Artificial intelligence-based method for predicting task travel time of heterogeneous robot systems

Abstract

The present invention relates to a method for predicting the task travel time of a heterogeneous robot system based on artificial intelligence, and belongs to the field of intelligent manufacturing. First, the topological structure of the neural network is determined according to the task travel time characteristics of the heterogeneous robot system, and the input / output dimension of the neural network is determined according to the dynamic input / output strategy. Secondly, according to the improved mayfly algorithm, the initial parameters of the neural network are determined, and the neural network parameters are updated by a selective update strategy. When the termination condition of the neural network is reached, the trained neural network model is obtained, and a test sample is input to test the performance of the network. Finally, the real-time information of the system is input into the tested neural network to predict the task travel time of the system. This method solves the problems of low prediction accuracy and poor adaptability that may be caused by the task travel time prediction method in the prior art, improves the prediction accuracy of the task travel time, and fills the gap in the task travel time prediction technology of the heterogeneous robot system.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control, and in particular to a method for predicting task travel time of a heterogeneous robot system based on artificial intelligence. Background Art

[0002] The application of robot systems in the field of intelligent manufacturing has achieved great success, greatly improving work efficiency and resource utilization. With the development of the manufacturing industry and the diversification of demand, heterogeneous robot systems are becoming more and more popular to solve practical problems such as a wide variety of tasks and large differences in task requirements. It consists of different types of robots, each type of robot has unique performance and functions to meet different task requirements. Therefore, compared with a single type of robot system, a heterogeneous robot system has better flexibility, stability and scalability.

[0003] Although heterogeneous robots can improve the flexibility and intelligence of the system, the task travel time is difficult to predict due to the dynamic and correlation of the tasks of the heterogeneous robot system. However, accurate prediction of the task travel time is extremely important for the efficient operation of the heterogeneous robot system. However, after reviewing the existing literature related to robot system scheduling, there are few studies on the task travel time prediction problem of the scheduling system, and the task travel time prediction of the scheduling system is based on empirical formulas, such as the Chinese invention patent with publication number CN114911198A. However, there is no relevant research literature on the prediction of robot task travel time based on data-driven and artificial intelligence methods. Summary of the invention

[0004] In order to solve the technical defects existing in the prior art, the present invention proposes a task travel time prediction method for a heterogeneous robot system based on artificial intelligence, which combines an improved neural network method and an intelligent optimization method to predict the robot's task travel time, thereby solving the problems of low prediction accuracy and poor adaptability that may be caused by the task travel time prediction method in the prior art, improving the task travel time prediction accuracy, and filling the gap in the task travel time prediction technology for heterogeneous robot systems.

[0005] The present invention is achieved through the following technical solutions:

[0006] A method for predicting task travel time of a heterogeneous robot system based on artificial intelligence includes the following steps:

[0007] Step 1: Determine the topology of the neural network based on the task travel time characteristics of the heterogeneous robot system;

[0008] Step 2: Determine the input / output dimension of the neural network according to the dynamic input / output strategy;

[0009] Step 3: Determine the initial parameters of the neural network according to the improved mayfly algorithm;

[0010] Step 4: Calculate the deviation between the output value of the neural network and the actual value based on the input training sample and the neural network parameters;

[0011] Step 5: Update the neural network parameters using a selective update strategy;

[0012] Step 6: Determine whether the neural network termination condition is met, if so, go to step 7, if not, go to step 4;

[0013] Step 7: Get the trained neural network model and input the test sample to test the performance of the network. Finally, input the real-time information of the system into the tested neural network to predict the mission travel time of the system.

[0014] Furthermore, the heterogeneous robot system is a system composed of robots with different functions and robots with different load capacities.

[0015] Furthermore, the robots with different functions include but are not limited to robots with different movement modes or different navigation modes.

[0016] Furthermore, the robots with different load capacities include single-load capacities robots and multi-load capacities robots.

[0017] Furthermore, the task travel time characteristic is that the task travel time of the heterogeneous robot system is not only related to the task information and the robot status, but also to the traffic flow in the system.

[0018] Furthermore, the neural network includes but is not limited to a BP neural network and a RBF neural network.

[0019] Furthermore, the topological structure of the neural network includes an input layer X, a hidden layer Z and an output layer Y, a total of three layers, with the factors affecting the travel as input and the task travel time as output. The input layer X represents the task status of the system at a certain moment, the traffic load of the path and the load of the robot. The output layer Y represents the travel time of each task in the system, and the tasks include the scheduled but unfinished tasks and the tasks to be scheduled in this round.

[0020] Furthermore, the factors affecting the trip include the priority of the task, the robot number assigned to the task, the execution order of the task on the robot, the path information of the robot, the task information being executed by the robot, and the task hierarchy.

[0021] Furthermore, the task level refers to the scheduled but unfinished tasks or the tasks to be scheduled in this round.

[0022] Furthermore, the dynamic input / output strategy sets the total number of elements n of X to the maximum number of tasks N that may exist in the system. n The product of the number of factors affecting the trip l, in X, every l elements from the front to the back together represent the factors affecting the trip of a task, and the total number of elements q of Y is set to the number of existing tasks N in the system n The number of neurons in the hidden layer is determined by an empirical formula.

[0023] Furthermore, the improved mayfly algorithm comprises the following steps:

[0024] Step 1.1: Determine the dimension of the mayfly population according to the dimension of the neural network parameters, and initialize the mayfly population;

[0025] Step 1.2: Convert each individual in the mayfly population into a neural network parameter to form a neural network, input a training sample to each neural network for training, and use the mean square error obtained from the neural network training as the fitness value;

[0026] Step 1.3: Evaluate the fitness value of each mayfly, and update the historical optimal position of the individual mayfly and the global optimal position of the mayfly population;

[0027] Step 1.4: Update the speed and position of male and female mayflies according to the update formula of the mayfly algorithm;

[0028] Step 1.5: Use the improved crossover operation and multi-level mutation operation to form a descendant population;

[0029] Step 1.6: Use the elite retention strategy to form the next generation parent population;

[0030] Step 1.7: Determine whether the termination condition of the Mayfly algorithm is met. If so, convert the global optimal solution into the initial parameters of the neural network. If not, go to step 1.2.

[0031] Furthermore, the dimension of the mayfly population is the product of the total number of neural network parameters and the dimension of each parameter.

[0032] Furthermore, the mean square error calculation method is:

[0033]

[0034] Where N is the total number of training samples, a is the training sample index, is the true value of the kth neuron, y k is the output value of the kth neuron, which is obtained according to the calculation formula of the neural network output value.

[0035] Furthermore, the improved crossover operation comprises the following steps:

[0036] Step 2.1: Select a parent from the male population and a parent from the female population according to the selection method;

[0037] Step 2.2: Use the mating method of the traditional mayfly algorithm as the first crossover operation to obtain two offspring individuals;

[0038] Step 2.3: Randomly select a dimension rand from all dimensions D of the parent mayfly individual vector as the intersection point;

[0039] Step 2.4: Assign the value of the rand dimension in the male mayfly to the first offspring, and the values ​​of the remaining dimensions to the second offspring. Similarly, assign the value of the rand dimension in the female mayfly to the second offspring, and the values ​​of the remaining dimensions to the first offspring.

[0040] Step 2.5: Determine whether the number of selected intersections is greater than or equal to the crossover probability P C and the product of the individual dimension D of the mayfly. If yes, proceed to step 2.6. If no, proceed to step 2.3.

[0041] Step 2.6: Get two offspring individuals of the second crossover operation;

[0042] Step 2.7: Determine whether the number of offspring individuals is sufficient. If so, proceed to step 2.8. If not, proceed to step 2.1.

[0043] Step 2.8: Compare the fitness values ​​of the individuals in the offspring populations obtained from the two crossover operations, and obtain the optimal offspring population from the two crossover operations.

[0044] Furthermore, the selection method includes, but is not limited to, a random selection method and a selection method based on fitness value.

[0045] Furthermore, the individual vector of the mayfly is r=1,2,…,N A , where r is the individual index of mayflies, N A is the size of the male or female population.

[0046] Furthermore, the multi-level mutation operation comprises the following steps:

[0047] Step 3.1: Randomly select an individual from the offspring population obtained by the improved crossover operation as the parent of the mutation operation;

[0048] Step 3.2: Introduce the mutation operation of the genetic algorithm to perform a single-level mutation operation to obtain a progeny individual;

[0049] Step 3.2: Determine whether the optimal individual of the mayfly population has not changed for consecutive f iterations. If so, proceed to step 3.3; if not, proceed to step 3.4.

[0050] Step 3.3: Use the offspring formed by the single-level mutation as the parent of the multi-level mutation, and perform the mutation operation again, that is, the multi-level mutation operation, to form a new offspring individual;

[0051] Step 3.4: Determine whether the number of offspring individuals is sufficient. If so, proceed to step 3.5. If not, proceed to step 3.1.

[0052] Step 3.5: Compare the fitness values ​​of the offspring population individuals obtained by the two mutation operations, and obtain the optimal offspring population from the two mutation operations;

[0053] Step 3.6: Calculate the fitness value of this iteration and obtain the optimal individual of the mayfly population. If the optimal individual has not changed, f=f+1, otherwise, f=1, and record the optimal individual and its fitness value.

[0054] Furthermore, the single-stage mutation operation is to perform a mutation operation only once, and the operator of the mutation operation is a genetic mutation operator.

[0055] Furthermore, the elite retention strategy is to first merge the updated male population with the first half of the individuals in the mutated offspring population into a new parent male population. The individuals in the new male population are ranked by fitness, and the optimal N A The individuals are used as the male population of the next generation parents. Similarly, the updated female population and the second half of the individuals of the mutated offspring population are merged, and the fitness values ​​are sorted and screened to obtain the female population of the next generation parents.

[0056] Furthermore, the termination conditions of the mayfly algorithm include reaching a set number of iterations and the optimal solution remaining unchanged for f1 consecutive generations, f1>f.

[0057] Furthermore, the selective update strategy is to ignore invalid data in each input sample during network training and only update the weight parameters of the rows with valid data in each sample.

[0058] Furthermore, the invalid data is that the topology of the neural network model is fixed, and its input dimension and output dimension are determined by the maximum number of tasks that may exist in the system. At different times, the number of tasks in the system is different. Therefore, when the actual number of tasks in the system is less than the maximum possible number of tasks in the system, for some samples, the data of some rows in the input matrix may be all zero. These zero data cannot describe any system characteristics and are called invalid data.

[0059] Furthermore, the valid data refers to the data in some rows of the input matrix in the input sample that are not all zero, and the data in these rows are called valid data.

[0060] Furthermore, the neural network termination conditions include reaching a set number of iterations and the mean square error reaching a set value.

[0061] Compared with the prior art, the present invention has at least the following beneficial effects or advantages:

[0062] The present invention proposes a method for predicting the task travel time of a heterogeneous robot system based on artificial intelligence, which takes into account the task characteristics of the heterogeneous robot system and the defects of the traditional neural network, solves the problems of low prediction accuracy and poor adaptability that may be caused by the traditional task travel time prediction method based on empirical formulas, improves the prediction accuracy of the task travel time, and fills the gap in the task travel time prediction technology of the heterogeneous robot system. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The present invention will be further described in detail below with reference to the accompanying drawings:

[0064] Figure 1 The present invention is a flowchart of a method for predicting task travel time of a heterogeneous robot system based on artificial intelligence.

[0065] Figure 2 It is the topological structure of the neural network of the present invention.

[0066] Figure 3 The flowchart of the improved mayfly algorithm of the present invention is shown in FIG. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] Since the heterogeneous robot system is relatively complex, the following assumptions are made for the system:

[0069] 1. The robot is a mobile robot. All multi-load robots are loaded at a fixed location, and then the multi-load robots carry and transport at a uniform speed according to the planned unloading sequence;

[0070] 2. Once the robot starts transporting, it will not change the path or adjust the order of task unloading on its own;

[0071] 3. The idle multi-load robot stays at a specific task waiting point to wait for the next material task, and does not affect the passage of other multi-load robots in the path network;

[0072] 4. The system path network is a one-way single lane and does not support overtaking;

[0073] 5. The multi-capacity robot loads trailers from front to back and unloads trailers in a first-in-last-out manner.

[0074] 6. The unloading time of multi-capacity robots is fixed. If there are multiple robots at the same loading and unloading point, they need to wait in line.

[0075] 7. Do not use path replanning methods during traffic control to avoid motion conflicts in the robot system.

[0076] Figure 1 The flowchart of the method for predicting the task travel time of a heterogeneous robot system based on artificial intelligence of the present invention is as follows. The method for predicting the task travel time of a heterogeneous robot system based on artificial intelligence specifically comprises the following steps:

[0077] Step 1: Determine the topology of the neural network based on the task travel time characteristics of the heterogeneous robot system;

[0078] The task travel time of the heterogeneous robot system has the characteristics of being affected by the task information, the robot state and the traffic flow in the system at the same time. According to the task travel time characteristics of the heterogeneous robot system, the topological structure of the neural network is determined. The topological structure of the neural network includes an input layer X, a hidden layer Z and an output layer Y, a total of 3 layers. The task travel influencing factors are used as inputs, and the task travel time is used as outputs. The input layer X represents the task status of the system at a certain moment, the traffic load status of the path and the load status of the robot. The output layer Y represents the travel time of each task in the system. The tasks include the scheduled but unfinished tasks and the tasks to be scheduled in this round. In the embodiment of the present invention, the travel influencing factors include the priority of the task, the robot number assigned to the task, the execution order of the task on the robot, the path information of the robot, the task information being executed by the robot and the task level. The task level is the scheduled but unfinished tasks or the tasks to be scheduled in this round.

[0079] In an embodiment of the present invention, a heterogeneous robot system is a system composed of robots with different functions and robots with different load capacities. The robots with different functions include but are not limited to robots with different movement modes or different navigation modes, such as omnidirectional robots, differential robots, trailer robots, visually guided robots, magnetically guided robots, and laser-guided robots. Robots with different load capacities include single-load capacities robots and multi-load capacities robots. Neural networks include but are not limited to BP neural networks and RBF neural networks.

[0080] It should be noted that a heterogeneous robot system may only include robots with different functions or robots with different load capacities. The robots with different functions may not be the robots listed in the embodiments of the present invention. For example, the robots may also be laser robots, visual robots, and magnetically guided robots. The robots with different load capacities may not be the robots listed in the embodiments of the present invention, such as collaborative transport robots, in which multiple robots transport one workpiece.

[0081] Step 2: Determine the input / output dimension of the neural network according to the dynamic input / output strategy;

[0082] The factors affecting the travel are used as the input factors of the prediction model. If the factors affecting the travel are different, the input factors of the prediction model will also be different. Since the input / output dimensions of the heterogeneous robot system will change with the dynamic changes in the number of tasks, the input / output dimensions of the neural network are determined according to the dynamic input / output strategy. Specifically, the total number of elements n of X is set to the maximum number of tasks N that may exist in the system. n The product of the number of factors affecting the trip l, in X, every l elements from the front to the back together represent the factors affecting the trip of a task, and the total number of elements q of Y is set to the number of existing tasks N in the system n The number of neurons in the hidden layer is determined by an empirical formula.

[0083] Step 3: Determine the initial parameters of the neural network according to the improved mayfly algorithm;

[0084] After the topological structure and input / output dimensions of the neural network are determined, the initial parameters of the neural network are determined according to the improved mayfly algorithm. After the initial parameters optimized by the improved mayfly algorithm are obtained, step 4 is entered.

[0085] Step 4: Calculate the deviation between the output value of the neural network and the actual value based on the input training sample and the neural network parameters;

[0086] After the initial parameters of the neural network are determined, the training samples are input to start the neural network training. First, the deviation between the output value of the neural network and the actual value is calculated based on the input training samples and the neural network parameters.

[0087] It should be noted that the calculation formula of the neural network output value varies depending on the selected neural network. For example, the calculation formula of the neural network output value of the RBF neural network is related to the radial basis function.

[0088] Step 5: Update the neural network parameters using a selective update strategy;

[0089] The topological structure of the neural network model is fixed, and its input dimension and output dimension are determined by the maximum number of tasks that may exist in the system. At different times, the number of tasks in the system is different. Therefore, when the actual number of tasks in the system is less than the maximum possible number of tasks in the system, for some samples, the data of some rows in the input matrix may be all zero. These zero data cannot describe any system characteristics and are called invalid data. On the contrary, the data of some rows in the input matrix in the input sample are not all zero, and the data of these rows are called valid data. Therefore, a selective update strategy is used to update the neural network parameters, and the specific update method is the gradient descent method. The selective update strategy is specifically to ignore the invalid data in each input sample during the network training process, and only update the weight parameters of the rows with valid data in each sample.

[0090] Step 6: Determine whether the neural network termination condition is met, if so, go to step 7, if not, go to step 4;

[0091] Determine whether the neural network termination condition is met. If the termination condition is met, go to step 7. If the termination condition is not met, return to step 4 to continue iterative optimization.

[0092] In the embodiment of the present invention, the neural network termination conditions include reaching a set number of iterations and the mean square error reaching a set value.

[0093] It should be noted that the neural network termination condition may also only include reaching a set number of iterations or the mean square error reaching a set value, and the neural network termination condition may not be the termination condition listed in the embodiments of the present invention, such as all training samples are repeated several times to complete the training.

[0094] Step 7: Obtain the trained neural network model and input test samples to test the performance of the network. Finally, input the real-time information of the system into the tested neural network to predict the task travel time of the system.

[0095] If the neural network termination condition is met, the neural network training is terminated, the trained neural network model is obtained, and the test sample is input to test the performance of the network. Finally, the real-time information of the system is input into the tested neural network to predict the task travel time of the system.

[0096] It should be noted that when training, testing neural networks and using neural network predictions, only valid data of input samples are used to calculate mean square error and prediction value, which improves the calculation speed of RBF neural network.

[0097] The improved mayfly algorithm includes the following steps:

[0098] Step 1.1: Determine the dimension of the mayfly population according to the dimension of the neural network parameters, and initialize the mayfly population;

[0099] The dimension of the mayfly population is determined according to the dimension of the neural network parameters, that is, the sum of the dimensions of all neural network parameters. The dimension of the mayfly population is the product of the total number of neural network parameters and the dimension of each parameter. The mayfly population is initialized according to the dimension of the mayfly population, including the size of the male and female mayfly populations, the position and speed of each individual mayfly, and the range of the position and speed values ​​is set.

[0100] Step 1.2: Convert each individual in the mayfly population into a neural network parameter to form a neural network, input a training sample to each neural network for training, and use the mean square error obtained from the neural network training as the fitness value;

[0101] Each mayfly individual represents a set of neural network parameters. Each individual in the mayfly population is converted into a neural network parameter to form a neural network. Each neural network is trained by inputting training samples, and the mean square error obtained from the neural network training is used as the fitness value.

[0102] The mean square error is calculated as:

[0103]

[0104] Where N is the total number of training samples, a is the training sample index, is the true value of the kth neuron, y k is the output value of the kth neuron, which is obtained according to the calculation formula of the neural network output value.

[0105] Step 1.3: Evaluate the fitness value of each mayfly, and update the historical optimal position of the individual mayfly and the global optimal position of the mayfly population;

[0106] Evaluate the fitness value of each mayfly, that is, compare the current fitness value of each individual mayfly with its historical optimal fitness value. If the current fitness value is better, then this value becomes the new historical optimal fitness value of the individual mayfly, and the historical optimal position of the individual mayfly is updated to the current position of the individual mayfly; then compare the optimal fitness values ​​of all individuals in the mayfly population to determine the global optimal mayfly individual and the optimal fitness value, and update the global optimal position of the mayfly population to the current position of the global optimal mayfly individual.

[0107] Step 1.4: Update the speed and position of male and female mayflies according to the update formula of the mayfly algorithm;

[0108] According to the updating formula of the mayfly algorithm, the speed and position of the male mayfly and the speed and position of the female mayfly are updated respectively. The speed updating formulas of the male mayfly individual and the female mayfly individual in the updating formula of the mayfly algorithm are different.

[0109] Step 1.5: Use the improved crossover operation and multi-level mutation operation to form a descendant population;

[0110] The crossover operation of the mayfly algorithm itself has certain limitations. In order to increase the diversity of the population, an improved crossover operation is proposed based on the existing crossover operator of the mayfly algorithm. After the offspring population is obtained by the improved crossover operation, the offspring population is subjected to multi-level mutation operations to form a new offspring population.

[0111] Step 1.6: Use the elite retention strategy to form the next generation parent population;

[0112] After obtaining the mutated offspring population, in order to select better individuals for the next iteration, the elite retention strategy is used to form the next generation parent population.

[0113] In the embodiment of the present invention, the elite retention strategy is to first merge the updated male population with the first half of the individuals in the mutated offspring population into a new parent male population. The individuals in the new male population are ranked by fitness, and the optimal N A The individuals are used as the male population of the next generation parents. Similarly, the updated female population and the second half of the individuals of the mutated offspring population are merged, and the fitness values ​​are sorted and screened to obtain the female population of the next generation parents.

[0114] Step 1.7: Determine whether the termination condition of the Mayfly algorithm is met. If so, convert the global optimal solution into the initial parameters of the neural network. If not, go to step 1.2.

[0115] After the algorithm completes one iteration, it is determined whether the termination condition of the mayfly algorithm is met. If so, the global optimal mayfly individual, that is, the global optimal solution, is determined according to the fitness value of the mayfly individual, and the global optimal solution is converted into the initial parameters of the neural network. The training samples are input to start the neural network training. If not, go to step 1.2.

[0116] In the embodiment of the present invention, the termination conditions of the mayfly algorithm include reaching a set number of iterations and the optimal solution remaining unchanged for f1 consecutive generations, f1>f, where f is the number of generations in which the optimal individual of the mayfly population has not changed continuously.

[0117] It should be noted that the termination condition of the mayfly algorithm may also only include reaching a set number of iterations or the optimal solution remaining unchanged for f1 consecutive generations, and the neural network termination condition may not be the termination condition listed in the embodiment of the present invention, such as the optimal fitness value of the mayfly algorithm reaching a set value.

[0118] The improved crossover operation includes the following steps:

[0119] Step 2.1: Select a parent from the male population and a parent from the female population according to the selection method;

[0120] Since the offspring population of the mayfly algorithm is obtained by mating male mayfly individuals and female mayfly individuals, it is necessary to select a parent from the male population and the female population respectively according to the selection method before mating.

[0121] In the embodiment of the present invention, the selection method includes but is not limited to a random selection method and a selection method based on fitness value. In the selection method based on fitness value, the best female mayfly individual mates with the best male mayfly individual, the second best female mayfly individual mates with the second best male mayfly individual, and so on.

[0122] Step 2.2: Use the mating method of the traditional mayfly algorithm as the first crossover operation to obtain two offspring individuals;

[0123] The improved crossover operation includes two crossover methods. The mating method of the traditional mayfly algorithm is used as the first crossover operation. Each crossover can obtain two offspring individuals offspring2 and offspring2.

[0124] The mating method of the traditional mayfly algorithm is as follows:

[0125]

[0126] Where male is the father, female is the mother, L is a random number in the range [-1,1], and · represents the matrix multiplication.

[0127] Step 2.3: Randomly select a dimension rand from all dimensions D of the parent mayfly individual vector as the intersection point;

[0128] After the first crossover operation is completed, the second crossover operation begins. First, a dimension rand is randomly selected from all dimensions D of the parent generation mayfly individual vector as the crossover point, that is, rand∈[1,D]

[0129] In the embodiment of the present invention, the individual vector of the mayfly can be expressed as r=1,2,…,N A , where r is the individual index of mayflies, N A is the size of the male or female population.

[0130] Step 2.4: Assign the value of the rand dimension in the male mayfly to the first offspring, and the values ​​of the remaining dimensions to the second offspring. Similarly, assign the value of the rand dimension in the female mayfly to the second offspring, and the values ​​of the remaining dimensions to the first offspring.

[0131] After determining the crossover point, the value of the rand dimension in the male mayfly is assigned to the first offspring, and the values ​​of the remaining dimensions are assigned to the second offspring. Similarly, the value of the rand dimension in the female mayfly is assigned to the second offspring, and the values ​​of the remaining dimensions are assigned to the first offspring, thus producing two offspring individuals.

[0132] Step 2.5: Determine whether the number of selected intersections is greater than or equal to the crossover probability P C and the product of the individual dimension D of the mayfly. If yes, proceed to step 2.6. If not, proceed to step 2.3.

[0133] The second crossover operation requires multiple crossovers, and it is necessary to determine whether the number of crossovers is sufficient, that is, whether the number of selected crossover points is greater than or equal to the crossover probability P. C and the product of the dimension D of the mayfly individual. If so, the second crossover operation is completed and the process goes to step 2.6. If not, the two obtained offspring individuals are used as the fathers and the process goes to step 2.3 to perform crossover again.

[0134] Step 2.6: Get two offspring individuals of the second crossover operation;

[0135] When the number of selected crossover points is greater than or equal to the crossover probability P C When the product of the dimension D of the mayfly individual is multiplied by , the second crossover operation ends, and two offspring individuals of the second crossover operation are obtained.

[0136] Step 2.7: Determine whether the number of offspring individuals is sufficient. If so, proceed to step 2.8. If not, proceed to step 2.1.

[0137] After the two crossover operations are completed, it is necessary to determine whether the number of offspring individuals generated by the two crossover operations is sufficient. If so, proceed to step 2.8. If not, proceed to step 2.1 and perform a new round of crossover operations again to generate more offspring individuals.

[0138] Step 2.8: Compare the fitness values ​​of the individuals in the offspring populations obtained from the two crossover operations, and obtain the optimal offspring population from the two crossover operations.

[0139] When the number of offspring individuals generated by the two crossover operations meets the requirements, a offspring population of the two crossover operations is obtained. The fitness values ​​of the offspring population individuals obtained by the two crossover operations are compared, and the optimal offspring population is obtained from the two crossover operations and used as the offspring population for this improved crossover operation.

[0140] In the embodiment of the present invention, the number of individuals in the offspring population of the improved crossover operation is half the number of individuals in the offspring population of the original two crossover operations.

[0141] It should be noted that the number of individuals in the offspring population of the improved crossover operation can be half of the number of individuals in the offspring population of the original two crossover operations, or can be other values, which are determined according to requirements.

[0142] The multi-level mutation operation includes the following steps:

[0143] Step 3.1: Randomly select an individual from the offspring population obtained by the improved crossover operation as the parent of the mutation operation;

[0144] After the improved crossover operation is completed, the multi-level mutation operation is started on the individuals in the offspring population obtained by the improved crossover operation. First, an individual is randomly selected from the offspring population obtained by the improved crossover operation as the parent of the mutation operation.

[0145] It should be noted that the parent selection method of the mutation operation may not be the random selection method listed in the embodiment of the present invention, such as a selection method based on fitness value.

[0146] Step 3.2: Introduce the mutation operation of the genetic algorithm to perform a single-level mutation operation to obtain a progeny individual;

[0147] A genetic algorithm mutation operation is introduced to the parent mayfly individual, that is, a single-stage mutation operation is performed to obtain a child individual.

[0148] In the embodiment of the present invention, the single-stage mutation operation is to perform a mutation operation only once, and the operator of the mutation operation is a genetic mutation operator.

[0149] Step 3.2: Determine whether the optimal individual of the mayfly population has not changed for consecutive f iterations. If so, proceed to step 3.3; if not, proceed to step 3.4.

[0150] If the single-level mutation search range is small and cannot jump out of the local optimal solution, multiple mutation operations are required. First, it is necessary to judge whether the single-level mutation can jump out of the local optimal solution based on the change of the optimal individual of the mayfly population, that is, to judge whether the optimal individual of the mayfly population has not changed for f consecutive iterations. If so, go to step 3.3 and perform multi-level mutation operations. If not, go to step 3.4.

[0151] Step 3.3: Use the offspring formed by the single-level mutation as the parent of the multi-level mutation, and perform the mutation operation again, that is, the multi-level mutation operation, to form a new offspring individual;

[0152] When the optimal individual of the mayfly population has not changed for f consecutive iterations, multi-level mutation is performed. That is, the offspring formed by single-level mutation is used as the parent of multi-level mutation, and the mutation operation is performed again to form a new offspring individual.

[0153] Step 3.4: Determine whether the number of offspring individuals is sufficient. If so, proceed to step 3.5. If not, proceed to step 3.1.

[0154] After each single-stage mutation operation or multi-stage mutation operation, a progeny individual will be obtained. It is necessary to determine whether the number of all progeny individuals obtained by the mutation operation is sufficient. If so, go to step 3.5. If not, go to step 3.1 and perform a new round of mutation operation again.

[0155] Step 3.5: Compare the fitness values ​​of the offspring population individuals obtained by the two mutation operations, and obtain the optimal offspring population from the two mutation operations;

[0156] When the number of all offspring individuals obtained by the mutation operation is sufficient, a offspring population of two mutation operations is obtained. The fitness values ​​of the offspring population individuals obtained by the two mutation operations are compared to obtain the optimal offspring population from the two mutation operations.

[0157] Step 3.6: Calculate the fitness value of this iteration and obtain the optimal individual of the mayfly population. If the optimal individual has not changed, f=f+1, otherwise, f=1, and record the optimal individual and its fitness value.

[0158] Calculate the fitness value of the offspring mayfly individuals in this round of iteration, and obtain the best individuals in the parent mayfly population and the offspring mayfly population. If the best individual has not changed, that is, the global best mayfly individual has not changed, then the number of generations in which the best individual of the mayfly population has not changed continuously is f=f+1. If not, the global best mayfly individual has changed, f=1, and start recording the number of generations in which the best individual of the mayfly population has not changed continuously, and record the best individual and its fitness value.

[0159] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be regarded as within the protection scope of the present invention.

Claims

1. A method for predicting task travel time of heterogeneous robot systems based on artificial intelligence, characterized in that: Includes steps: Step 1: Determine the topology of the neural network based on the task travel time characteristics of the heterogeneous robot system; Step 2: Determine the input / output dimension of the neural network according to the dynamic input / output strategy; Step 3: Determine the initial parameters of the neural network according to the improved mayfly algorithm, the improved mayfly algorithm; Step 4: Calculate the deviation between the output value of the neural network and the actual value based on the input training sample and the neural network parameters; Step 5: Update the neural network parameters using a selective update strategy; Step 6: Determine whether the neural network termination condition is met, if so, go to step 7, if not, go to step 4; Step 7: Get the trained neural network model, input the test sample to test the performance of the network, and finally input the real-time information of the system into the tested neural network to predict the task travel time of the system; The task travel time characteristic is that the task travel time of the heterogeneous robot system includes task information, robot status, and traffic flow in the system; The topology of the neural network, including the input layer , hidden layer and output layer , a total of 3 layers, with the factors affecting the trip as input and the task travel time as output. It represents the task status of the system at a certain moment, the traffic load of the path, and the load of the robot. The output layer It represents the travel time of each task in the system, including the scheduled but unfinished tasks and the tasks to be scheduled in this round; The dynamic input / output strategy is The total number of elements Set to the maximum number of tasks that may exist in the system Number of factors affecting the itinerary The product of From the beginning to the end, The elements together represent the factors affecting the itinerary of a task. The total number of elements Set to the number of existing tasks in the system Equal, the number of neurons in the hidden layer is determined by an empirical formula; The improved mayfly algorithm comprises the following steps: Step 1.1: Determine the dimension of the mayfly population according to the dimension of the neural network parameters, and initialize the mayfly population; Step 1.2: Convert each individual in the mayfly population into a neural network parameter to form a neural network, input a training sample to each neural network for training, and use the mean square error obtained from the neural network training as the fitness value; Step 1.3: Evaluate the fitness value of each mayfly, and update the historical optimal position of the individual mayfly and the global optimal position of the mayfly population; Step 1.4: Update the speed and position of male and female mayflies according to the update formula of the mayfly algorithm; Step 1.5: Use the improved crossover operation and multi-level mutation operation to form a descendant population; Step 1.6: Use the elite retention strategy to form the next generation parent population; Step 1.7: Determine whether the termination condition of the Mayfly algorithm is met. If so, convert the global optimal solution into the initial parameters of the neural network. If not, go to step 1.2; The improved crossover operation comprises the following steps: Step 2.1: Select a parent from the male population and a parent from the female population according to the selection method; Step 2.2: Use the mating method of the traditional mayfly algorithm as the first crossover operation to obtain two offspring individuals; Step 2.3: From all dimensions of the parent mayfly individual vector D Randomly select a dimension As a point of intersection; Step 2.4: Separate the male mayflies into The value of the dimension is assigned to the first offspring, and the value of the remaining dimension is assigned to the second offspring. Similarly, the value of the first dimension in the female mayfly is assigned to the second offspring. The values ​​on the dimension are assigned to the second child, and the values ​​on the remaining dimensions are assigned to the first child; Step 2.5: Determine whether the number of selected intersections is greater than or equal to the crossover probability and the ephemera individual dimension D If yes, go to step 2.6, if no, go to step 2.3; Step 2.6: Get two offspring individuals of the second crossover operation; Step 2.7: Determine whether the number of offspring individuals is sufficient. If so, proceed to step 2.

8. If not, proceed to step 2.

1. Step 2.8: Compare the fitness values ​​of the individuals in the offspring populations obtained from the two crossover operations, and obtain the optimal offspring population from the two crossover operations; The multi-level mutation operation includes the following steps: Step 3.1: Randomly select an individual from the offspring population obtained by the improved crossover operation as the parent of the mutation operation; Step 3.2: Introduce the mutation operation of the genetic algorithm to perform a single-level mutation operation to obtain a progeny individual; Step 3.2: Determine whether the optimal individuals of the mayfly population are continuous No change occurs in the iteration. If yes, go to step 3.

3. If no, go to step 3.

4. Step 3.3: Use the offspring formed by the single-level mutation as the parent of the multi-level mutation, and perform the mutation operation again, that is, the multi-level mutation operation, to form a new offspring individual; Step 3.4: Determine whether the number of offspring individuals is sufficient. If so, proceed to step 3.

5. If not, proceed to step 3.

1. Step 3.5: Compare the fitness values ​​of the offspring population individuals obtained by the two mutation operations, and obtain the optimal offspring population from the two mutation operations; Step 3.6: Calculate the fitness value of this iteration and obtain the optimal individual of the mayfly population. If the optimal individual does not change, ,if not, , and record the best individual and the fitness value of the best individual.

2. The method for predicting task travel time of heterogeneous robot system based on artificial intelligence according to claim 1 is characterized in that: The heterogeneous robot system is a system composed of robots with different functions and robots with different load capacities.

3. The method for predicting task travel time of heterogeneous robot system based on artificial intelligence according to claim 1, characterized in that: The factors affecting the trip include the priority of the task, the robot number assigned to the task, the execution order of the task on the robot, the path information of the robot, the task information being executed by the robot, and the task level.

4. The method for predicting task travel time of heterogeneous robot system based on artificial intelligence according to claim 1, characterized in that: The selective update strategy is to ignore invalid data in each input sample during network training and only update the weight parameters of the rows with valid data in each sample. The topological structure of the neural network model is fixed, and its input dimension and output dimension are determined by the maximum number of tasks in the system. At different times, the number of tasks in the system is different. When the actual number of tasks in the system is less than the maximum number of tasks in the system, the sample input matrix has R Zero data for rows where R It is the difference between the maximum number of tasks and the actual number of tasks.

Images