Multi-underwater robot task allocation method and system based on self-organizing neural network

Abstract

The present invention discloses a multi-underwater robot task allocation method and system based on a self-organizing neural network, belonging to the technical field of multi-agent system collaborative control. The method constructs a multi-constrained task allocation optimization model for a multi-heterogeneous underwater robot system based on multiple constraints such as ocean current influence, resource consumption, load balancing, and limited resources. The task allocation optimization model is a self-organizing neural network; the task allocation optimization model completes task allocation through competition and learning between neurons corresponding to robot path points, ultimately generating the required circular path for each robot. The method has the advantages of fast solution speed and high solution accuracy, and can provide a good task allocation solution within a limited number of iterations. The advantages of the method proposed by the present invention are particularly evident in complex large-scale task scenarios.

Description

Technical Field

[0001] The present invention belongs to the field of multi-agent system collaborative control, and in particular relates to a multi-underwater robot task allocation method and system based on a self-organizing neural network. Background Art

[0002] Underwater robots (AUVs), one of the primary tools for human exploration of marine resources, environmental monitoring, and information collection, are gaining widespread attention and application. However, due to inherent limitations of a single AUV, such as limited onboard resources and limited internal power, they often struggle to meet increasingly complex mission demands. Compared to a single AUV, multi-AUVs have emerged as a new application. By collaborating, they can more conveniently, quickly, and reliably complete complex system tasks that a single AUV cannot independently accomplish. Furthermore, in large ocean environments, collaborative task allocation can significantly improve task completion efficiency and practicality. In large-scale mission scenarios, optimizing task allocation within an AUV system directly impacts system efficiency. Therefore, task allocation within multi-AUV systems has become a hot topic in AUV research.

[0003] Underwater robot task allocation—that is, how to assign tasks and rationally implement task planning for a group of underwater robots targeting multiple targets in an uncertain ocean environment—is a highly challenging problem for multi-underwater robot systems. This problem is a hot topic in multi-underwater robot system research, reflecting the self-organizing operational mechanism of underwater robot systems and fundamental to their goal achievement. On the one hand, with the increasing scale and complexity of tasks to be assigned in large-scale mission scenarios, existing self-organizing neural networks converge slowly and take a long time to converge. On the other hand, with limited resources, resource coordination among multiple underwater robot systems makes load balancing a difficult problem, requiring a rational distribution of task loads to maximize system performance. Therefore, it is crucial to design a multi-underwater robot task allocation method that is efficient, load-balanced, and suitable for large-scale mission scenarios. Summary of the Invention

[0004] In order to solve the above-mentioned defects, the present invention proposes a multi-underwater robot task allocation method and system based on a self-organizing neural network.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] The multi-underwater robot task allocation method based on self-organizing neural network includes:

[0007] Based on the multiple constraints of ocean current influence, resource consumption, load balancing and limited resources, a multi-constrained task allocation optimization model for multi-heterogeneous underwater robot systems is constructed. The task allocation optimization model is a self-organizing neural network.

[0008] The task allocation optimization model completes task allocation through competition and learning of neurons corresponding to the robot path points, and finally generates the circular path required by each robot.

[0009] A further improvement of this method is that: according to the multiple constraints of underwater current influence, resource consumption, load balancing and limited resources, a multi-constrained multi-heterogeneous underwater robot system task allocation optimization model is constructed, wherein the task allocation optimization model is a self-organizing neural network including:

[0010] When constructing the task allocation optimization model for a multi-underwater robot system, the influence of ocean currents is first considered, and the robot speed and ocean current speed are comprehensively considered;

[0011] According to the energy consumption of each robot, its path is divided into segments and converted into the calculation of its cruise time in the case of velocity vector sum to accurately measure its load. At the same time, its required cruise time must be less than the maximum cruise time.

[0012] With the goal of minimizing the resource consumption of the underwater robot system and maximizing the task reward acquisition of the robot system, a fitness function is constructed and a task allocation optimization model is established.

[0013] A further improvement of this method is that the task allocation optimization model completes task allocation through competition and learning of neurons corresponding to robot path points, and ultimately generates the circular path required by each robot, including:

[0014] At the beginning, each robot uses its base as the center point and generates several neurons to form a neuron ring. The neurons on the neuron ring act as output neurons and participate in the competition and update of the self-organizing neural network. Through iterative learning, the neurons that can access tasks are eventually retained as path points. These path points form a circular path. By moving along the path in sequence, all assigned tasks can be completed and finally return to the base.

[0015] A further improvement of this method is that the process and characteristics of the task allocation optimization model through competition and learning corresponding to the robot path points are as follows:

[0016] Parameter initialization: Setting the parameters of the self-organizing neural network, including the learning rate, neighborhood radius, iteration rounds, decay coefficient, and optimization function parameters; input and output neurons are determined according to the number of tasks and robots;

[0017] Task input: At the beginning of each round, the order of tasks is randomized and input into the self-organizing neural network in sequence.

[0018] Winning neuron selection and load balancing mechanism: The proposed speed synthesis method converts the travel path into a cruising time calculation, accurately measuring its load. Load balancing parameters are proposed based on the different working capabilities of robots and act on the proposed winning neuron selection mechanism. The speed synthesis method, which combines the robot speed and the ocean current speed, converts the travel path into a cruising time calculation, accurately measuring its load. The load balancing parameters of the robots are used to optimize the neuron selection mechanism.

[0019] Neuron update: Through the neuron addition mechanism and historical information guidance mechanism, an efficient self-organizing neural network is obtained. The process is as follows:

[0020] Neuron addition mechanism process: setting the winning parameter γ for each neuron ki , where γ ki is the number of times the neuron is the winning neuron, if the parameter γ ki If it is greater than or equal to the threshold, the neuron information is copied as a new neuron, and the original neuron is set to γ ki =0, the newly added neurons do not participate in the update of this task;

[0021] The neuron history information guidance process synthesizes the update vectors used in the history of neighboring neurons into a history information guidance vector, which is added to the current update as additional learning information. The specific calculation is as follows:

[0022]

[0023]

[0024]

[0025]

[0026] Among them, the task set Each task in selects neuron k in the past update i As the winning neuron, this task is performed with neuron k i The corresponding distance is parameter Depending on the order of the task in the task set, the weight parameter ψ is the overall weight of the historical information guidance vector;

[0027] During the neuron deletion mechanism, in each neuron ring, the cumulative error of each neighboring neuron is calculated. The cumulative error is used to measure the mapping ability of the neuron. The calculation method is to add up the difference between the update degree of each neighboring neuron and the update degree of the winning neuron each time it is updated as a neighboring neuron, and then divide it by the number of times it is a neighboring neuron. Then, the average cumulative error of this ring is calculated, and neurons with a larger average cumulative error of this ring are deleted. The specific formula is:

[0028]

[0029]

[0030]

[0031] Among them, the task set Each task in the i As a neighboring neuron, and Determine the tasks Update strength, this ratio reflects the update error, indicating the difference in update ability between the neighboring neurons and the winning neurons; the average of the neurons is The average error of the updating ability of all neurons in the ring is err i,ave .

[0032] The multi-underwater robot task allocation system based on self-organizing neural network includes:

[0033] A construction module is used to construct a multi-constrained multi-heterogeneous underwater robot system task allocation optimization model based on multiple constraints such as ocean current influence, resource consumption, load balancing, and resource limitations. The task allocation optimization model is a self-organizing neural network.

[0034] The task allocation module is used to complete the task allocation through the competition and learning of neurons corresponding to the robot path points in the task allocation optimization model, and generate the circular path required by each robot.

[0035] Due to the adoption of the above technical solutions, the technical progress achieved by the present invention is as follows: (1) considering the influence of ocean currents, the energy consumption of the robot is converted into the calculation of its cruising time, and constraints such as task completion rate, load balancing, and limited resources are proposed, and a complex optimization model is established. Compared with the common task allocation optimization model, the model established by the present invention considers more comprehensive factors, which is conducive to improving the working efficiency of the robot system; (2) when using the self-organizing neural network algorithm to solve the task allocation problem, a neuron self-layout mechanism is proposed, which includes neuron deletion, neuron addition and historical information guidance mechanism to comprehensively optimize the quality of the neuron population, and at the same time improves the original update method to achieve a balanced relationship between development and utilization in different learning stages, thereby improving the solution efficiency; (3) in a multi-robot system, the different energy consumption of robots is fully considered, and their energy consumption is accurately measured under the influence of ocean currents. At the same time, a neuron winning mechanism is proposed to achieve load balancing, which can give full play to robots with different working capabilities and improve their continuous working time. Therefore, the task allocation method based on the self-organizing neural network algorithm proposed in the present invention has the advantages of fast solution speed and high solution accuracy, and can provide a good task allocation solution within a limited number of iterations. Especially in complex large-scale task scenarios, the advantages of the method proposed in the present invention are more obvious. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A diagram showing the structure of a self-organizing neural network according to the present invention;

[0037] Figure 2 is a flow chart of the method of the present invention;

[0038] Figure 3 This is a flow chart of the self-organizing neural network algorithm proposed in the present invention;

[0039] Figure 4 This is a comparison diagram of the utility function process of the method of the present invention, the traditional self-organizing neural network SOM, the smoothed self-organizing neural network SSOM and the improved self-organizing neural network ISOM in a small-scale task scenario;

[0040] Figure 5 The utility function process comparison diagram of the method of the present invention with the traditional self-organizing neural network SOM, smoothed self-organizing neural network SSOM and improved self-organizing neural network ISOM in a medium-scale task scenario;

[0041] Figure 6 The utility function process comparison diagram of the method of the present invention with the traditional self-organizing neural network SOM, smoothed self-organizing neural network SSOM and improved self-organizing neural network ISOM in large-scale task scenarios;

[0042] Figure 7This is a comparison chart of the change in the number of neurons in the method of the present invention, the traditional self-organizing neural network SOM, the smoothed self-organizing neural network SSOM, and the improved self-organizing neural network ISOM in a small-scale task scenario;

[0043] Figure 8 This is a comparison chart of the change in the number of neurons in the method of the present invention, the traditional self-organizing neural network SOM, the smoothed self-organizing neural network SSOM, and the improved self-organizing neural network ISOM in a medium-scale task scenario;

[0044] Figure 9 This is a comparison chart of the change process of the number of neurons in the method of the present invention, the traditional self-organizing neural network SOM, the smoothed self-organizing neural network SSOM and the improved self-organizing neural network ISOM in large-scale task scenarios. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below with reference to the accompanying drawings:

[0046] like Figure 2 As shown, the present invention mainly includes the following steps:

[0047] ⑴Establish a multi-heterogeneous robot task allocation system model

[0048] A task allocation scenario is constructed, where the robot set I = {1, 2, ..., i, ..., I} and the task (sensor) set M = {1, 2, ..., m, ..., M}, where i and m represent the i-th robot and the m-th task, respectively, and I and M represent the number of robots and tasks, respectively. Each task needs to be executed by a robot, that is, the robot needs to obtain data from each sensor. All robots perform tasks together and return to their home base after completing the assigned tasks. To verify the wide applicability of the method proposed in this invention, we set up three task scenarios of different scales: small-scale, medium-scale, and large-scale. In the task scenarios, tasks are randomly distributed, and robots have fixed bases. The scale of each scenario, the number of tasks, the number of robots, the base, and the corresponding maximum cruising time are shown in Table 3.

[0049] Table 3

[0050]

[0051] ⑵Construct a multi-robot system task allocation optimization model with multiple constraints

[0052] The present invention takes into account the limited resources, load balancing and task access requirements of a multi-robot system under the influence of ocean currents, and constructs an optimization model that comprehensively considers the marine environment, robot system resource consumption and load balancing.

[0053] ① Convert the multi-robot task allocation problem into an optimization model with multiple constraints. In underwater operating environments, the influence of ocean currents is unavoidable, so the robot's energy consumption depends on both the length of its travel distance and its heading. The method proposed in this paper synthesizes the robot's speed and the speed of the ocean current by speed, and calculates the time spent on the travel distance as the robot's energy consumption. The proposed energy consumption formula is defined as:

[0054]

[0055]

[0056] in is the velocity vector of the robot, is the velocity vector of the ocean current, is the velocity vector sum, pointing to the current target direction, is the current distance, so The time consumption corresponding to the current path.

[0057] For robot i, its path during the task can be divided into different waypoints, so these waypoints can form a circular path. At the same time, the adjacent waypoints maintain the same heading, and its speed is the sum of the above velocity vectors. Then the energy consumption of the robot when performing the assigned task can be calculated as:

[0058]

[0059] where K' i ={1' i ,2' i ,...,k' i ,...K' i}, i∈Ι are all the path points of robot i, and path point 1' i With the path point K' i Both are the base points of the robot, ensuring that it returns to the base after completing the mission.

[0060] For robot i, since its consumable energy is limited, its energy consumption cannot be greater than its maximum working capacity in task allocation, which can be expressed as:

[0061]

[0062] Its maximum cruising time.

[0063] For a task m∈Μ, it has a reward value of θ m , the communication radius is r m, if and only if the distance between the robot and the task is less than the communication radius, the data can be successfully transmitted and the corresponding task reward value can be obtained. Therefore, for the robot system, after completing the task, its optimization model is established as:

[0064]

[0065] Where α and β are coefficients, which can be adjusted according to actual conditions, as defined in formula (4), F i is the task reward obtained by robot i and C i is the cost consumed by robot i. Based on the above constraints and objective function, an optimization model for multi-robot task allocation is established. In this model, by minimizing the energy consumption function and maximizing the task reward, the robot system is guaranteed to obtain the maximum task reward while minimizing resource consumption, while also complying with the energy consumption constraint.

[0066] ⑶ Solve the constructed optimization model using the algorithm based on the enhanced self-organizing neural network. The simple flow chart of the algorithm based on the enhanced self-organizing neural network proposed in this invention is shown in the attached figure. Figure 3 shown.

[0067] ①Neural network parameter initialization

[0068] Parameter initialization primarily involves two steps: first, inputting and setting the neural network's own parameters, primarily including the learning rate, neighborhood radius, number of iterations, decay coefficient, and optimization function parameters. Second, the input and output neurons are determined based on the number of tasks and robots. In this paper, each task is treated as a separate input neuron, and each robot corresponds to a ring of output neurons. The total number of output neurons equals the number of tasks, and the initial number of output neurons is evenly divided according to the total number of robots. The specific values of the various parameters used in this paper are shown in Table 4.

[0069] Table 4

[0070] parameter Learning rate μ Neighborhood radius σ Neighborhood attenuation coefficient a Maximum iteration round t α β Numerical 0.6 0.1M 0.02 50 3 -1

[0071] The initial learning rate is 0.6, and the learning rate is The iteration round is one round in which all tasks are input once, and the neighborhood radius change relationship is σ = σ(1-ta).

[0072] ②Task input

[0073] For a given task scenario, at the beginning of each round, the order of tasks is randomized to avoid falling into local extremes. The first task in the current order is entered, and after the update, the entered task numbers are deleted. According to the above definition, the round ends when all tasks in the round have been entered once.

[0074] ③Winning neuron selection and load balancing mechanism

[0075] In a multi-robot system, robots with varying capabilities and base locations can prematurely exhaust their energy, while others struggle to obtain tasks. This leads to an unbalanced load across the system, which can also prolong task completion time. Therefore, considering the influence of ocean currents, we propose a method for load balancing robots with heterogeneous capabilities.

[0076] Measuring real capacity consumption

[0077]

[0078]

[0079] As shown above, by taking into account the ocean current speed and the robot's heading speed, the calculation of the robot's consumed work capacity at the corresponding distance is more comprehensive and accurate than the distance calculation in other studies.

[0080] Load balancing parameters

[0081] Due to the different working capabilities of robots in a robot system, the key to load balancing is that robots with higher working capabilities perform more tasks, while robots with lower working capabilities perform fewer tasks. In other words, the load is proportional to their working capabilities. Therefore, the load balancing parameters are proposed as follows:

[0082]

[0083] Winning neuron selection mechanism

[0084]

[0085]

[0086]

[0087] Among them, the distance The output neuron k on robot i is represented by i The distance from the current task, and The direction of the current position points to the target position. Therefore, in the process of selecting the winning neuron, not only the distance between the neuron and the task is considered, but also the working ability of the robot. Under the influence of the load balancing parameters, the difference in the working ability of the robots is resolved, thus achieving true task allocation based on working ability.

[0088] ④Neural renewal

[0089] During the update process, a neuron addition mechanism and a historical information guidance mechanism are proposed. The neuron addition mechanism is designed to achieve rapid updates of the output neuron population, thereby introducing high-quality neurons to accelerate the iteration and convergence of the algorithm. The historical information guidance mechanism adds additional update information to the updated neighboring neurons, thereby achieving a balance between algorithm development and utilization during the update process. The details are as follows:

[0090] Neuron increase mechanism

[0091] The self-organizing neural network algorithm is to input neurons and continuously adjust the output neurons to achieve accurate mapping of the input neurons. According to the winning neuron selection mechanism mentioned above, the winning neuron indicates that it has the highest mapping accuracy in the current neuron population. However, the update mechanism updates it to a mapping point with higher value. Although the current mapping position is still valuable, it is abandoned. Therefore, a neuron addition mechanism is proposed to optimize the neuron population, which is implemented as follows:

[0092] Assign parameters to each winning neuron Record the number of times it is the winning neuron, up to Then copy the neuron information as a new neuron, and set the original neuron The newly added neurons do not participate in the update of this task.

[0093] Selection and determination of neighboring neurons

[0094]

[0095]

[0096] distance is the distance between the winning neuron and the neurons in the same neuron ring. According to the formula, when the distance is less than the neighborhood radius, it is selected as a neighboring neuron. When it is greater than the distance, it does not participate in this update. is the update strength, so the farther the neighboring neurons are from the Vulcan neuron, the lower their update strength.

[0097] Historical information guidance mechanism

[0098] For neighboring neurons, since their update intensity is not as strong as that of the winning neurons, the overall purpose of the algorithm is to enhance the diffusion of the neuron population and explore better mapping positions. Therefore, the update of the winning neurons is for utilization, while the main purpose of the update of the neighboring neurons is for development. However, for neighboring neurons, they may have been updated as winning neurons in the previous update, but this time the update as a neighboring neuron ignores the previous update, which may cause confusion in their update and affect their original update for development. Based on this, the present invention proposes a historical information guidance mechanism, which synthesizes the update vectors of the neighboring neurons that have been used in the past into a historical information guidance vector, and adds it to this update as additional learning information. The specific calculation is as follows:

[0099]

[0100]

[0101]

[0102] The task set Each task in selects neuron k in the past update i As the winning neuron, the task is similar to that of neuron k i The corresponding distance is parameter It depends on the order of the task in the task set. All tasks are sorted according to the time sequence. Therefore, the more recent the task, the greater its weight coefficient, which is in line with the current update trend. The weight parameter ψ is the overall weight of the historical information guidance vector. It can be seen that with the increase in the number of iterations, the weight parameter gradually increases. This trend is in line with the characteristics of the algorithm that focuses on development in the early stage and gradually transitions to focusing on utilization.

[0103] Neuronal renewal

[0104] Based on the historical information guidance mechanism proposed above, the update formulas for the winning neuron and the neighboring neurons are as follows:

[0105]

[0106] After updating based on the proposed update formula, the response to the input of the current task is completed, and then the next task is input according to the task sequence, and the iterative process is repeated.

[0107] ⑤ Neuron deletion mechanism

[0108] After all tasks in the task set are input once, this round of iteration is completed. According to the role of the output neurons in the original update, all output neurons can be divided into three categories: static neurons, neurons that only serve as neighborhood neurons, and winning neurons. In order to improve the iterative efficiency and learning ability of the algorithm, the neuron population needs to be optimized. The proposed neuron deletion mechanism is to first delete all static neurons. These neurons do not participate in the update, indicating that they are worthless. All winning neurons are retained due to their high value. The deletion criteria for neighborhood neurons are as follows: in each neuron ring, the cumulative error of each neighborhood neuron is calculated. The cumulative error is used to measure the mapping ability of the neuron. The calculation method is that for each time it serves as a neighborhood neuron, the difference between it and the winning neuron during the update is accumulated and then divided by the number of times it serves as a neighborhood neuron. Then the average cumulative error of this ring is calculated, and neurons greater than the average cumulative error are deleted. The specific formula is:

[0109]

[0110]

[0111]

[0112] Task Collection Each task in the i As a neighboring neuron, and Determine the tasks Update strength, this ratio is the update error, which shows the difference in update ability between the neighboring neurons and the winning neurons. The average error of the updating ability of all neurons in the ring is err i,ave , and then determine the neurons that need to be deleted according to the formula.

[0113] Under the proposed neuron deletion mechanism, the neuron mapping ability is quantitatively measured, thus ensuring that only low-value neurons are deleted, while neurons with higher value in the neighborhood are preserved. This neither leads to a large-scale reduction in the neuron population nor optimizes the quality of the neuron population in a reasonable and efficient manner, thereby improving the efficiency of the algorithm while reducing the computational burden brought by redundant neurons.

[0114] ⑥Simulation

[0115] In this paper, we provide a task allocation method based on an enhanced self-organizing neural network, where the communication radius of the task is any value in {1, 2, 3} and the task reward is 30≤θ m≤80, and the remaining simulation parameters are as shown above. To evaluate the algorithm's convergence speed and load balancing effectiveness, we conducted a large number of Monte Carlo simulation experiments. Table 1 compares the simulation results of the proposed method with those of a traditional self-organizing neural network (SOM), a smoothed self-organizing neural network (SSOM), and an improved self-organizing neural network (ISOM) for three scale task scenarios. Figure 4 The utility function process comparison diagram of the method of the present invention and SOM, SSOM and ISOM in three scale task scenarios. Figure 5 Three types of regulations

[0116] Comparison of the changes in the number of neurons between the method of the present invention and SOM, SSOM and ISOM in the simulation task scenario.

[0117] Table 1 Simulation result data comparison table

[0118]

[0119] Table 2 is a comparison table of load balancing among multi-robot systems in three scale task scenarios according to the method of the present invention.

[0120] Table 2 Load balancing comparison table

[0121]

[0122] It can be seen from the simulation results that the task allocation method proposed in this invention has higher performance than the commonly used task allocation methods (SOM, SSOM and ISOM) in terms of convergence speed, simulation time and task allocation rate. At the same time, the present invention can achieve good load balancing between robots with different working capabilities.

[0123] The specific implementation process of the improved enhanced self-organizing neural network is summarized as follows:

[0124] (1) Initialize the algorithm parameters, input the task and robot information, and build a ring composed of output neurons with the base of each robot as the center.

[0125] (2) The order of tasks is random, and the tasks are entered in sequence from the random order.

[0126] (3) The output neurons of different robots compete with each other, and the best one is selected as the winning neuron according to the proposed winning rule. The proposed winning rule can achieve load balancing regardless of whether the cruising time between robot systems is the same or different.

[0127] (4) For the winning neuron, its winning times are measured. If it meets the proposed replication standard, that is, the number of times it becomes a winning neuron is the same as the proposed replication standard, the neuron is replicated. However, in this update, the new neuron obtained by the replication is not updated, and the winning times of the original winning neuron are reset to zero.

[0128] (5) According to the winning neuron, the neuron ring to which the corresponding robot belongs is determined, and its neighboring neurons are determined through the domain formula. For the neighboring neurons, a historical information guidance mechanism is proposed to guide their update.

[0129] (6) Update the winning neuron and its neighboring neurons according to the proposed neighborhood update formula.

[0130] (7) After the update is completed, the update of the task is completed, and the next task is input in sequence until all tasks are input once, and the current input era ends.

[0131] (8) Selectively delete the neurons on the neuron ring corresponding to each robot. The specific deletion criteria are as follows: first, delete the static neurons in this round. For the neighboring neurons that have participated in the update, calculate their cumulative error and the average cumulative error of all neighboring neurons on the ring. Delete the neurons with a cumulative error greater than the average cumulative error.

[0132] (9) Repeat the input task to continue updating the iteration until the iteration termination condition is met.

[0133] In summary, the present invention is a task allocation method based on an enhanced self-organizing neural network. Aiming at underwater environmental constraints, robot body design constraints, and large-scale task scenario limitations, an optimization model with multiple constraints such as load balancing and limited resources is constructed. Aiming at the problems of insufficient solution accuracy and slow solution speed of existing optimization algorithms when solving complex optimization problems, an improved self-organizing neural network is designed. By rationally optimizing the population quality and enriching the learning information of neurons, the efficiency of the algorithm is improved, so that it can quickly and accurately solve both small-scale and large-scale task scenarios, meeting the system task requirements. At the same time, in order to address the influence of ocean currents, a speed synthesis method is proposed to convert the path into cruising time to reasonably measure the load of different robots, while achieving good load balancing.

[0134] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A multi-underwater robot task allocation method based on a self-organizing neural network, characterized in that: include: Based on the multiple constraints of ocean current influence, resource consumption, load balancing and limited resources, a multi-constrained task allocation optimization model for multi-heterogeneous underwater robot systems is constructed. The task allocation optimization model is a self-organizing neural network. The task allocation optimization model completes task allocation through competition and learning of neurons corresponding to the robot path points, and ultimately generates the circular path required by each robot; The process and characteristics of the task allocation optimization model through competition and learning corresponding to the robot path points are as follows: Parameter initialization: Setting the parameters of the self-organizing neural network, including the learning rate, neighborhood radius, iteration rounds, decay coefficient, and optimization function parameters; input and output neurons are determined according to the number of tasks and robots; Task input: At the beginning of each round, the order of tasks is randomized and input into the self-organizing neural network in sequence. Winning neuron selection and load balancing mechanism: The proposed speed synthesis method converts the travel path into a cruising time calculation, accurately measuring its load. Load balancing parameters are proposed based on the different working capabilities of robots and act on the proposed winning neuron selection mechanism. The speed synthesis method, which combines the robot speed and the ocean current speed, converts the travel path into a cruising time calculation, accurately measuring its load. The load balancing parameters of the robots are used to optimize the neuron selection mechanism. Neuron update: Through the neuron addition mechanism and historical information guidance mechanism, an efficient self-organizing neural network is obtained. The process is as follows: Neuron addition mechanism process: setting the winning parameters for each neuron in, is the number of times the neuron is the winning neuron, if the parameter If the value is greater than or equal to the threshold, the neuron information is copied as a new neuron, and the original neuron is set Newly added neurons do not participate in the update of this task; The neuron history information guidance process synthesizes the update vectors used in the history of neighboring neurons into a history information guidance vector, which is added to the current update as additional learning information. The specific calculation is as follows: Among them, the task set Each task in selects neuron k in the past update i As the winning neuron, this task is performed with neuron k i The corresponding distance is parameter Depending on the order of the task in the task set, the weight parameter ψ is the overall weight of the historical information guidance vector; During the neuron deletion mechanism, in each neuron ring, the cumulative error of each neighboring neuron is calculated. The cumulative error is used to measure the mapping ability of the neuron. The calculation method is to add up the difference between the update degree of each neighboring neuron and the update degree of the winning neuron each time it is updated as a neighboring neuron, and then divide it by the number of times it is a neighboring neuron. Then, the average cumulative error of this ring is calculated, and neurons with a larger average cumulative error of this ring are deleted. The specific formula is: Among them, the task set Each task in the i As a neighboring neuron, and Determine the tasks Update intensity, and The ratio of reflects the update error, indicating the difference in updating ability between the neighboring neurons and the winning neurons; the average of the neurons is It fully reflects its updating ability; the average error of the updating ability of all neurons in the ring is err i,ave .

2. The multi-underwater robot task allocation method based on self-organizing neural network according to claim 1 is characterized in that: According to the multiple constraints of ocean current influence, resource consumption, load balancing and limited resources, a multi-constrained multi-heterogeneous underwater robot system task allocation optimization model is constructed, wherein the task allocation optimization model is a self-organizing neural network including: When constructing the task allocation optimization model for a multi-underwater robot system, the influence of ocean currents is first considered, and the robot speed and ocean current speed are comprehensively considered; According to the energy consumption of each robot, its path is divided into segments and converted into the calculation of its cruise time in the case of velocity vector sum to accurately measure its load. At the same time, its required cruise time must be less than the maximum cruise time. With the goal of minimizing the resource consumption of the underwater robot system and maximizing the task reward acquisition of the robot system, a fitness function is constructed and a task allocation optimization model is established.

3. The multi-underwater robot task allocation method based on self-organizing neural network according to claim 1 is characterized in that: The task allocation optimization model completes task allocation through competition and learning of neurons corresponding to robot path points, and ultimately generates the circular path required by each robot, including: At the beginning, each robot uses its base as the center point and generates several neurons to form a neuron ring. The neurons on the neuron ring act as output neurons and participate in the competition and update of the self-organizing neural network. Through iterative learning, the neurons that can access tasks are eventually retained as path points. These path points form a circular path. By moving along the path in sequence, all assigned tasks can be completed and finally return to the base.

4. A multi-underwater robot task allocation system based on a self-organizing neural network based on the allocation method according to any one of claims 1 to 3, characterized in that: include: A construction module is used to construct a multi-constrained multi-heterogeneous underwater robot system task allocation optimization model based on multiple constraints such as ocean current influence, resource consumption, load balancing, and resource limitations. The task allocation optimization model is a self-organizing neural network. The task allocation module is used to complete the task allocation through the competition and learning of neurons corresponding to the robot path points in the task allocation optimization model, and generate the circular path required by each robot.

Images