Multi-agv scheduling method, improved bp neural network training method and related products

Abstract

The application discloses a multi-AGV scheduling method, a training method of an improved BP neural network and related products. In the AGV scheduling method, a work scene feature map is determined according to a work scene panoramic view, an optimal grid length is determined according to the work scene feature map and a vehicle body length, square grid division is performed on the work scene feature map according to the optimal grid length, and an optimal grid map is obtained. A proper grid length is provided to balance between path length and path planning time consumption.

Description

Technical Field

[0001] This invention belongs to the field of automatic control technology, specifically relating to a multi-AGV scheduling method, an improved BP neural network training method, and related products. Background Technology

[0002] Automated Guided Vehicles (AGVs) are increasingly widely used in warehousing and logistics, inspection and security, dock loading and unloading, and automated parking. The AGVs' good adaptability and high flexibility enable them to play an important role in these fields.

[0003] Chinese patent CN111474926B discloses a waste smoke recycling method based on a multi-AGV time window path optimization algorithm. Chinese patent CN115629614B discloses a warehouse multi-AGV path planning method based on a genetic algorithm; however, the methods in these two patents have limited application scenarios. Chinese patent CN109557886B discloses a grid map and a multi-AGV scheduling method based on the grid map, but the grid granularity is imprecise. Furthermore, this method does not consider the shortest path problem. Summary of the Invention

[0004] This invention provides a multi-AGV scheduling method, an improved BP neural network training method, and related products.

[0005] The present invention adopts the following technical solution: a multi-AGV scheduling method, comprising: acquiring a panoramic view of the work scene, feature data of the AGVs, and task objectives of each AGV, wherein the feature data includes vehicle length and a set constant speed; determining a work scene feature map based on the panoramic view of the work scene, determining an optimal grid length based on the work scene feature map and vehicle length, dividing the work scene feature map into square grids based on the optimal grid length to obtain an optimal grid map; using path length as a performance indicator, planning the shortest path for each AGV based on the constant speed and the task objectives of all AGVs in the optimal grid map.

[0006] This invention adopts the following technical solution: a training method for an improved BP neural network, comprising: providing a training sample set, wherein each training sample in the training sample set includes a historical working scene feature map, vehicle length, and optimal grid length, wherein the improved BP neural network is a three-layer neural network, consisting of an input layer, a hidden layer, and an output layer; a single-round iteration step, wherein the historical working scene feature map and vehicle length from each training sample are sequentially input into the improved BP neural network for forward propagation to obtain the output value of each training sample, the cumulative error of the output values ​​of all samples is calculated, and backpropagation is performed based on the cumulative error. The loss function on all samples calculates the partial derivatives of the connection weights of the improved BP neural network. Based on the set learning rate and the calculated partial derivatives of the weights, the neural network weight values ​​are obtained and updated. The loss function includes cumulative error, a first regularization function, and a second regularization function. The first regularization function is used to limit the amplitude of the connection weights of each neuron in the input layer and each neuron in the hidden layer to prevent the weight parameters from being too large. The second regularization function is used to limit the amplitude of the connection weights of each neuron in the hidden layer and each neuron in the output layer to prevent the weight parameters from being too large. The single-round iteration steps are repeated until the cumulative error meets the set accuracy, at which point the iteration stops.

[0007] The present invention adopts the following technical solution: a multi-AGV scheduling system, including a memory and a processor, wherein the memory stores a program and the processor runs the program to execute the aforementioned multi-AGV scheduling method.

[0008] The present invention adopts the following technical solution: an improved BP neural network training device, comprising a memory and a processor, wherein the memory stores a program and the processor runs the program to execute the aforementioned improved BP neural network training method.

[0009] The present invention adopts the following technical solution: a program product that executes the aforementioned multi-AGV scheduling method or the aforementioned improved BP neural network training method during runtime.

[0010] The aforementioned AGV scheduling method and the improved BP neural network training method can provide an appropriate grid length to achieve a balance between path length and path planning time. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the multi-AGV scheduling method according to an embodiment of the present invention.

[0012] Figure 2 This is a flowchart illustrating the improved BP neural network training method in an embodiment of the present invention.

[0013] Figure 3 This is a structural block diagram of a multi-AGV scheduling system according to an embodiment of the present invention.

[0014] Figure 4 This is a structural block diagram of the training device for the improved BP neural network according to an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0016] refer to Figure 1 This invention provides a multi-AGV scheduling method. From a software perspective, the execution entity of this method can be a program product. From a hardware perspective, the execution entity of this method can be a processor, an application-specific integrated circuit (ASIC), or a programmable logic device, etc. The method specifically includes the following steps.

[0017] Step 101: Obtain a panoramic view of the work scene, the feature data of the AGV, and the task objectives of each AGV. The feature data includes the vehicle length and the set constant speed.

[0018] The characteristic data of the AGV includes: vehicle body length l1 and set constant speed v. In the embodiments of the present invention, the characteristic data of each AGV is the same.

[0019] The AGV's mission objectives include: a starting point and a target endpoint. The starting points of all AGVs constitute the starting point set s0. The target endpoints of all AGVs constitute the target endpoint set s0. e .

[0020] The panoramic view P of the work scene was captured by a camera.

[0021] Step 102: Determine the feature map of the work scene based on the panoramic view of the work scene, determine the optimal grid length based on the feature map of the work scene and the length of the vehicle body, and divide the feature map of the work scene into square grids based on the optimal grid length to obtain the optimal grid map.

[0022] A grid is used to discretize the environment into a basic grid, and environmental information is obtained by describing the grid. The environment is divided into the road where the AGV travels and obstacles. In the scenario considered in this invention, the obstacles are fixed.

[0023] Raster mapping can provide a precise and detailed description of complex working environments. The more grid divisions and the smaller the grid length, the higher the accuracy of path planning. However, its computational time and space complexity are relatively high, and the computational load increases exponentially with the number of grids. The optimal raster map refers to the raster map obtained by the minimum number of grid divisions while meeting accuracy requirements.

[0024] Step 102 includes steps 1021 and 1022.

[0025] Step 1021: Input the panoramic view of the work scene into the convolutional neural network to obtain the work scene feature map P2 corresponding to the panoramic view of the work scene.

[0026] Work scene feature map P2: Multiple two-dimensional grids with environmental information. The grids are composed of black and white colors. The black grid has an attribute value of 1, which indicates the area where the obstacle is located, and the white grid has an attribute value of 0, which indicates the feasible area of ​​the AGV.

[0027] Convolutional neural networks extract feature information of roads and obstacles from a panoramic view of the work scene through operations such as convolution and pooling, thereby obtaining a feature map P2 of the work scene.

[0028] Step 1022: Input the work scene feature map and the AGV feature data (vehicle body length l1) into the improved BP (backpropagation) neural network to obtain the optimal grid length r. o (i.e., the grid has a side length of r) o The optimal grid image is obtained by dividing the feature map of the work scene into grids according to the optimal grid length (a square).

[0029] The following is for reference Figure 2 This paper introduces an improved training method for BP neural communication.

[0030] Step 201: Provide a training sample set. Each training sample in the training sample set includes a historical work scene feature map, vehicle length, and optimal grid length. The improved BP neural network is a three-layer neural network, consisting of an input layer, a hidden layer, and an output layer.

[0031] The optimal grid length in the training samples is determined as follows:

[0032] Based on the grid length, the map accuracy loss function of the historical work scene feature map is obtained, where the map accuracy loss function is proportional to the square of the grid length;

[0033] Based on the grid length, a performance function representing the computation time of path planning is calculated, wherein the performance function is inversely proportional to the square of the grid length;

[0034] Within a numerical range greater than or equal to the vehicle length, the optimal grid length is determined by searching for the grid length that minimizes the overall performance evaluation function, which includes both a map accuracy loss function and a performance function.

[0035] Specifically, define the map accuracy loss function G. e For: G e =k1r 2 / λ1, where r represents the grid length, k1 represents the scaling factor, and λ1 represents the weight adjustment parameter. It can be seen that the smaller the grid length, the higher the map accuracy, and the map accuracy loss function G... e The smaller the value, the worse the map accuracy. Conversely, the larger the grid length, the worse the map accuracy loss function G. e The larger the value.

[0036] The smaller the grid length, the longer the computation time for path planning (i.e., subsequent step 103). When the area of ​​the feature map of the working scene is fixed, the performance function G representing the computation time of path planning... T It should be inversely proportional to the square of the grid length r, that is:

[0037]

[0038] The optimal grid length should minimize the map accuracy loss function G. e and the performance function G representing the computation time of path planning T Make them as small as possible, that is

[0039] G = G e +G T

[0040] r o =arg min r (G e +G T )

[0041] r≥l1

[0042] In the formula, G represents the overall performance evaluation function, and l1 is the vehicle body length.

[0043] In applications, the computation time of path planning is difficult to describe precisely. The above formula can be used to calculate and run multiple sets of tests for different grid lengths *r* to minimize the overall performance evaluation function, and then plot the corresponding Gr curves to obtain the optimal grid length. The algorithmic time consumption of path planning cannot be described by a specific formula; only a qualitative description is possible here. Therefore, in the same scenario and for the same type of AGV, the optimal grid length also needs to be determined based on the Gr curve.

[0044] Step 202, Single-round iteration step, wherein the historical working scene feature map and vehicle length in each training sample are sequentially input into the improved BP neural network for forward propagation to obtain the output value of each training sample, the cumulative error of the output values ​​of all samples is calculated, backpropagation is performed based on the cumulative error, the partial derivatives of the connection weights of the improved BP neural network are calculated based on the loss function on all samples, and the neural network weight values ​​are obtained and updated based on the set learning rate and the calculated partial derivatives of the weights. The loss function includes the cumulative error, a first regularization function and a second regularization function. The first regularization function is used to limit the amplitude of the connection weights of each neuron in the input layer and each neuron in the hidden layer to prevent the weight parameters from being too large. The second regularization function is used to limit the amplitude of the connection weights of each neuron in the hidden layer and each neuron in the output layer to prevent the weight parameters from being too large.

[0045] Using historical work scene feature maps, AGV body lengths, and corresponding optimal grid lengths as training samples, a sample dataset S = {(P1, L1, r o1 ),(P2,L2,r o2 ),……,(P m ,L m ,r om The improved BP neural network is trained to obtain the trained improved BP neural network. Here, the subscripts 1 to m represent sample numbers, and P1 to P... m Represents the feature map of historical work scenarios, L1 to L m The length r of the AGV body is indicated by the number of AGVs. o1 to r om This indicates the corresponding optimal grid length.

[0046] Through repeated training, the improved BP neural network has the ability to accurately simulate the complex nonlinear relationship between the optimal grid length and the feature map of the working environment and the length of the AGV body. Under the framework of a three-layer network model (input layer, hidden layer and output layer), it accurately and automatically outputs the optimal grid length.

[0047] Here, an improved BP neural network model is designed. Compared with the traditional BP neural network, the improved BP neural network model can effectively solve the problem of getting trapped in local minima due to unreasonable learning rate settings, thus failing to achieve the desired results.

[0048] Define a three-layer neural network: the input layer has n1 neurons, the hidden layer has n2 neurons, and the output layer has n3 neurons. The connection weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer is ω. ih The connection weight between the h-th neuron in the hidden layer and the j-th neuron in the output layer is ω. hj The input to the h-th neuron in the hidden layer is xi The output of the i-th neuron in the input layer is given, and the input of the j-th neuron in the output layer is given. Among them, b h This represents the output of the h-th neuron in the hidden layer.

[0049] For the training set S = {(P1,L1,r} o1 ),(P2,L2,r o2 ),……,(P m ,L m ,r om )}, Q={(P1,L1),(P2,L2),……,(P m ,L m As the input to the improved BP neural network, Y = {r} o1 ,r o2 ,……,r om} as the expected output. To improve the output value of the BP neural network, the loss function J of the network on all samples is defined as:

[0050]

[0051] in, This represents the cumulative error of all sample output values. This represents the first regularization function, used to apply the connection weights ω between the i-th neuron in the input layer and the h-th neuron in the hidden layer. ih Amplitude limiting is applied. This represents the second regularization function, used to adjust the connection weights ω between the h-th neuron in the hidden layer and the j-th neuron in the output layer. hj Amplitude limiting is applied.

[0052] m represents the number of samples, f represents a weighting parameter used to balance the weights of the first regularization function, the second regularization function, and the cumulative error, n1 represents the number of neurons in the input layer, n2 represents the number of neurons in the hidden layer, and ω... ih ω represents the connection weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer. hj r represents the connection weights between the h-th neuron in the hidden layer and the j-th neuron in the output layer. oi This indicates the expected output. This represents the output value of the improved BP neural network.

[0053] By designing a first regularization function and a second regularization function, the neural network model can learn smaller weights, thereby reducing the complexity of the model, enabling faster convergence, and reducing the probability of overfitting.

[0054] Step 203: Repeat the single-round iteration steps until the cumulative error meets the set accuracy, then stop the iteration.

[0055] The performance function G described above is primarily used to qualitatively analyze the optimal grid length, thereby obtaining the corresponding input-output training samples for the improved BP neural network. The cumulative error E is a function used to measure the difference between the predicted and actual values ​​of the improved BP neural network. The loss function J plays a role in adjusting the neural network weight parameters; minimizing the loss function J is the model optimization objective, and the prediction model parameters are iteratively solved.

[0056] The steps to improve the BP neural network algorithm are as follows.

[0057] Step 1: Initialize the weights of the BP neural network.

[0058] Step 2: Generate training samples Q = {(P1,L1),(P2,L2),……,(P...} m ,L m The inputs to the improved BP neural network are then processed through forward propagation to obtain the output of the improved BP neural network.

[0059] Step 3: Calculate the cumulative error E of all sample output values. The calculation formula is shown below:

[0060]

[0061] Step 4: Based on the cumulative error E, start backpropagation and calculate the partial derivatives of the connection weights based on the loss function J.

[0062] Step 5: Calculate and update the weight parameter values ​​of the improved BP neural network based on the given learning rate λ and the partial derivatives of the weights.

[0063]

[0064]

[0065] Where, ω ih These are the connection weights between the i-th neuron in the input layer and the h-th neuron in the hidden layer. It is ω ih The gradient value, ω hj These are the connection weights between the h-th neuron in the hidden layer and the j-th neuron in the output layer. It is ω hj The gradient values ​​are given by λ, where λ represents the learning rate, u represents the first exponential smoothing of the partial derivatives of the connection weights, r represents the second exponential smoothing of the partial derivatives of the connection weights, and ε is the accuracy parameter.

[0066] Step 6: Repeat steps 2, 3, 4, and 5 iteratively (gradient descent strategy) until the accumulated error meets the set precision, i.e., E < ε.

[0067] The improved BP neural network model can effectively solve the problem that the traditional BP neural network is prone to getting trapped in local minima due to unreasonable learning rate settings, thus failing to achieve the desired results.

[0068] The overall performance function G, as described above, is primarily used to qualitatively analyze the optimal grid length, thereby obtaining the corresponding input-output training samples for the improved BP neural network. The cumulative error E is a function used to measure the difference between the predicted and actual values ​​of the improved BP neural network. The loss function J plays a role in adjusting the neural network weight parameters; minimizing the loss function is the model optimization objective, and the prediction model parameters are iteratively solved.

[0069] Step 103: Using path length as a performance indicator, plan the shortest path for each AGV in the optimal grid map based on the task objectives of all AGVs.

[0070] Step 103 includes steps 1031 to 1035.

[0071] Step 1031: Obtain the starting position s of each AGV. i Let i be the AGV number, and these starting points form a set s0. Obtain the target endpoint t for each AGV. i These target endpoints constitute the target endpoint set s e Set the first set H. o Used to store determined nodes, set the second set H. e Used to store nodes that need to be determined.

[0072] Here, a node represents a grid cell. To find the shortest path, all grid cells are divided into two parts, the first set H. o The nodes in the second set H have already been searched. e The nodes in the diagram need to be further explored and determined.

[0073] Step 1032: For each AGV, starting from the starting point s i Explore towards the endpoint from the center, and calculate the distance from the starting point s. i Find the shortest path node i, then move node i from the second set H. e Take it out and put it into the first set H. o middle.

[0074] For each AGV, each path consists of j pairs of adjacent grid cells between the starting point and the destination. Therefore, the path for an AGV to perform a task can be represented as a set L of ordered grid cell lengths. i ={L 0→1 L 1→2 , ..., L i→j L j→e}

[0075] Step 1033: Calculate the time window for passing through nodes to determine whether the AGV has a conflict.

[0076] Time window: The time interval for passing through each grid cell. The time window is used to calculate the time period during which AGVs occupy path resources. By replanning the occupancy time of road segments, conflicts between multiple AGVs can be avoided. The horizontal axis of the time window represents the time period, and the vertical axis represents the resources, i.e., nodes and paths.

[0077] Set the initial time t of the task o According to step 1032, the starting point s can be calculated. i The distance between node i and the shortest path node is L 0→j Then from the starting point s i The time window to node i can be described as

[0078] Understandably, by analogy, the i-th AGV occupies grid L. i→j The time window is defined as T ij :

[0079]

[0080] Among them, T ij The i-th AGV occupies grid L. i→j Time window, t i Let v be the initial time when the AGV enters the i-th grid node, v represent the AGV's speed, and L represent the initial time. i→j This represents the distance between the i-th grid node and the j-th grid node in the AGV's travel route.

[0081] Step 1034 performs conflict detection.

[0082] The time windows of multiple AGVs are compared to check for conflicts with other AGVs. If there are no conflicts, path information is sent to multiple AGVs and the time windows are updated, and multiple AGVs start to execute tasks. If there are conflicts, step S1035 is executed to resolve the conflicts.

[0083] Each grid cell corresponds to a set of time windows. The set of time windows corresponding to grid cell j is represented as: T j ={T 1j T 2j T 3j ,…,T nj}, where n is the number of time windows corresponding to grid j.

[0084] Whenever a task is assigned to an AGV, a task time window T is added to all the grid cells occupied by the task. lj Perform the following judgment: If If the time windows do not overlap, it indicates no conflict, and path information is sent to multiple AGVs, allowing multiple AGVs to begin executing their tasks; if This indicates that the time windows overlap, causing a conflict in the AGV paths. At this point, step S1035 is executed to resolve the conflict.

[0085] Step 1035 Conflict Resolution

[0086] According to step 1034, when This indicates that a conflict occurs between the l-th AGV and the i-th AGV paths in the j-th grid path. Further conflict resolution is required.

[0087] Determine AGV priority: based on the remaining time t for the AGV to complete its currently executing task. go Determine the priority of indicators, t go The smaller the value, the higher the AGV priority.

[0088] when When time windows overlap, the priority of the l-th AGV and the i-th AGV is determined. The AGV with lower priority needs to wait for passage until the time windows no longer overlap. Then, path information is sent to multiple AGVs, and multiple AGVs start to execute tasks.

[0089] Based on the same inventive concept, and referring to Figure 3 The present invention also provides a multi-AGV scheduling system, including a memory 1a and a processor 2a, wherein the memory 1a stores a program and the processor 2a runs the program to execute the aforementioned multi-AGV scheduling method.

[0090] Based on the same inventive concept, and referring to Figure 4 The present invention also provides an improved BP neural network training device, including a memory 1b and a processor 2b, wherein the memory 1b stores a program and the processor 2b runs the program to execute the aforementioned improved BP neural network training method.

[0091] The present invention adopts the following technical solution: a program product that executes the aforementioned multi-AGV scheduling method or the aforementioned improved BP neural network training method during runtime.

[0092] The various embodiments in this invention are described in a progressive manner. For the same or similar parts between the various embodiments, please refer to each other. Each embodiment focuses on describing the differences from other embodiments.

[0093] The scope of protection of this invention is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its scope and spirit. If these modifications and variations fall within the scope of the claims of this invention and their equivalents, then the intent of this invention also includes these modifications and variations.

Claims

1. A multi-AGV scheduling method, characterized in that, include: Acquire a panoramic view of the work scene, feature data of the AGVs, and task objectives of each AGV. Feature data includes vehicle length and set constant speed. The work scene feature map is determined based on the panoramic view of the work scene. Based on the work scene feature map and the vehicle length, the work scene feature map and the vehicle length are input into an improved BP neural network to determine the optimal grid length. The work scene feature map is then divided into square grids based on the optimal grid length to obtain the optimal grid map. The improved BP neural network is a three-layer neural network, consisting of an input layer, a hidden layer, and an output layer. The input layer has n1 neurons, the hidden layer has n2 neurons, and the output layer has n3 neurons. The connection weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer is ω. ih The connection weight between the h-th neuron in the hidden layer and the j-th neuron in the output layer is ω. hj The input to the h-th neuron in the hidden layer is x i The input layer is the output of the i-th neuron, and the output layer is the input of the j-th neuron. , where b h This represents the output of the h-th neuron in the hidden layer. The optimal grid length is determined as follows: Based on the grid length, the map accuracy loss function of the historical work scene feature map is obtained, where the map accuracy loss function is proportional to the square of the grid length; Based on the grid length, a performance function representing the computation time of path planning is calculated, wherein the performance function is inversely proportional to the square of the grid length; Within a numerical range greater than or equal to the vehicle length, the optimal grid length is determined by searching for the grid length that minimizes the overall performance evaluation function, which includes both a map accuracy loss function and a performance function. Among them, when the area of ​​the feature map of the working scene is fixed, the performance function G representing the computation time of path planning is... T It should be inversely proportional to the square of the grid length r, that is: ； Using path length as a performance indicator, and based on the constant speed and the task objectives of all AGVs, the shortest path is planned for each AGV in the optimal grid diagram, including the following steps: Obtain the starting position s of each AGV i Let i be the AGV number, and these starting points form a set s0. Obtain the target endpoint t for each AGV. i These target endpoints constitute the target endpoint set s e Set the first set H o Used to store determined nodes, set the second set H. e Used to store nodes to be determined; For each AGV, starting from the starting point s i Explore towards the endpoint from the center, and calculate the distance from the starting point s. i Find the shortest path node i, then move node i from the second set H. e Take it out and put it into the first set H. o middle; Calculate the time window for passing through nodes to determine whether the AGV has a conflict; the time window is the time interval when the AGV passes through each grid, the horizontal axis of the time window represents the time period, and the vertical axis represents the resources, including nodes and paths; Set the initial time t of the task o From the starting point s i The time window to node i is T0: ； For the i-th AGV occupying a grid The time window is defined as T ij : ； Among them, L 0→j Starting point s i The distance between the shortest path node i and T; ij The i-th AGV occupies grid L. i→j Time window, t i Let v be the initial time when the AGV enters the i-th grid node, v represent the AGV's speed, and L represent the initial time. i→j This represents the distance between the i-th grid node and the j-th grid node in the AGV's travel route; Each grid cell corresponds to a set of time windows. The set of time windows corresponding to grid cell j is represented as follows: , where n is the number of time windows corresponding to grid j; Whenever a task is assigned to an AGV, a task time window T is added to all the grid cells occupied by the task. lj Make the following judgment: if If the time windows do not overlap, it means there is no conflict. if This indicates that the time windows overlap and the AGV paths conflict. If there is no conflict, path information is sent to multiple AGVs and the time window is updated, and multiple AGVs begin to execute tasks; If there is a conflict, conflict resolution is performed; the conflict resolution includes: Based on the remaining time t for the AGV to complete the currently executing task go Determine the priority of indicators, t go The smaller the value, the higher the AGV priority; The AGVs with lower priority are instructed to wait for passage until the time windows do not overlap. Then, path information is sent to multiple AGVs, and they begin to execute their tasks.

2. An improved training method for a BP neural network, characterized in that, include: A training sample set is provided, where each training sample includes a historical work scene feature map, vehicle length, and optimal grid length. The improved BP neural network is a three-layer neural network, consisting of an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is denoted as n1, the number of neurons in the hidden layer as n2, and the number of neurons in the output layer as n3. The connection weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer is denoted as ω. ih The connection weight between the h-th neuron in the hidden layer and the j-th neuron in the output layer is ω. hj The input value of the h-th neuron in the hidden layer is calculated as follows: x i Let be the output of the i-th neuron in the input layer, and let be the input value of the j-th neuron in the output layer. , where b h This represents the output value of the h-th neuron in the hidden layer. The optimal grid length in the training samples is determined as follows: based on the grid length, the map accuracy loss function of the historical work scene feature map is obtained, where the map accuracy loss function is proportional to the square of the grid length; Based on the grid length, a performance function representing the computation time of path planning is calculated, wherein the performance function is inversely proportional to the square of the grid length; Within a numerical range greater than or equal to the vehicle length, the optimal grid length is determined by searching for the grid length that minimizes the overall performance evaluation function, which includes both a map accuracy loss function and a performance function. Among them, when the area of ​​the feature map of the working scene is fixed, the performance function G representing the computation time of path planning is... T It should be inversely proportional to the square of the grid length r, that is: ； In a single iteration step, the historical working scene feature map and vehicle length from each training sample are sequentially input into the improved BP neural network for forward propagation to obtain the output value of each training sample. The cumulative error of the output values ​​of all samples is calculated, and backpropagation is performed based on the cumulative error. The partial derivatives of the connection weights of the improved BP neural network are calculated based on the loss function on all samples. The neural network weight values ​​are obtained and updated based on the set learning rate and the calculated partial derivatives of the weights. The loss function includes the cumulative error, a first regularization function, and a second regularization function. The first regularization function is used to limit the connection weights of each neuron in the input layer and each neuron in the hidden layer to prevent the weight parameters from being too large. The second regularization function is used to limit the connection weights of each neuron in the hidden layer and each neuron in the output layer to prevent the weight parameters from being too large. Repeat the single-round iteration steps until the cumulative error meets the set accuracy, then stop the iteration.

3. The training method according to claim 2, characterized in that, The map accuracy loss function is denoted as G. e G e =k1r 2 / λ1, where r represents the grid length, k1 represents the scaling factor, and λ1 represents the weight adjustment parameter; The overall performance evaluation function is denoted as G, and G = G e +G T .

4. The training method according to claim 2, characterized in that, Q={(P1,L1),(P2,L2),……,(P m ,L m As the input to the improved BP neural network, Y={r} o1 ,r o2 ,……,r om } represents the expected output, where m represents the number of samples, and P1 to P2 are the values ​​of the samples. m For historical work scene feature maps, L1 to L m Let r be the length of the vehicle body. o1 to r om To achieve the optimal grid length, To improve the output value of the BP neural network, the loss function J of the improved BP neural network on all samples is defined as: in, This represents the cumulative error of all sample output values. This represents the first regularization function, used to apply the connection weights ω between the i-th neuron in the input layer and the h-th neuron in the hidden layer. ih Amplitude limiting is applied. This represents the second regularization function, used to adjust the connection weights ω between the h-th neuron in the hidden layer and the j-th neuron in the output layer. hj Amplitude limiting is applied, where f represents a weighting parameter used to balance the proportions of the first regularization function, the second regularization function, and the cumulative error.

5. The training method according to claim 4, characterized in that, ； ； in, yes gradient value, yes gradient value, Let represent the learning rate, u represent the first exponential smoothing of the partial derivatives of the connection weights and γ represent the second exponential smoothing of the partial derivatives of the connection weights, and ε be the accuracy parameter.

6. A multi-AGV scheduling system, characterized in that, It includes a memory and a processor, the memory storing a program, and the processor running the program to perform the method according to claim 1.

7. A training device for an improved BP neural network, characterized in that, It includes a memory and a processor, the memory storing a program, and the processor running the program to perform the training method according to any one of claims 2 to 5.

8. A program product, characterized in that, It executes the method according to claim 1 or the training method according to any one of claims 2 to 5 during runtime.

Images