CN108985580B - Multi-robot disaster search and rescue task allocation method based on improved BP neural network - Google Patents

Abstract

The invention discloses a multi-robot disaster search and rescue task assignment method based on an improved BP neural network. When the auctioneer does not have enough ability to complete the current search and rescue mission, the information of the mission will be broadcast to the bidders; the bidders will respond according to their current status; if the bidders are free or can complete the search and rescue mission, they will send The auctioneer responds to receive this task; otherwise, the bidder will not participate in the auction; if the auctioneer does not receive bids during the entire auction process, the auction process ends; otherwise, the auctioneer will pay for these bid prices After normalization, the improved BP neural network is used to train the processed bid prices. The invention can effectively make the critically injured get preferential treatment, and the method has the robustness to the robot failure. Computer simulation verifies the training effect of the improved BP neural network.

Description

Multi-robot disaster search and rescue task assignment method based on improved BP neural network

technical field

The invention belongs to the technical field of disaster rescue, and relates to a multi-robot disaster search and rescue task assignment method based on an improved BP (Back Propagation) neural network.

Background technique

The present invention takes a multi-robot system (multi-robot team) composed of multiple autonomous wheeled mobile robots with search and rescue functions as the research object. Carry out life search and rescue. In this process, the assignment of tasks between robots is the key to ensuring that the critically injured are given priority treatment and the rescue is completed efficiently. At present, the methods widely used in multi-robot task assignment mainly include: market auction method and its improvement, behavior-based method, linear programming-based method, etc. The prior art proposes an improved auction-based multi-agent task assignment algorithm, which enables multi-robots to coordinate with each other and complete tasks in the shortest time in a complex and unmanned dynamic environment. There are also existing technologies that combine the multi-robot coordination method of exchange tree on the basis of the contract network agreement, and propose a multi-robot dynamic task allocation algorithm based on Pareto improvement, which improves the efficiency of task allocation and reduces the time required for completing tasks. time. The behavior-based method is to find the robot-task pair with the greatest utility, and then assign the task to the corresponding robot. This method has strong real-time and fault tolerance; In various practical applications based on auction methods, it is difficult to find a fusion and decision-making method when each robot gives multiple bid prices for a task without mentioning the side (distance, speed, energy consumption, etc.) to determine the bid winner.

On the other hand, neural networks can be used for information fusion. The BP neural network algorithm based on gradient descent has a solid theoretical basis and strong generality, but it has the shortcomings of easy to fall into local minima, slow learning convergence process and difficult to determine the network structure. Therefore, the present invention applies the improved BP neural network to realize the fusion of bidding information of the robot, and designs a task assignment method to treat the critically wounded with priority based on the characteristics of the task and the robot in the disaster environment.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the defects existing in the prior art, and propose a multi-robot disaster search and rescue task assignment method based on an improved BP neural network. When there is a new task to be completed, the auctioneer first publishes the task information, fuses the bidding information of each robot through the neural network, and then uses the contract network protocol auction algorithm to realize the task allocation of multiple robots. The robot that finds a task or initiates an auction during the auction acts as both an auctioneer and a bidder, and other robots can only be bidders.

Its technical solutions are as follows:

The multi-robot disaster search and rescue task assignment method based on improved BP neural network includes the following steps:

1) Each robot of the multi-robot system (or multi-robot team) starts to search the environment according to the principle of depth first.

2) When a robot finds a new rescue mission in a given direction, auction the mission. This auction bot is called the auctioneer.

3) The auctioneer will broadcast the task information that needs to be completed to all robots in the team. The task information includes the location of the wounded, the deadline for task bidding, the number of robots n _c required for the task, and the priority of the task. level etc.

4) Each robot in the team is a potential bidder. After receiving the search and rescue task information released by the auctioneer, they respond according to their current status:

①If the current robot is in an idle state, participate in the bidding;

②If the current robot is performing a task, but the priority of the task being performed is lower than the task of the auction, it will participate in the bidding;

③ If the current robot is performing a task, but the priority of the task being executed is higher than the task of the auction, it will not participate in the bidding.

5) If an auction conflict occurs, the auction conflict resolution strategy is invoked.

6) If the number of bidding robots received by the auctioneer before the auction deadline is less than the number of robots required to complete the task, the auction process ends, and the task is stored in their respective unfinished task lists by all robots in the team.

7) Otherwise, the auctioneer will normalize these bid prices, and then use the trained improved BP neural network to comprehensively sort the multiple bid prices given by the robots that intend to participate in the task, and select n _c winners By.

8) The auctioneer informs the n _c winners that the task will be performed, and the bidder sends the contract confirmation message to the auctioneer after receiving the notification. Otherwise, the auctioneer will think that the bidder gave up the task due to failures and other reasons, and then select other robots to participate in the task in turn according to the evaluation size of the bid value from other robots in the bid.

9) If the auctioneer receives more than the minimum number of robots required to complete the task, the contract is established and the rescue mission begins.

10) If the number of contract confirmation information received by the auctioneer is less than the minimum number of robots required to complete the task, the contract cannot be established; all robots store the task in the unfinished task list, and the idle robots continue to search in the predetermined direction. The robot continues to perform the original rescue mission.

11) If the contract is established and a robot that accepts a task is executing a task with a lower priority, the lower priority task will be suspended and the lower priority task will be stored in the unfinished task list.

12) When a robot completes the rescue task it is undertaking, if the uncompleted task list is empty, go to step 1); otherwise, select the task with the highest priority in the uncompleted task list for auction, go to step 3) .

Further, in step 5, the auction conflict is that only robot 1 receives the information of task 2 released by robot 2 during the auction of task 1. At this time, two auctioneers compete for an idle robot in the system. In order to To avoid deadlocks, the solution adopted by the present invention is that two auctioneers compare the priorities of two tasks, the tasks with higher priorities continue to be assigned, the tasks with lower priorities are stored in the list of unfinished tasks, and the corresponding auctions are higher Robots with priority tasks act as auctioneers in the subsequent allocation process.

In step 7, the improved BP neural network refers to the BP neural network based on the Levenberg-Marquardt (LM) algorithm, which not only has the local convergence of the Gauss-Newton method, but also has the global convergence of the gradient method. sex. Let x ^(k) be the network weight vector of the k-th iteration, and the k+1-th weight vector x ^(k+1) can be obtained by the following formula:

x ^(k+1) = x ^(k) + Δx ^(k+1) (1)

In the Gauss-Newton method calculation rules

Δx ^(k+1) = -[J ^T (x)J(x)] ^-1 J(x)e(x) (2)

where e(x)=[e ₁ (x) … e _N (x)] ^T is the output error vector, where x represents x ^k , the superscript is omitted without causing confusion, the same below; N is the neural The dimension of the network output, J(x) is the Jacobian matrix of the error to the weight, and its calculation formula is

In the weight update value formula (2), J(x) ^T J(x) is irreversible, so the algorithm may not converge. To solve this problem, the following form of weight update is used

Δx ^(k+1) = -[J ^T (x)J(x)+λI] ^-1 J(x)e(x) (4)

where λ is a constant, and I is an n×n unit matrix.

The beneficial effects of the present invention are:

The present invention is a multi-robot task assignment method based on the combination of contract network protocol and improved BP algorithm. Compared with the traditional task assignment method, it has the following advantages: First, the method considers the involvement of robots when completing tasks. multiple parameters, such as distance, speed, energy consumption, task priority, etc.; secondly, the method integrates and sorts the bidding values of each robot through neural network, so that the tasks that can be completed by a single robot can be assigned, and the tasks that can be completed by a single robot can also be assigned to multiple robots. The tasks that can only be completed by each robot are allocated; thirdly, the comprehensive sorting also ensures the robustness of the task allocation method proposed by the present invention to the robot failure; finally, the method realizes the rescue process of each robot by setting the priority of each task. The dynamic adjustment of the committed tasks in China ensures that those with serious injuries can be given priority treatment. The traditional BP neural network has the disadvantage of slow convergence speed. The simulation results show that the training process of the improved BP neural network is significantly accelerated. The invention can effectively deal with the fusion problem of multiple bidding parameters, and the proposed rescue task allocation strategy can ensure that the critically injured can be treated in the shortest time.

Description of drawings

Figure 1 is a flowchart of the task allocation and rescue operations of the auctioneer robot in the process of disaster search and rescue;

Figure 2 is the auction conflict resolution strategy;

Figure 3 is a flowchart of the task response and rescue operations of the bidder robot in the process of disaster search and rescue;

Figure 4 is the traditional BP neural network training process;

Figure 5 shows the training process of the BP neural network based on the Levenberg-Marquardt algorithm.

Detailed ways

The technical solutions of the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

1 Rescue mission bidding and bidding information fusion

1.1BP neural network

BP neural network, namely back-propagation network, is a feed-forward network using error back-propagation training algorithm, and it is one of the most widely used neural network models at present. The bidding information B of multiple robots in the team together constitute the input of the BP neural network, B = [B ₁ B ₂ ... B _M ], B _i (i = 1, ..., M) is the bidding price of robot i, take B _i =[d _i s _i e _i ], where d _i is the shortest distance from robot i to the target point, s _i is the fastest moving speed of robot i, e _i is the current remaining power of robot i, and M is the robot participating in the bidding number. The invention adopts a three-layer BP neural network, with 15 neurons in the input layer, 10 neurons in the hidden layer, and 5 neurons in the output layer; that is, for a task, only the five robots with the fastest response are accepted for bidding. In order to reduce the computational complexity, the hidden layer adopts a unipolar Log-Sigmoid function, and the output layer adopts a linear function.

1.2 Normalization

In order to avoid the influence of different input data amplitudes and different physical meanings on the output of the BP network, the input data is normalized. When the input data needs to be transformed to [0, 1], the following formula is used:

When the input data needs to be transformed to [-1, 1], the following formula is used:

_Among them, _{x i} (i=1, _. is the input value of the neural network.

1.3Levenberg-Marquardt BP algorithm

The improved BP neural network refers to the BP neural network based on the Levenberg-Marquardt (LM for short) algorithm, which not only has the local convergence of the Gauss-Newton method, but also the global convergence of the gradient method. Let x ^(k) be the network weight vector of the k-th iteration, and the k+1-th weight vector x ^(k+1) can be obtained by the following formula:

x ^(k+1) = x ^(k) + Δx ^(k+1) (7)

In the Gauss-Newton method calculation rules

Δx ^(k+1 )=-[J ^T (x)J(x)] ^-1 J(x)e(x) (8)

where e(x)=[e ₁ (x) … e _N (x)] ^T is the output error vector, where x represents x ^k , the superscript is omitted without causing confusion, the same below; N is the neural The dimension of the network output, J(x) is the Jacobian matrix of the error to the weight, and its calculation formula is

In the weight update value formula (2), J(x) ^T J(x) is irreversible, so the algorithm may not converge. To solve this problem, the following form of weight update is used

Δx ^(k+1) = -[J ^T (x)J(x)+λI] ^-1 J(x)e(x) (10)

where λ is a constant, and I is an n×n unit matrix.

Neural networks need to be trained on artificially constructed data before being used for bidding information fusion.

2 The task allocation method of search and rescue

The contract network protocol is a type of auction algorithm. The three main auction systems are First-Price Sealed Auction, Vickrey Auction, and Dutch Auction. The traditional contract network protocol adopts the First-Price sealed auction system. When a bidder submits a sealed bid to the auctioneer, the bidder's bid price to other bidders is unknown at this time. The invention adopts the contract network protocol auction algorithm to realize the task point distribution of the multi-robots. The robot that finds the task during the auction or the robot that initiates the auction is the auctioneer, and the robot that may complete the task is the bidder. But only one robot can act as a single auctioneer, and it can also be a bidder. The task allocation flow chart is shown in Figure 1.

A multi-robot disaster rescue task assignment method based on improved BP neural network is proposed. Firstly, the bidding information of each robot is fused through neural network, and then the contract network protocol auction algorithm is used to realize the assignment of multi-robot task points.

Its technical solutions are as follows:

A multi-robot disaster rescue task assignment method based on an improved BP neural network, comprising the following steps:

1) Each robot of the multi-robot system (or multi-robot team) starts to search the environment according to the principle of depth first.

2) When a robot finds a new rescue mission in a given direction, auction the mission. This auction bot is called the auctioneer.

3) The auctioneer will broadcast the task information that needs to be completed to all robots in the team. The task information includes the location of the wounded, the deadline for task bidding, the number of robots n _c required for the task, and the priority of the task. level etc.

4) Each robot in the team is a potential bidder. After receiving the search and rescue task information released by the auctioneer, they respond according to their current status:

①If the current robot is in an idle state, participate in the bidding;

②If the current robot is performing a task, but the priority of the task being performed is lower than the task of the auction, it will participate in the bidding;

③ If the current robot is performing a task, but the priority of the task being executed is higher than the task of the auction, it will not participate in the bidding.

5) If an auction conflict occurs, the auction conflict resolution strategy is invoked.

6) If the number of bidding robots received by the auctioneer before the auction deadline is less than the number of robots required to complete the task, the auction process ends, and the task is stored in their respective unfinished task lists by all robots in the team.

7) Otherwise, the auctioneer will normalize these bid prices, and then use the trained improved BP neural network to comprehensively sort the multiple bid prices given by the robots that intend to participate in the task, and select n _c winners By.

8) The auctioneer informs the n _c winners that the task will be performed, and the bidder sends the contract confirmation message to the auctioneer after receiving the notification. Otherwise, the auctioneer will think that the bidder gave up the task due to failures and other reasons, and then select other robots to participate in the task in turn according to the evaluation size of the bid value from other robots in the bid.

9) If the auctioneer receives more than the minimum number of robots required to complete the task, the contract is established and the rescue mission begins.

10) If the number of contract confirmation information received by the auctioneer is less than the minimum number of robots required to complete the task, the contract cannot be established; all robots store the task in the unfinished task list, and the idle robots continue to search in the predetermined direction. The robot continues to perform the original rescue mission.

11) If the contract is established and a robot that accepts a task is executing a task with a lower priority, the lower priority task will be suspended and the lower priority task will be stored in the unfinished task list.

12) When a robot completes the rescue task it is undertaking, if the uncompleted task list is empty, go to step 1); otherwise, select the task with the highest priority in the uncompleted task list for auction, go to step 3) .

The auctioneer's work process and conflict resolution strategy are shown in Figure 1 and Figure 2, and the bidder's response and work process are shown in Figure 3.

3 Methods Comparative Analysis

Table 1 Comparison of search and rescue task allocation methods

The present invention proposes a multi-robot disaster search and rescue task allocation method based on an improved BP neural network. Compared with the existing disaster rescue task allocation method, the task allocation strategy designed by the present invention has obvious advantages in the following aspects:

(1) The most important goal of operation in disaster search and rescue is to save the life and safety of the injured. Because each injured has a different degree of injury, the critically injured must be given priority treatment. The present invention sets the task priority method, Ensure that this objective is achieved as a priority;

(2) The situation at the disaster site varies widely, some rescue tasks can be completed by one robot, and some rescue tasks can be completed only by the collaborative operation of multiple robots. The prior art ignores this feature, but the present invention can effectively handle this situation;

(3) The environmental conditions of the rescue site are harsh, and the failure of the search and rescue robot may occur at any time. The task allocation strategy of the present invention can be the most suitable robot to replace the faulty robot, and the robustness of the designed method to the failure of the robot is realized;

(4) The traditional method starts the bidding based on the difference between the benefit and the cost, but the conversion coefficient between the benefit and the cost is very darkly determined. Remaining energy and other factors, so the present invention establishes a multi-parameter bidding method.

(5) Distributed processing of tasks is a feature of all rescue tasks, and the traditional method takes distance as the cost. In order to treat the critically injured as soon as possible, the present invention also takes the maximum speed of the robot into consideration when the robot bids on the task;

(6) The process of discovering tasks by the robots in the team has asynchrony and concurrency. The present invention deals with asynchrony by default. When multiple robots discover different tasks at the same time, the present invention adopts the method of priority comparison to deal with critical injuries in a limited manner. allocation of rescue tasks.

(7) There is a solution to the auction conflict in the prior art, but it is only a simple time delay method. The present invention solves the conflict based on the priority of the characters, and has a very obvious meaning to life rescue.

The following is a simulation comparison between the traditional BP neural network and the BP neural network based on the L-M algorithm. Simulation equipment: laptop, CPU: i5-4210U, memory: 4G, system Windows8.1 Professional Edition, simulation software is Matlab 2015b. The input data is the normalized bidding value of each robot, in which each output neuron corresponds to the evaluation of the benefit of the robot with the same label to complete the task. The larger the value, the more advantageous the corresponding robot is to complete the task. Use the newff function to create a BP neural network in Matlab 2015b, and then use the train function to train the network on the normalized data. The traditional BP neural network algorithm uses the trainingd function to train the input data, and the BP neural network based on the L-M algorithm uses the self-compiled function for training. The training results are shown in Figure 4 and Figure 5. The training accuracy and The training times are better than the traditional method.

When a robot finds a new task in the current search direction, the task will be auctioned; or after completing a task, the task with the highest priority stored in the uncompleted task list will be auctioned; when the auction starts, the auctioneer will use the broadcast Publish information to potential bidders in a way; bidders respond according to their current status and the priority of the task they are undertaking; if the bidder is idle or the priority of the rescue task being undertaken is lower than that of the task being auctioned, Then respond to the auctioneer to receive this task and bid according to the current state; otherwise, the bidder will not participate in the auction. If the bidder received by the auction deadline is less than the minimum number of robots required to complete the task, the auction process ends; otherwise, the auctioneer will normalize these bid prices and then use the trained improved BP The neural network comprehensively sorts the processed bidding information, and selects the robot to perform the task according to the sorting size. The invention can effectively make the critically injured get preferential treatment, and the method has the robustness to the robot failure. Computer simulation verifies the training effect of the improved BP neural network.

The above are only preferred specific embodiments of the present invention, and the protection scope of the present invention is not limited thereto. Any person skilled in the art can obviously obtain the simplicity of the technical solution within the technical scope disclosed in the present invention. Variations or equivalent substitutions fall within the protection scope of the present invention.

Claims (2)

1. The multi-robot disaster search and rescue task allocation method based on the improved BP neural network is characterized by comprising the following steps of:

1) each robot of the multi-robot system starts to search the environment according to a depth-first principle;

2) when a certain robot finds a new rescue task in a given direction, performing auction on the task; the auction robot is called an auctioneer;

3) the auctioneer distributes the task information to all robots in the team by broadcasting the information needing to complete the task;

4) each robot in the team is a potential auctioneer, and after receiving search and rescue task information issued by the auctioneer, the robots respond according to the current state of the robots:

if the current robot is in an idle state, participating in bidding;

② if the robot is currently performing a task, but the task performed has a lower priority than the auctioned task,

participation in bidding;

third, if the robot is currently performing a task, but the task performed has a higher priority than the auctioned task,

then not participate in bidding;

5) if the auction conflict occurs, calling an auction conflict resolution strategy;

6) if the number of the bidding robots received by the auctioneer before the auction deadline is less than the number of the robots required for completing the task, the auction process is ended, and the task is stored in respective incomplete task lists by all the robots in the team;

7) otherwise, the auctioneer normalizes the bid price, then uses the trained improved BP neural network to comprehensively sort the multiple bid prices given by the robot to participate in the task, and selects n _c A winner;

8) the auctioneer notifies n _c The winner will execute the task, and the auction receives the notice and sends contract confirmation information to the auctioneer; otherwise, the auctioneer considers that the auctioneer abandons the task because of the fault, and then sequentially selects other robots to participate in the task from other bidding robots according to the evaluation size of the bid value;

9) if the number of the auctioneers receiving the contract confirmation information is larger than the minimum number of robots required for completing the task, establishing a contract and starting to execute the rescue task;

10) if the number of the messages confirmed by the auctioneer after receiving the contract is less than the minimum number of robots required for completing the task, the auctioneer cannot establish the contract; all robots store the tasks into an incomplete task list, the idle robots continue to search in the preset direction, and the rescue robots continue to execute the original rescue tasks;

11) if the contract is established and a certain robot receiving the task is executing the task with lower priority, suspending the task with lower priority and storing the task with lower priority into an uncompleted task list;

12) after a certain robot finishes the undertaken rescue task, if the uncompleted task list is empty, turning to step 1); otherwise, selecting the task with the highest priority from the uncompleted task list for auction, and turning to the step 3);

in the step 7), improving the BP neural network, namely the BP neural network based on the Levenberg-Marquardt algorithm, and setting x ^(k) Is the network weight vector of the kth iteration, the weight vector x of the (k + 1) th iteration ^(k+1) Obtained by the following formula:

x ^(k+1) ＝x ^(k) +Δx ^(k+1) (1)

in the calculation rule of Gauss-Newton method

Δx ^(k+1) ＝-[J ^T (x)J(x)] ^-1 J(x)e(x) (2)

Wherein e (x) is [ e ] ₁ (x)…e _N (x)] ^T Is the output error vector, where x represents x ^k The upper corner marks are omitted under the condition of not causing confusion, and the same is applied below; n is the dimension of the neural network output, J (x) is the Jacobian matrix of error versus weight, which is calculated as

In the weight update value formula (2), J (x) ^T J (x) presence of irreversibleThe following form of weight update is used:

Δx ^(k+1) ＝-[J ^T (x)J(x)+λI] ^-1 J(x)e(x) (4)

in the formula, λ is a constant, and I is an n × n unit matrix.

2. The method for distributing the multi-robot disaster search and rescue task based on the improved BP neural network as claimed in claim 1, wherein in step 3), the tasks published by the auctioneer include the geographical location of the wounded, the task bid deadline time, and the number of robots n required by the task _c Information of the priority of the task.

Images