing a scheduling scheme, the encoding structure of the chromosome including task execution order, assigned robot identifier, and task start time; Calculate the fitness function value of each chromosome in the initial population, the fitness function comprehensively evaluates the total job time, total energy consumption and job quality penalty term of the scheduling scheme; Based on the fitness function value, the next generation population is generated by performing selection, crossover and mutation operations, the selection operation calculates the selection probability according to the reciprocal of the chromosome fitness function value, the crossover operation exchanges gene fragments at the chromosome encoding position, and the mutation operation randomly disturbs the gene bits in the chromosome encoding; Determine whether the termination condition is met, if yes, output the optimal chromosome as the global job task scheduling scheme, otherwise continue iteration.

[0008] Further, the calculation method of the fitness function is: The total job time, total energy consumption and job quality penalty term of the scheduling scheme are weighted and summed, wherein the total job time is determined by calculating the maximum time for each robot to complete all assigned tasks, the total energy consumption is determined by accumulating the execution energy consumption and moving energy consumption of each task and is normalized, and the job quality penalty term is assigned a penalty value according to whether the robot power constraint, mechanical arm load constraint, building slope constraint and safety constraint are violated; The termination condition is to reach the maximum number of iterations or the population optimal fitness function value does not improve continuously for a preset number of generations.

[0009] Further, the processing flow of the reinforcement learning algorithm includes: Define the state space, action space and reward function of reinforcement learning, the state space includes robot position, robot pose, environmental perception data and task execution progress, the action space includes robot movement action and job action, and the reward function comprehensively evaluates the task completion reward, energy consumption penalty and safety reward; The deep Q network algorithm is used to train the reinforcement learning model, a deep neural network is used to approximate the action value function, an input layer of the deep neural network receives a state vector, and an output layer outputs Q values of each action, a job task is performed in a simulation environment, and state transition data is stored in an experience replay buffer, a batch of data is randomly sampled from the experience replay buffer, and a time difference error is used to update parameters of the deep neural network, and an optimal job control strategy is obtained; Based on real-time environment perception data and the optimal job control strategy, input the current state into the trained reinforcement learning model to select the optimal action, and convert the optimal action into a real-time control instruction and send it to the robot.

[0010] Further, the calculation method of the time difference error is: The time difference error is obtained by using the current obtained reward plus the product of the discount factor and the maximum Q value of the next state, and subtracting the Q value of the current action performed in the current state; Based on the time difference error, a loss function is constructed, the loss function is defined as the mean square error expectation value of the time difference error, and an optimization algorithm is used to update the parameters of the deep neural network to make the loss function converge.

[0011] Further, it further comprises: In the step of generating a real-time control instruction, when the wind speed or obstacle distance in the real-time environment perception data changes suddenly, a rapid response adjustment of the job control strategy is triggered, when the wind speed exceeds a safety threshold, a safety braking action is forced to be executed, and when the obstacle distance is less than a safety threshold, an obstacle avoidance action is forced to be executed and a local path is re-planned.

[0012] Further, it further comprises: The strategy experience learned by the reinforcement learning algorithm is fed back to the genetic algorithm optimization module, experience data is accumulated during the execution of the job task by the robot, and the optimization law of the job task scheduling is extracted, the optimization law is converted into a correction parameter or a constraint condition of the fitness function of the genetic algorithm, and a global job task scheduling scheme is generated by using the corrected fitness function in the next round of job task scheduling, realizing the closed-loop cooperation of the genetic algorithm and the reinforcement learning algorithm.

[0013] Further, it further comprises: When the robot has a partial function failure, a fault diagnosis report of the failed robot is obtained, the remaining functional modules are analyzed based on the fault diagnosis report, and the number of remaining available degrees of freedom, the volume of the remaining working space, and the remaining perception ability index are calculated, and a remaining function evaluation result is generated; Based on the remaining function evaluation result and the function-task mapping relationship, the task queue of the faulty robot is classified, tasks meeting the remaining function executable condition are marked as degradable execution tasks, and tasks not meeting the condition are marked as transfer tasks, a degradable execution task set and a transfer task set are generated; The task in the transfer task set is calculated with the task transfer cost between each normal robot, and the robot with the minimum cost is selected as the receiving robot, the task in the degradable execution task set is generated according to the degradation rule, and the expected quality score is calculated, when the expected quality score is greater than the minimum quality threshold, the degradation scheme is accepted, and the task re-allocation scheme under the fault condition is integrated.

[0014] Further, the calculation method of the task transfer cost is: The migration cost is determined by the spatial distance between the task job location and the current location of the normal robot and is normalized, and the load influence cost is determined by the sum of the current task load of the normal robot and the expected time consumption of the to-be-transferred task divided by the maximum task load capacity of the robot, the migration cost and the load influence cost are weighted and summed to obtain the task transfer cost; The calculation method of the expected quality score is: the ratio of the degraded job area to the original job area, the ratio of the original job accuracy requirement to the degraded job accuracy requirement are weighted and summed to obtain the expected quality score.

[0015] The application provides a dual-algorithm-driven multi-modal high-altitude operation system and an intelligent control system, which are used for executing the intelligent control method. The mobile and positioning module is used for realizing the movement and positioning of the robot on the building surface. The operation execution module includes a multi-degree-of-freedom mechanical arm and a replaceable end effector, and is used for executing the high-altitude operation task. The environment perception and safety module includes a wind speed sensor, an ultrasonic sensor, a vision sensor and a collision detection sensor, and is used for realizing real-time environment perception and safety protection. The data transmission and control module includes a wireless communication module and an industrial computer, and is used for data transmission and robot control. The energy supply module is used for providing energy support for the robot.

[0016] The beneficial effects of the present application are that the present application solves the intelligent control and decision-making problem of high-altitude operation robots by using the dual-algorithm driving strategy of the fusion of genetic algorithm and reinforcement learning, and achieves the following technical effects: first, the genetic algorithm is used to globally optimize the operation task parameters, the optimal solution is selected from a large number of scheduling schemes through the evolution process of coding-selection-crossover-mutation, the global optimization of multi-dimensional parameters is realized, the defect that the traditional scheduling method falls into local optimization is avoided, and the operation efficiency is improved; second, the reinforcement learning algorithm is used to perceive the environmental changes in real time and adjust the operation control strategy through the interaction process of state-action-reward, the defect that the static scheduling method is difficult to cope with dynamic environmental changes is overcome, and the safety and reliability of the operation are improved; third, through the closed-loop cooperation of the genetic algorithm and the reinforcement learning algorithm, an intelligent scheduling closed loop combining global planning and dynamic adaptation is formed, the defect that a single algorithm lacks cooperation ability is overcome, the task parallelism, process connection and load balancing of the multi-robot cluster are realized, and the multi-task cooperation performance is improved; fourth, through the fault robot task degradation transfer method based on the remaining function evaluation, the remaining operation capacity of the fault robot is quantitatively analyzed and the task classification degradation transfer is realized, the defect that the traditional method regards the robot as a binary state and causes the remaining function to be idle and wasted is overcome, and the resource utilization rate and operation continuity of the cluster under the fault condition are improved. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of the dual-algorithm driven multi-modal high-altitude operation system and intelligent control method of the present application. DETAILED DESCRIPTION

[0018] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can include changes, modifications, or additions of elements to the functions and arrangements of the elements discussed without departing from the scope of the present disclosure. Various examples can omit, substitute, or add various procedures or components as appropriate, or modify aspects of the elements in a variety of ways. Additionally, features described with respect to some examples can be combined in other examples.

[0019] Example 1 The present embodiment provides a dual-algorithm driven multi-modal high-altitude operation system and intelligent control method, which comprises the following steps: Step 100: Obtain high-altitude operation environment perception data and robot state parameters, and generate a set of operation task parameters.

[0020] The high-altitude operation environment sensing data is collected by the environment sensing and safety module, including building surface three-dimensional model data, obstacle position data, and wind speed data; the robot state parameters are collected by the movement and positioning module and the operation execution module, including the current position coordinates of the robot, the remaining battery capacity, and the mechanical arm load state; and the operation task information to be executed is obtained, including the operation area range, the operation type (cleaning, detection, and maintenance), and the operation quality requirement. The above data is integrated into the operation task parameter set , wherein represents the operation area characteristic parameters, including the operation area area, the building surface material type, and the building surface slope; represents the robot state parameters, including the robot position coordinates , the battery remaining capacity percentage , and the current load force of the mechanical arm ; represents the task attribute parameters, including the task type identifier (1 for cleaning, 2 for detection, and 3 for maintenance), the task priority , and the task estimated time consumption .

[0021] It should be noted that the above building surface three-dimensional model data refers to a three-dimensional geometric model constructed by scanning the building surface through a laser radar and a vision sensor, using a point cloud data fusion algorithm. The model contains information such as the geometric shape, surface normal vector, and curvature of the building surface. Specifically, laser radar point cloud data and vision sensor image data are collected, the point cloud data is denoised and downsampled, multiple frames of point cloud data are registered to a unified coordinate system through an iterative closest point algorithm (ICP), a Poisson surface reconstruction algorithm is used to generate a three-dimensional grid model of the building surface, and the surface normal vector and curvature information of each grid element in the grid model are extracted to form the building surface three-dimensional model data.

[0022] The aforementioned iterative closest point algorithm takes multiple frames of point cloud data sets as input and outputs point cloud data in a unified coordinate system after registration; the algorithm calculates the closest point correspondence and coordinate transformation matrix between point clouds through iteration, minimizes the distance error between point clouds, and realizes spatial alignment of multiple frames of point cloud data.

[0023] The aforementioned Poisson surface reconstruction algorithm takes the registered point cloud data and its normal vector information as input and outputs a three-dimensional grid model of the building surface; the algorithm converts the surface reconstruction problem into a Poisson equation solving problem, calculates the best fitting of the gradient field of the indicator function and the point cloud normal vector field, and generates a smooth three-dimensional grid model.

[0024] It should be noted that the robot arm load state in the above robot state parameters refers to the current load force and torque state of each joint of the robot arm. Specifically, joint torque data is collected by torque sensors at each joint of the robot arm. ( , (Number of joints in the robotic arm), and calculate the load force borne by the end effector based on the kinematic model of the robotic arm. The kinematic model of the robotic arm is represented as follows: in Let Jacobian matrix be the value of the robotic arm. For the angle vectors of each joint, For the torque vector of each joint, This is the transpose of the Jacobian matrix.

[0025] Step 200: Input the job task parameter set into the genetic algorithm optimization module to generate a global job task scheduling scheme.

[0026] Using genetic algorithms to analyze the set of job parameters Perform global optimization to generate a global job scheduling scheme. Global job scheduling scheme This includes task sequence, task path planning, and task execution time window.

[0027] The processing flow of the genetic algorithm optimization module includes: Step 201: Based on the job task parameter set This generates the initial population.

[0028] The job scheduling scheme is encoded as chromosomes, with each chromosome representing a feasible scheduling scheme. The chromosome encoding structure is as follows: ,in The number of tasks to be executed. ( ) indicates the first The scheduling information for each task includes the task execution order, the assigned robot identifier, and the task start time. (Randomly generated) 1 chromosome, forming the initial population: in Population size.

[0029] Step 202: Calculate the fitness function value of each chromosome in the initial population.

[0030] fitness function The term used to evaluate the merits of a scheduling scheme is defined as: wherein, is the total operation time of the scheduling scheme , is the total energy consumption, is the operation quality penalty term; , , is a weight coefficient, satisfying . The smaller the fitness function value is, the better the scheduling scheme is.

[0031] The calculation method of is as follows: wherein, is the time for the i-th robot to complete all assigned tasks, which is calculated by accumulating the time for the robot to execute each task and the moving time between tasks.

[0032] The calculation method of is as follows: wherein, is the energy consumption for executing the i-th task, including the energy consumption for task execution and the energy consumption for movement. To eliminate the dimensional influence, is normalized, and the normalization method is wherein, is a preset maximum energy consumption reference value.

[0033] is the operation quality penalty term, which is assigned a penalty value when the scheduling scheme violates the constraint condition. The constraint conditions include: robot power constraint ( is the minimum power threshold, which is 20%), mechanical arm load constraint ( is the maximum load of the mechanical arm), building slope constraint (operation is prohibited or the moving speed is reduced when the slope exceeds 60°), and safety constraint (operation is stopped when the wind speed exceeds 15 m / s). If the scheduling scheme violates any constraint condition, is assigned a larger penalty value , otherwise .

[0034] Step 203: Based on the fitness function value, selection, crossover, and mutation operations are performed to generate the next generation population.

[0035] ​​​​​​The selection operation uses a roulette wheel selection algorithm, calculating the probability of selection based on the reciprocal of the chromosome's fitness function value; chromosomes with smaller fitness function values ​​have a higher probability of being selected. The crossover operation uses a single-point crossover algorithm, randomly selecting two parent chromosomes and exchanging gene segments at a specific location in their encoding to generate two offspring chromosomes. The mutation operation randomly perturbs a gene location in the chromosome's encoding, such as changing the task execution order or adjusting the task start time. The crossover probability is set. Probability of mutation .

[0036] Step 204: Determine if the termination condition is met. If it is, output the optimal chromosome as the global job scheduling scheme. Otherwise, return to step 202 and continue iterating.

[0037] The termination condition is: reaching the maximum number of iterations. or continuous The optimal fitness function value of the population did not improve. This implementation sets... , When the termination condition is met, the chromosome with the smallest fitness function value in the current population is selected as the optimal chromosome, and the global job scheduling scheme is obtained by decoding this chromosome. .

[0038] It should be noted that the above global task scheduling scheme It also includes multi-robot cluster collaborative scheduling information. When the aerial work system contains multiple robots, the genetic algorithm optimization module encodes the task sequences, spatial paths, and time windows of multiple robots into a chromosome, realizing multi-robot task parallelism, process connection, and load balancing. Specifically, the chromosome encoding structure is expanded to... ,in For the number of robots, Indicates assignment to the first A sequence of tasks for each robot. To be assigned to the The number of tasks for each robot. In the fitness function calculation, a spatial conflict penalty term between robots is added. When the work areas of multiple robots overlap within the same time window, a penalty value is assigned to avoid spatial conflicts.

[0039] Step 300: Based on the global job task scheduling scheme Based on real-time environmental perception data, reinforcement learning algorithms are used to dynamically adjust the operation control strategy and generate real-time control commands.

[0040] Global job scheduling scheme As an initial job control strategy, a reinforcement learning algorithm is used to dynamically adjust the job control strategy according to real-time environmental changes to generate real-time control instructions. The real-time control instructions include robot movement speed instructions, mechanical arm action instructions, and end effector control instructions.

[0041] The processing flow of the reinforcement learning algorithm includes: Step 301: Define the state space, action space, and reward function of reinforcement learning.

[0042] State space Including the current position of the robot, the pose of the robot, the environmental perception data (wind speed, obstacle distance), and the task execution progress. The state vector is represented as where is the robot position coordinates, is the robot pose angle (roll angle, pitch angle, yaw angle), is the current wind speed, is the distance to the nearest obstacle, is the task execution progress percentage.

[0043] Action space Including robot movement actions and job actions. Movement actions include forward, backward, left turn, right turn, and stop, identified by discrete actions ; job actions include angle adjustment instructions ( ) of each joint of the mechanical arm and end effector control instructions (such as cleaning brush rotation speed, detection probe activation).

[0044] Reward function is used to evaluate the immediate income of executing action in state , defined as: where, is the task completion reward, which is a positive reward when the robot completes a sub-task, and the reward value is related to the task quality; is the energy consumption penalty, represented as , where is the energy consumed by executing action , is the penalty coefficient; is the safety reward, which is a positive reward when the robot avoids collision or works within a safe wind speed range, and a negative penalty when safety braking is triggered or collision occurs.

[0045] Step 302: Train the reinforcement learning model to obtain the optimal job control strategy.

[0046] The reinforcement learning model is trained using the Deep Q-Network (DQN) algorithm. The DQN algorithm uses a deep neural network to approximate the action-value function. This function represents the state. Next action The expected value of long-term cumulative rewards.

[0047] The input layer of a deep neural network receives a state vector. The output layer outputs the Q-values ​​of each action. ( , (This refers to the action space size); the hidden layer uses a fully connected layer structure, and the activation function is the ReLU function.

[0048] The training process adopts an offline training mode, which includes: the robot performing tasks in a simulation environment, and observing the environmental state transition after each action is performed. ,in This is the current state. For the action to be performed, For the reward, The state transition is defined as follows: store the state transition data in the experience replay buffer; randomly sample batch data from the experience replay buffer and update the deep neural network parameters using the time difference (TD) error.

[0049] TD error is defined as ,in This is the discount factor, and its value range is... In this embodiment, The loss function is defined as the mean squared error of the TD error: in, For the current network parameters, For the target network parameters, This serves as an experience replay buffer. The Adam optimization algorithm is used to update the parameters of the deep neural network, and training is iteratively until the loss function converges to obtain the optimal job control strategy. .

[0050] Step 303: Generate real-time control commands based on real-time environmental perception data and optimal operation control strategies.

[0051] During the robot's task execution, environmental perception data is collected in real time to obtain its current status. ( (for time steps); the current state The input is fed into the trained reinforcement learning model, based on the optimal job control strategy. Select the optimal action ; the optimal action is converted into real-time control instructions, which are sent to the movement and positioning module and the work execution module of the robot through the data transmission and control module, to control the robot to perform corresponding movement actions and work actions.

[0052] It should be noted that the above reinforcement learning algorithm also includes a rapid response to dynamic environmental changes. When the wind speed or the distance to the obstacle in the real-time environmental perception data changes abruptly, a rapid response is triggered immediately to adjust the work control strategy. Specifically, a wind speed safety threshold and an obstacle distance safety threshold are set, and when , a safety braking action is forced to be performed, and the robot stops moving and remains in the adsorbed state; when , an obstacle avoidance action is forced to be performed, and the robot replans a local path to bypass the obstacle. The response time of the rapid response is not more than 1 second.

[0053] In the embodiments of the present application, in order to improve the accuracy of path planning and dynamic obstacle avoidance capability, in step 300, based on the environmental perception data of the laser radar and the vision sensor, a three-dimensional grid map of the building surface is constructed, and a dynamic window method (DWA) is used to realize dynamic obstacle avoidance in the movement process. Specifically, the three-dimensional model of the building surface is discretized into a three-dimensional grid map, and each grid cell is marked as a passable area or an obstacle area; during the movement of the robot, according to the current speed and acceleration constraints of the robot, the speed space that can be reached within a future time window is calculated, which is called a dynamic window; a plurality of groups of speed candidate values are sampled within the dynamic window, the motion trajectory of the robot within a future period of time is simulated for each group of speed candidate values, and the evaluation function value of the trajectory is calculated; the evaluation function includes three parts: target direction evaluation, obstacle distance evaluation, and speed evaluation, the target direction evaluation measures the deviation of the end of the trajectory from the target direction, the obstacle distance evaluation measures the distance between the trajectory and the nearest obstacle, and the speed evaluation prefers a higher moving speed; the speed candidate value with the highest evaluation function value is selected as the optimal speed instruction at the current time, which is sent to the movement and positioning module.

[0054] The aforementioned dynamic window method takes the current position of the robot, the current speed, the acceleration constraint, the target position, and the three-dimensional grid map of the building surface as inputs, and outputs the optimal speed instruction at the current time; the algorithm searches for the optimal speed within the speed space that satisfies the kinematic constraints of the robot, and comprehensively considers the target approach, obstacle avoidance, and motion smoothness through the evaluation function, to realize local path planning in a dynamic environment.

[0055] In the embodiments of the present application, in order to improve the safety of the robot operation process, in step 300, a safety evaluation model is established, multi-source data such as wind speed, position, attitude, and adsorption force are fused, the safety state of the robot is evaluated in real time, and a hierarchical safety response is automatically triggered when a safety risk is detected. Specifically, the safety state score of the safety evaluation model is defined as: is the wind speed safety score, which is calculated according to the ratio of the current wind speed to the safety threshold, is the adsorption force safety score, which is calculated according to the ratio of the actual adsorption force of the mobile mechanism to the minimum adsorption force requirement, is the attitude safety score, which is calculated according to the attitude angle deviation, is the attitude angle deviation, is the maximum allowed attitude angle deviation; is the load safety score, which is calculated according to the ratio of the current load force of the robot to the maximum load force, is the weight coefficient, which satisfies The value range of the safety state score is , and the higher the score, the better the safety.

[0056] Three safety response thresholds are set: when , the safety state is good, and the robot operates normally; when , a first-level safety response is triggered, the robot speed is reduced to 50% of the normal speed, and the sensing frequency is increased; when , a second-level safety response is triggered, the robot suspends operation and performs an emergency return to the safety area. Through hierarchical safety response, the safety of the robot operation process is ensured.

[0057] ​​​​​​​​​​​​​In the embodiments of the present application, in order to realize the closed-loop cooperation of the genetic algorithm and the reinforcement learning algorithm, in step 300, the excellent strategy experience learned by the reinforcement learning algorithm is fed back to the genetic algorithm optimization module to optimize the generation of the next round of global job task scheduling scheme. Specifically, in the process of executing the job task by the robot, the reinforcement learning algorithm accumulates experience data by interacting with the environment, including the optimal action selection, task execution time, energy consumption and the like in different job scenes; the accumulated experience data is analyzed to extract the optimization law of the job task scheduling, for example, the law that the spiral cleaning path is more energy-saving than the reciprocating path for the curved surface area of the glass curtain wall is identified; the extracted optimization law is converted into a correction parameter or a constraint condition of the fitness function of the genetic algorithm, for example, a path form preference item is added to the fitness function, and a higher fitness evaluation is given to the spiral path; in the next round of job task scheduling, the genetic algorithm optimization module generates the global job task scheduling scheme by using the corrected fitness function, so as to realize the continuous optimization and performance improvement of the algorithm. Through the closed-loop cooperation of the genetic algorithm and the reinforcement learning algorithm, the intelligent scheduling closed loop of ‘genetic algorithm global planning initial optimal scheduling → reinforcement learning dynamic adaptation to environmental changes → reinforcement learning experience feedback to genetic algorithm optimization direction’ is formed.

[0058] Embodiment 2 The present embodiment further solves the task reassignment problem in the process of multi-robot cooperative work under fault conditions, including the following steps: Step 400: Obtain the fault diagnosis report of the faulty robot, analyze the remaining functional modules based on the fault diagnosis report, and generate the remaining function evaluation result.

[0059] When the fault diagnosis module of the robot detects a functional module failure, a fault diagnosis report is generated and a task reassignment process is triggered. The fault diagnosis report includes fault module identification, fault type, fault occurrence time, and remaining functional module list. The fault module identification is represented in a hierarchical coding manner, for example, ‘ARM-J4’ represents the failure of the 4th joint of the mechanical arm, and ‘SENSOR-LIDAR’ represents the failure of the laser radar sensor. The fault type includes complete failure (the module cannot work) and performance degradation (the module can work but the performance index is reduced).

[0060] Based on the remaining functional module list in the fault diagnosis report, the remaining work capacity of the faulty robot is analyzed. For the failure of the mechanical arm, the number of remaining available degrees of freedom is calculated , and the remaining workspace volume ; for the sensor failure, the remaining perception ability index is calculated, which comprehensively considers parameters such as positioning accuracy and environmental perception range; for the failure of the mobile mechanism, the remaining movement ability index is calculated, including the maximum movement speed and the range of building slopes that can be passed.

[0061] Generate the remaining function evaluation result wherein is a list of remaining function modules. The remaining function evaluation result quantifies the work capability of the faulty robot that can still be executed under the condition of partial function failure.

[0062] It should be noted that the above-mentioned remaining work space volume refers to the volume of the space region that the end effector of the robot arm can reach under the constraint of the remaining available joints. Specifically, based on the kinematics model of the robot arm, the Monte Carlo sampling method is used to generate random angle combinations of the remaining joints, the position coordinates of the end effector are calculated for each angle combination, and the point cloud formed by all the position coordinates is subjected to convex hull calculation to obtain the boundary of the remaining work space, and the convex hull volume is calculated as For example, after the 4th joint of a 6-DOF robot arm fails, the remaining 3 DOFs, 10000 angle combinations are calculated by Monte Carlo sampling, and the corresponding end position point cloud is calculated. The convex hull is calculated using the Quick-Hull algorithm to obtain a remaining work space volume of about 35% of the normal work space volume.

[0063] The aforementioned Monte Carlo sampling method, the input is the angle range of the remaining joints and the sampling number, and the output is a set of random angle combinations of the remaining joints; this method generates a large number of joint angle combinations by randomly and uniformly sampling in the joint angle range, which is used to explore the reachable work space of the robot arm.

[0064] The aforementioned Quick-Hull algorithm, the input is the point cloud data in three-dimensional space, and the output is the convex hull of the point cloud and its volume; this algorithm constructs the outer surface of the convex hull through the recursive divide-and-conquer strategy, and quickly calculates the smallest convex polyhedron that encloses the point cloud.

[0065] It should be noted that the above-mentioned remaining perception capability index refers to the comprehensive perception capability score of the faulty robot under the condition of partial sensor failure. Specifically, is defined as: wherein and are the current positioning accuracy and the normal positioning accuracy, respectively, and are the current perception range and the normal perception range, respectively, and are the current data acquisition frequency and the normal acquisition frequency, respectively; , , are weight coefficients, satisfying . The value range of the remaining function evaluation result is The greater the value, the stronger the remaining perceptual ability.

[0066] Step 500: Based on the remaining function evaluation result and the function-task mapping relationship, the task queue of the faulty robot is classified, and a set of degradable execution tasks and a set of tasks to be transferred are generated.

[0067] All tasks to be executed in the current task queue of the faulty robot are obtained, and the task queue is represented as: Wherein is the number of tasks. For each task , the task complexity parameters are extracted, including the required minimum number of degrees of freedom , the required workspace volume , the work accuracy requirement , and the building structure complexity .

[0068] Based on the function-task mapping relationship, it is determined whether the task belongs to the type of task that can be executed by the remaining function of the faulty robot. The function-task mapping relationship is stored through a mapping table, and the construction of the mapping table is based on the correspondence between the task type and the function requirement. For example, a planar area cleaning task only requires a 3-degree-of-freedom robot arm and basic perception ability, and can be mapped to a simplified cleaning task; a complex curved surface cleaning task requires a 6-degree-of-freedom robot arm and high-precision perception, and cannot be simplified.

[0069] Each task in the task queue is classified and determined: if and and ( is the minimum perception ability required by the task), the task is marked as a degradable execution task and added to the degradable execution task set ; otherwise, the task is marked as a task to be transferred, and added to the task set to be transferred .

[0070] The classification result is generated, including the degradable execution task set and the task set to be transferred , satisfying and .

[0071] It should be noted that the above function-task mapping relationship also includes task degradation rules. The task degradation rules define how to simplify the original task into a degraded task that matches the remaining functions. Specifically, for the case of reduced degrees of freedom of the robot arm, the degradation rules include: degrading complex surface cleaning to planar area cleaning, adjusting the work area range to cover only the planar part reachable by the robot; degrading the precision detection task to a rough detection task, reducing the detection accuracy requirement from to ; degrading the multi-angle shooting task to a single-angle shooting task, reducing the end effector pose adjustment requirement. The degradation rules are stored in the form of a rule base, including the task type before degradation, the task type after degradation, the degradation condition, and the degradation parameter adjustment method.

[0072] Step 600: executing a task transfer algorithm on the tasks in the task set to be transferred, generating a task transfer scheme; executing a task degradation algorithm on the tasks in the executable task set, generating task degradation parameters; integrating the task reassignment scheme under fault conditions.

[0073] For each task in the task set to be transferred , execute the task transfer algorithm to select the optimal receiver robot. The processing flow of the task transfer algorithm is as follows: Step 601: obtain the state information of all normally working robots in the cluster, including position coordinates, remaining battery power, current task load, and available degrees of freedom of the robot arm.

[0074] Step 602: for the task to be transferred , calculate the task transfer cost between it and each normally working robot . The task transfer cost includes the migration cost and the load impact cost, and is defined as: wherein is the spatial distance between the work position of task and the current position of robot , representing the migration cost; is the current task load (the estimated total time consumption of the assigned tasks) of robot , is the estimated time consumption of task , is the maximum task load capacity of robot , representing the load impact; and are weight coefficients. To eliminate the dimensional effect, the distance is normalized, and the normalization method is , wherein This represents the maximum spatial span of the cluster's operational area.

[0075] Step 603: Select the robot with the lowest task transfer cost as the optimal receiving robot, that is: Generate task transfer instructions to transfer the task. Remove from the task queue of the faulty robot and add to the robot. In the task queue.

[0076] Degradable task set For each task in the algorithm, a task degradation algorithm is executed to generate degradation parameters. The processing flow of the task degradation algorithm is as follows: Step 604: Execute the degradeable task Based on the remaining function assessment results Based on the degradation rules, the degraded task parameters are generated. The degradation parameters include the adjusted work area range, the reduced work accuracy requirements, and the simplified work path.

[0077] Step 605: Calculate the expected quality score of the degradation task based on the degradation parameters. This score indicates the degree of completion of the downgraded task relative to the original task. Defined as: in, and These are the areas of the work area after the downgrade and the original area, respectively. and These are the downgraded and original operational accuracy requirements, respectively. and For the weighting coefficients, satisfying . The range of values ​​is The larger the value, the closer the completion rate of the downgraded task is to the original task.

[0078] Step 606: Determine the expected quality score Is it greater than the minimum quality threshold? (In this embodiment, we take) If the value is greater than 1, the degradation plan is accepted, a task degradation instruction is generated, and the degradation parameters are updated in the task queue of the faulty robot; if the value is less than 1, the degradation plan is rejected, and the task is... Move from the set of tasks that can be degraded to the set of tasks that need to be transferred, and execute the task transfer process.

[0079] Integrate task transfer schemes and task degradation parameters to generate task redistribution schemes in case of failure. The task reassignment scheme includes: a list of tasks to be transferred and corresponding receiver robot identifiers, a list of tasks to be degraded and corresponding degradation parameters, an updated task queue of the faulty robot, and an updated task queue of each normal robot.

[0080] It should be noted that the above task transfer cost calculation also includes task type matching degree evaluation. Different robots may be equipped with different types of end effectors or have different work ability specialties, and the matching degree of task type and receiver robot ability needs to be considered when transferring tasks. Specifically, a matching degree penalty term is added to the task transfer cost: wherein, is the type matching degree of the task and the robot , and the value range is higher, indicating that the robot is more suitable for executing the task; is the matching degree penalty weight. The calculation of the matching degree is based on a corresponding table of task types and robot capabilities, for example, cleaning tasks are preferentially assigned to robots equipped with cleaning brushes, and detection tasks are preferentially assigned to robots equipped with high-precision sensors.

[0081] In the embodiments of the present application, in order to improve the dynamic adaptability of task reassignment, in step 600, real-time quality monitoring of the degraded tasks is performed, and when it is detected that the actual work quality is lower than expected, further degradation or task transfer is triggered. Specifically, during the execution of the degraded task by the faulty robot, the work quality data is collected in real time by the work quality evaluation module, including the cleanliness score of the completed area, the detection accuracy deviation and other indicators; the deviation rate of the actual quality score from the expected quality score is calculated ; it is judged whether the deviation rate exceeds a threshold value , if it exceeds, it means that the degradation scheme is not applicable and needs to be further adjusted. The adjustment strategy includes: if , it means that the remaining functions of the faulty robot cannot meet the expected quality, the task is terminated and transferred to a normal robot; if , it means that the degradation parameter is too conservative, and the work range or accuracy requirement of the degraded task can be appropriately improved to fully utilize the remaining functions. Through dynamic quality monitoring and adjustment, the actual completion quality of the degraded task is ensured to meet the requirements.

[0082] In the embodiments of the present application, in order to optimize task reassignment in the case of multiple failures, in step 600, when multiple robots in the cluster are simultaneously in failure, a joint optimization algorithm is used to coordinate the task reassignment of all failed robots, avoiding the overall efficiency loss caused by local optimization. Specifically, the residual function evaluation results and the task queue of all failed robots are obtained, and a cluster-level task reassignment optimization problem is constructed, and the objective function is the weighted sum of minimizing the total task transfer cost and maximizing the degraded task utilization: wherein, is the cluster-level task reassignment scheme, and is the weight coefficient. The constraint conditions include: each task that needs to be transferred can only be assigned to one normal robot, the task load of the normal robot does not exceed its maximum capacity, and the quality score of the degraded task is not lower than the minimum threshold. The simulated annealing algorithm is used to solve the optimization problem to obtain the globally optimal task reassignment scheme. Through joint optimization, task transfer conflicts (multiple failed robots transferring tasks to the same normal robot causing overload) and resource waste (some normal robots have too light load while other robots are overloaded) that may occur when multiple failed robots make independent decisions are avoided.

[0083] The aforementioned simulated annealing algorithm takes the objective function and the constraint conditions of the cluster-level task reassignment optimization problem as input, and outputs the globally optimal task reassignment scheme; the algorithm simulates the physical annealing process, accepts inferior solutions with a certain probability to escape from local optimum, and finally converges to the globally optimal solution as the temperature parameter gradually decreases.

[0084] In the embodiments of the present application, in order to realize the cooperation of task reassignment and the scheduling algorithm of specific embodiment 1, after generating the task reassignment scheme in step 600, the updated task queue and robot state information are fed back to the genetic algorithm optimization module to regenerate the global job scheduling scheme. Specifically, the residual function evaluation results and the degraded task information of the failed robots are input into the genetic algorithm as constraint conditions, the capability limitation of the failed robots is marked during chromosome coding, and a degraded quality penalty term is added to the tasks assigned to the failed robots during the calculation of the fitness function; the genetic algorithm re-evolves to generate an optimal scheduling scheme considering the failure constraints, which includes both normal task allocation before failure and degraded task and transferred task allocation after failure; the new scheduling scheme is issued to all robots in the cluster to guide the execution of subsequent job tasks. Through the cooperation of task reassignment and scheduling algorithm, the smooth transition of cluster jobs under failure is realized, and the overall job interruption caused by failure is avoided.

[0085] It can be understood that the data preprocessing method known to those skilled in the art includes data cleaning, data conversion, data reduction, wherein the data conversion includes type conversion and normalization and standardization, although the dimension and type of the data are ignored in the foregoing embodiment description, the data preprocessing is the technical knowledge known to those skilled in the art and the preprocessing step of data processing, and therefore the foregoing does not independently describe the known data preprocessing step.

[0086] The foregoing describes the embodiments of the present application, but the embodiments are not limited to the specific implementation described above, and the specific implementation described above is only illustrative but not restrictive, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which all belong to the protection of the embodiments.

Claims

1. A dual-algorithm driven multimodal aerial work system and intelligent control method, characterized in that, Includes the following steps: Acquire high-altitude operation environment perception data and robot status parameters to generate a set of operation task parameters. The environment perception data includes three-dimensional model data of building surface, obstacle location data, and wind speed data. The robot status parameters include robot position coordinates, remaining battery power, and robotic arm load status. The set of job task parameters is input into the genetic algorithm optimization module. By encoding the job task scheduling scheme into chromosomes, the scheduling scheme is evaluated using the fitness function and selection, crossover, and mutation operations are performed to generate a global job task scheduling scheme. Based on the global task scheduling scheme and real-time environmental perception data, the task control strategy is dynamically adjusted using a reinforcement learning algorithm to generate real-time control commands. The reinforcement learning algorithm is trained to obtain the optimal task control strategy based on the robot's current state, motion space, and reward function.

2. The method according to claim 1, characterized in that, The processing flow of the genetic algorithm optimization module includes: An initial population is generated based on the set of job task parameters. The job task scheduling scheme is encoded as a chromosome. Each chromosome represents a scheduling scheme. The encoding structure of the chromosome includes the task execution order, the assigned robot identifier, and the task start time. Calculate the fitness function value of each chromosome in the initial population. The fitness function comprehensively evaluates the total operation time, total energy consumption, and operation quality penalty of the scheduling scheme. The next generation population is generated by performing selection, crossover, and mutation operations based on the fitness function value. The selection operation calculates the probability of being selected based on the reciprocal of the chromosome fitness function value. The crossover operation exchanges gene segments at the chromosome coding position. The mutation operation randomly perturbs the gene positions in the chromosome coding position. Determine if the termination condition is met. If it is, output the optimal chromosome as the global job task scheduling scheme; otherwise, continue iterating.

3. The method according to claim 2, characterized in that, The fitness function is calculated as follows: The total operation time, total energy consumption, and operation quality penalty of the scheduling scheme are weighted and summed. The total operation time is determined by calculating the maximum time for each robot to complete all assigned tasks. The total energy consumption is determined by summing the execution energy consumption and movement energy consumption of each task and then normalizing it. The operation quality penalty is assigned a penalty value based on whether the robot's power constraints, robotic arm load constraints, building slope constraints, and safety constraints are violated. The termination condition is reaching the maximum number of iterations or the population's optimal fitness function value not improving after a preset number of consecutive generations.

4. The method according to claim 1, characterized in that, The processing flow of the reinforcement learning algorithm includes: Define the state space, action space, and reward function for reinforcement learning. The state space includes robot position, robot posture, environmental perception data, and task execution progress. The action space includes robot movement actions and task actions. The reward function comprehensively evaluates task completion rewards, energy consumption penalties, and safety rewards. A reinforcement learning model is trained using a deep Q-network algorithm. The deep neural network approximates the action value function. The input layer of the deep neural network receives the state vector, and the output layer outputs the Q value of each action. The task is executed in the simulation environment and the state transition data is stored in the experience replay buffer. Batch data is randomly sampled from the experience replay buffer and the parameters of the deep neural network are updated using the time difference error to obtain the optimal task control strategy. Based on real-time environmental perception data and the optimal operation control strategy, the current state is input into the trained reinforcement learning model to select the optimal action, and the optimal action is converted into a real-time control command and sent to the robot.

5. The method according to claim 4, characterized in that, The method for calculating the time difference error is as follows: The time difference error is obtained by multiplying the current reward plus the discount factor by the maximum Q value of the next state, and then subtracting the Q value of the current action in the current state. A loss function is constructed based on the time difference error. The loss function is defined as the expected value of the mean square error of the time difference error. An optimization algorithm is used to update the parameters of the deep neural network to make the loss function converge.

6. The method according to claim 5, characterized in that, Also includes: In the step of generating real-time control commands, when the wind speed or obstacle distance in the real-time environmental perception data changes abruptly, a rapid response is triggered to adjust the operation control strategy. When the wind speed exceeds the safety threshold, a safety braking action is forcibly executed. When the obstacle distance is less than the safety threshold, an obstacle avoidance action is forcibly executed and the local path is replanned.

7. The method according to claim 1, characterized in that, Also includes: The strategy experience learned by the reinforcement learning algorithm is fed back to the genetic algorithm optimization module. During the robot's task execution, experience data is accumulated and optimization rules for task scheduling are extracted. The optimization rules are transformed into correction parameters or constraints for the fitness function of the genetic algorithm. In the next round of task scheduling, the corrected fitness function is used to generate a global task scheduling scheme, thereby realizing the closed-loop collaboration between the genetic algorithm and the reinforcement learning algorithm.

8. The method according to claim 1, characterized in that, Also includes: When a robot experiences a partial functional failure, a fault diagnosis report of the failed robot is obtained. Based on the fault diagnosis report, the remaining functional modules are analyzed and the number of remaining available degrees of freedom, the volume of remaining workspace, and the remaining perception capability index are calculated to generate a remaining functional evaluation result. Based on the remaining function evaluation results and the function-task mapping relationship, the task queue of the faulty robot is classified. Tasks that meet the conditions for remaining function execution are marked as degradeable execution tasks, and tasks that do not meet the conditions are marked as tasks that need to be transferred, thus generating a set of degradeable execution tasks and a set of tasks that need to be transferred. For the tasks in the set of tasks to be transferred, calculate the task transfer cost between the tasks and each normal robot, and select the robot with the lowest cost as the receiving robot. For the tasks in the set of tasks that can be degraded, generate degrade parameters according to the degrade rules and calculate the expected quality score. When the expected quality score is greater than the minimum quality threshold, accept the degraded scheme and integrate to generate a task redistribution scheme in case of failure.

9. The method according to claim 8, characterized in that, The method for calculating the task transfer cost is as follows: The migration cost and load impact cost are calculated together. The migration cost is determined by the spatial distance between the task operation location and the current location of the normal robot and is normalized. The load impact cost is determined by the sum of the current task load of the normal robot and the estimated time of the task to be transferred, divided by the maximum task load capacity of the robot. The task transfer cost is obtained by weighted summation of the migration cost and the load impact cost. The expected quality score is calculated by comprehensively evaluating the ratio of the downgraded work area to the original work area and the ratio of the original work accuracy requirement to the downgraded work accuracy requirement, and then weighting and summing the two ratios to obtain the expected quality score.

10. A dual-algorithm driven multimodal aerial work system and intelligent control system, used to execute the intelligent control method according to any one of claims 1 to 9, characterized in that, The system includes: The movement and positioning module is used to enable the robot to move and position itself on the building surface. The task execution module, including a multi-degree-of-freedom robotic arm and a replaceable end effector, is used to perform high-altitude work tasks; The environmental perception and safety module includes a wind speed sensor, an ultrasonic sensor, a vision sensor, and a collision detection sensor, used for real-time environmental perception and safety protection. The data transmission and control module includes a wireless communication module and an industrial computer for data transmission and robot control. The industrial computer integrates a genetic algorithm optimization module and a reinforcement learning algorithm module. The energy supply module is used to provide energy support for the robot.

Images