Intelligent distribution robot system resource dynamic planning method based on multi-objective optimization

Abstract

The application discloses a kind of intelligent distribution robot system resource dynamic programming method based on multi-objective optimization, belong to intelligent distribution robot system resource dynamic deployment field.The application is related by pearsson chi-square test to effective sample data correlation analysis, filters out the characteristic data variable that has influence on deployment result, and constructs data set;Select L-BPNN neural network with linear mapping relationship to construct the deployment optimization model of intelligent distribution robot;Select particle swarm algorithm to solve the deployment model optimization problem of intelligent distribution robot, effectively solve the multi-objective coupling optimization problem of intelligent distribution robot deployment;Through genetic algorithm, the L-BPNN neural network parameters of intelligent distribution robot deployment optimization model, charging pile number optimization model and single-layer charging pile site selection model are respectively iteratively optimized, the prediction accuracy and efficiency are improved, and then high-precision and high-efficiency resource planning of distribution robot is realized, so that intelligent distribution robot resource reaches optimal deployment under different working conditions.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent distribution robot system resource dynamic deployment, and particularly relates to an intelligent distribution robot system resource dynamic planning method based on multi-objective optimization. BACKGROUND

[0002] In China, major hospitals are operating at high load, and the distribution of medical supplies has always been extremely trivial and heavy work. With the development of science and technology, new logistics equipment such as hospital intelligent distribution robots have solved the existing difficulties in hospital material distribution, and the maturity of 5G network has also promoted the rapid development of intelligent distribution robots.

[0003] Due to the characteristics of ordinary comprehensive hospitals, such as a large number of departments, tight key resources, separation of medicine and ward, and scattered distribution, when studying the intelligent distribution robot system resource planning and dynamic deployment, the optimal configuration target of key resources such as distribution robots and charging piles is needed, the current information state indicators of the hospital area are focused on, and the historical business demand of each department and its environmental background information are combined. Data processing and analysis methods are used to establish and solve the optimization problem model to deploy the key resources of the system. At the same time, a department distribution robot demand prediction model based on intelligent algorithm is established, and according to the change of state indicators, the key resources are timely and dynamically adjusted and planned, thereby improving the utilization efficiency of the key resources of the intelligent distribution robot system. In addition, the efficient distribution of the distribution robot is closely related to the number and location of the charging piles. The appropriate number of charging piles can ensure the power demand of the distribution robot, and the appropriate location can make the distribution robot charge in time and improve the charging efficiency. In addition, the planning of the number and location of the charging piles needs to meet the charging demand of the distribution robot to prevent power shortage, and cannot have redundant idle charging piles to cause resource waste. Therefore, the number and location planning of the charging piles becomes a key.

[0004] In the normal operation of the hospital intelligent distribution robot system, before selecting the distribution robot, the optimal selection scheme of the distribution robot needs to be determined according to the deployment state of the distribution robot in the current departments. Therefore, firstly, the historical business volume and background information of each department in different time ranges are evaluated under the influence of environmental changes such as seasons and climate, and reference is made to typical deployment cases of distribution robots in various situations. Using data processing and analysis methods such as feature data screening and association rule mining, a multi-objective coupled optimization problem model for solving the intelligent distribution robot deployment is designed to obtain the optimal deployment of the distribution robot in each time range. Then, by monitoring the characteristic indexes such as the number of inpatients and the demand for drug volume in the current hospital departments, wards and other places, the optimal deployment number of the distribution robot in each department in the current period is determined to meet the reasonable demand of each department for the distribution robot. Finally, due to the different business volumes of each department in different periods, in order to improve the service efficiency of the intelligent distribution robot in the future, the historical business volume under the influence of complex environmental changes and various characteristic data indexes of each department are analyzed, and intelligent analysis methods such as machine learning and model fitting are used to establish a demand prediction model of the intelligent distribution robot for each department in the future period, which provides a basis for the dynamic deployment of the distribution robot. At the same time, in order to cope with the sudden peak of patients, the real-time state change indexes and the demand prediction model of the intelligent distribution robot for each department under different conditions are used to design a dynamic deployment scheme of the intelligent distribution robot, which improves the business completion efficiency of each department in special periods.

[0005] Considering that the deployment scheme of the intelligent distribution robot is determined, in order to ensure the power supply of the distribution robot and enable the distribution system to run effectively, the number and location optimization of the charging pile of the distribution robot needs to be studied. The number of charging piles of the distribution robot determines the power supply, such as the number of distribution robots required for periodic diseases. When seasonal cold and flu occur, more distribution robots are needed, and the number of charging piles required is also larger. The location of the charging pile of the distribution robot is the key to efficient charging and operation of the distribution robot. For example, the urgency of the distribution task is different from the time efficiency. The urgency of critical treatment is much greater than that of rehabilitation treatment, so the charging pile should be more installed along the way of emergency distribution. In addition, the study of department layout and other characteristic factors, combined with the information of task and historical distribution robot charging time and location, is conducive to the dynamic deployment of the number and location of charging piles, thereby improving the utilization efficiency of charging piles and saving resources. Therefore, the dynamic deployment of the number and location of charging piles of the distribution robot is a prerequisite for the reliable operation of the system. Reasonable arrangement of charging piles can not only prevent power shortage, but also avoid the problem of redundant and idle charging piles, thereby realizing reasonable allocation and efficient utilization of resources. SUMMARY

[0006] In order to solve two technical difficulties in the research on dynamic planning of intelligent distribution robot system resources, the present application mainly aims to provide a dynamic planning method for intelligent distribution robot system resources based on multi-objective optimization, which can efficiently realize optimal deployment of intelligent distribution robots under the condition that the demands of various departments for the intelligent distribution robots are different, and realize optimal planning of the number and position of charging piles on this basis; in addition, the present application can continuously online iteratively optimize the multi-objective coupled intelligent distribution robot deployment optimization model, the charging pile number optimization model and the single-layer charging pile site selection model, thereby improving the prediction accuracy and efficiency of the dynamic planning of intelligent distribution robot system resources based on multi-objective optimization. The present application has the advantages of high prediction accuracy, high efficiency and strong robustness, and is especially suitable for the condition that the demands of various departments are constantly changing during the sudden patient peak period, and can be used in ordinary general hospitals with many departments, tight key resources, separation of medicine and patient rooms and scattered distribution.

[0007] The object of the present application is realized by the following technical solutions.

[0008] To solve two technical problems in the resource dynamic planning of an intelligent delivery robot system, the application discloses an intelligent delivery robot system resource dynamic planning method based on multi-objective optimization. The method analyzes and evaluates the delivery robot deployment data in various scenarios, selects delivery robot deployment sample data, and performs data preprocessing and screening on the sample data to obtain effective sample data that does not include data missing and abnormal problems. The effective sample data is subjected to correlation analysis through Pearson's chi-square test, and feature data variables that have an impact on the intelligent delivery robot deployment result are screened out, including historical business volume in different periods, delivery demand volume, delivery volume, average delivery time, position distribution of delivery robots, and number of delivery robots. A dataset for constructing a deployment optimization model of the intelligent delivery robot is constructed according to the feature data variables. The feature data variables and the intelligent delivery robot deployment result have both a nonlinear relationship and a linear relationship, and an L-BPNN neural network with a linear mapping relationship is selected to construct the deployment optimization model of the intelligent delivery robot, so that the deployment optimization model of the intelligent delivery robot is more suitable for the actual working conditions of the intelligent delivery robot. The deployment optimization model of the intelligent delivery robot is further improved in combination with constraint conditions based on actual conditions. The particle swarm algorithm is selected to solve the deployment model optimization problem of the intelligent delivery robot, effectively solving the multi-objective coupled optimization problem of the intelligent delivery robot deployment and improving the deployment precision of the intelligent delivery robot. Through a deep learning modeling method, an L-BPNN neural network with a linear mapping relationship is selected to construct a feature variable change prediction model to predict the deployment change of the delivery robot in each department in a future period, and the prediction result is used as prior knowledge. The deployment change of the delivery robot in each department in a future period is obtained by using the established delivery robot deployment optimization model. An L-PBNN neural network is used to construct a charging pile number optimization model, and the charging pile number planning is converted into an optimization problem. The momentum gradient descent method is used to solve the optimization problem to determine the number of charging piles. Under the constraints of the deployment environment of the charging piles and the task delivery timeliness, an L-PBNN neural network is used to construct a single-layer charging pile site selection model with the maximum utilization rate of resources as the target. The particle swarm algorithm is used to solve the appropriate position of the charging pile for the established charging pile site selection model. Based on the actual demand information of the delivery robot in each department, the genetic algorithm is used to iteratively optimize the L-BPNN neural network parameters of the intelligent delivery robot deployment optimization model, the charging pile number optimization model, and the single-layer charging pile site selection model, thereby improving the prediction accuracy and efficiency of the intelligent delivery robot deployment optimization model. The particle swarm algorithm is used to realize high-precision and high-efficiency resource planning of the delivery robot, so that the intelligent delivery robot resource can achieve optimal deployment under different working conditions.

[0009] The two technical problems refer to:

[0010] 1. Under the background of complex environment changing, the business volume of each department in different period is different, and the demand for intelligent delivery robot is different. It is urgent to analyze the optimal deployment mode of intelligent robot based on the change of business information in different period, and to meet the business demand of current robot delivery by reasonably deploying the number of current delivery robots according to the change of current information index. In case of emergency, it is urgent to provide an optimal deployment mode of robot based on the change of different state indexes to meet the urgent demand of robot delivery in emergency. Therefore, under the complex and changeable hospital environment, how to optimally deploy and dynamically plan the key resources of intelligent delivery system has become a key technical problem to be solved.

[0011] 2. The insufficient number of charging piles will cause power shortage, and too many charging piles will cause waste of charging resources. The location deployment of charging piles is related to whether the delivery robot can be charged efficiently and timely. At the same time, reasonable optimization and deployment of the number and location of charging piles can avoid problems such as charging waiting and equipment idling, and thus improve the utilization rate of resources, so as to promote the effective operation of the system. Therefore, it is urgent to establish a deployment optimization model of the number and location of charging piles for the hospital intelligent delivery system to solve the optimal deployment mode of charging piles and realize maximum resource utilization.

[0012] The application discloses an intelligent delivery robot system resource dynamic planning method based on multi-objective optimization, which comprises the following steps:

[0013] Step 1: analyze, preprocess and select the deployment data of the intelligent delivery robot to build a data set of the deployment optimization model of the intelligent delivery robot, and build the deployment optimization model of the intelligent delivery robot through an L-BPNN neural network.

[0014] The specific implementation method of the deployment optimization model of the intelligent delivery robot in step 1 is as follows:

[0015] Step 1.1: analyze the deployment data of the intelligent delivery robot and select sample data, and perform data preprocessing and selection on the sample data to obtain effective sample data not including data missing and abnormal problems, effectively select effective sample data from a large amount of data, and reduce the influence of invalid sample data on the establishment of the deployment optimization model of the intelligent delivery robot as much as possible.

[0016] The deployment data includes historical business volume in different periods, and corresponding period demand amount, delivery amount, average delivery time, position distribution of delivery robots, and number of delivery robots. For data in different working conditions, the selected historical sample data is cleaned through data preprocessing, and data problems in historical order samples caused by other reasons are processed. The data preprocessing method includes interpolation, fitting and rejection. The other reasons include data transfer and human recording errors. The data problems include data missing and data abnormality.

[0017] Step 1.2: For the valid sample data obtained in step 1.1, the correlation analysis of the valid sample data is carried out by Pearson chi-square test, the feature data variables which can have influence on the deployment result of intelligent delivery robots are screened out, and the data set is constructed. The feature data variables which have no influence on the deployment effect of intelligent delivery robots are effectively deleted, the learning speed of neural network is accelerated, and the efficiency of establishing the deployment optimization model of intelligent delivery robots is improved. The feature data variables include historical business volume in different periods, and corresponding period demand amount, delivery amount, average delivery time, position distribution of delivery robots, and number of delivery robots.

[0018] The feature data variables which can have influence on the deployment result are screened out by using Pearson chi-square test. The "null hypothesis" is that the historical business volume in different periods is statistically independent of the deployment result. Then a contingency table is arranged. The contingency table is r rows and c columns, and the theoretical number E i,j As follows:

[0019]

[0020] Where N is the sample size, and then the statistical value χ 2 :

[0021]

[0022] According to the set confidence level, the chi-square distribution critical value with degree of freedom df = rc-1 is found out, and it is compared with the statistical value χ 2 . If the statistical value is larger, the null hypothesis cannot be rejected, that is, the feature data variable is statistically independent of the deployment result. Similarly, the statistical independence of the deployment data of all intelligent delivery robots and the deployment result can be obtained, and the feature data variables which can have influence on the deployment result are screened out. The deployment data includes variables including historical business volume in different periods, and corresponding period demand amount, delivery amount, average delivery time, position distribution of delivery robots, and number of delivery robots.

[0023] According to the characteristic data variable having an influence on the deployment result, a data set for constructing a deployment optimization model of the intelligent delivery robot is constructed. The characteristic data variable having no influence on the deployment effect of the intelligent delivery robot is effectively pruned, the learning speed of the neural network is accelerated, and the efficiency of establishing the deployment optimization model of the intelligent delivery robot is improved.

[0024] Step 1.3: For the data set obtained in step 1.2, a deployment optimization model of the intelligent delivery robot is constructed by an L-BPNN neural network. The deployment optimization model of the intelligent delivery robot is more suitable for the actual working conditions of the deployment of the intelligent delivery robot.

[0025] The deployment optimization model of the intelligent delivery robot is established by a deep learning method. The BPNN has good nonlinear mapping ability, strong self-learning ability, good generalization ability, and excellent fault tolerance ability. Among them, the nonlinear mapping ability is its most prominent advantage, that is, without prior understanding of the specific mathematical relationship, only through modeling can the complex nonlinear mapping relationship between the input and the output be obtained. However, the relationship between the deployment effect and the characteristic data variable is not only a nonlinear relationship, but also a linear relationship. Through analysis, it is found that the characteristic data variable and the deployment result of the intelligent delivery robot have not only a nonlinear relationship, but also a linear relationship, which is a combination of nonlinear and linear relationships. At this time, the BPNN may not be able to accurately express completely. Therefore, it is necessary to improve the BPNN from the network topology structure to obtain a BPNN with linear mapping relationship, that is, L-BPNN. The L-BPNN neural network with linear mapping relationship is selected to construct the deployment optimization model of the intelligent delivery robot, so that the deployment optimization model of the intelligent delivery robot is more suitable for the actual working conditions of the deployment of the intelligent delivery robot.

[0026] The L-BPNN neural network structure is based on the BPNN, and the direct connection between the input layer and the output layer is added, reflecting the linear and nonlinear relationship between the input and the output. Among them: x1, x2, …, x l-1 ,x l represents input information, h1, h2, … h m-1 ,h m represents the hidden layer, y1, y2, … y n-1 ,y n represents the output value, w ji represents the weight between the input layer and the hidden layer, w kj represents the weight between the hidden layer and the output layer, represents the weight between the input layer and the output layer. Therefore, the output calculation expression of the L-BPNN hidden layer is still:

[0027]

[0028] where g(x) represents the activation function of the hidden layer, commonly sigmoid function, m represents the number of neurons in the hidden layer, b j represents the bias of the hidden layer.

[0029] The expression corresponding to the output layer is represented as:

[0030]

[0031] where n represents the number of neurons in the output layer, b k represents the bias of the output layer.

[0032] The deployment optimization model prediction running framework of the intelligent distribution robot based on the L-BPNN neural network also contains four parts: the establishment of network topology structure and the initialization of related parameters, the learning and training of L-BPNN, the acquisition of optimal network parameters, and the prediction of the deployment optimization model of the intelligent distribution robot.

[0033] After the deployment optimization model of the distribution robot is established using the L-BPNN neural network, a constraint model is established based on real-time information of feature variables, and optimization indexes based on distribution time and distribution efficiency as cost. The feature variables include historical business volume at different periods and distribution demand, distribution volume, average distribution time, position distribution of distribution robots, and the number of distribution robots at corresponding periods. The optimization sub-problems are defined as follows:

[0034]

[0035] and

[0036]

[0037] where J i (x i ) is the distribution efficiency cost of the i th distribution robot under deployment, T i (x i ) is the distribution time cost under deployment, M i is the current task volume, m i is the maximum load; t i is the single distribution time, D i is the overall distance of single distribution, E i is the current power.

[0038] Step two: select the particle swarm algorithm to solve the deployment model optimization problem of the intelligent distribution robot, effectively solve the multi-objective coupling optimization problem of the intelligent distribution robot deployment through the particle swarm algorithm, and improve the deployment accuracy of the intelligent distribution robot. A feature variable change prediction model is established to obtain the intelligent distribution robot deployment change in the future period.

[0039] Step 2.1: Select the particle swarm optimization algorithm to solve the deployment optimization model of the intelligent distribution robot obtained in step 1, and improve the deployment accuracy of the intelligent distribution robot.

[0040] The particle swarm optimization algorithm randomly generates a certain number of particles (the specific number will be discussed later) as effective solutions of the problem search space, then iteratively searches, determines the fitness value of the particles through the fitness function corresponding to the problem, and obtains the optimization result. The core of the particle swarm optimization algorithm is to use the information sharing of individuals in the group to make the movement of the entire group evolve from disorder to order in the problem solving space, so as to obtain the optimal solution of the problem.

[0041] The specific process is as follows:

[0042] ①Initialize all particles, that is, assign values to their speed and position, and set the historical optimal pBest of the individual as the current position, and the optimal individual in the group as the current gBest.

[0043] ②In the evolution of each generation, calculate the fitness function value of each particle.

[0044] ③If the current fitness function value is better than the historical optimal value, update pBest.

[0045] ④If the current fitness function value is better than the global historical optimal value, update gBest.

[0046] ⑤Update the speed and position of each particle i in the dth dimension according to the following formulas respectively:

[0047]

[0048]

[0049] Where ω is the inertia weight, c1 and c2 are acceleration coefficients, and are two random numbers in [0, 1].

[0050] Step 2.2: Through the deep learning modeling method, select the L-BPNN neural network with linear mapping relationship to construct a feature variable change prediction model to predict the deployment change of the intelligent distribution robot in the future period. The prediction result is used as prior knowledge, and the deployment optimization model of the distribution robot is used to obtain the deployment change of the distribution robot in each department in the future period, and the robustness of the resource dynamic planning method is improved. Further effectively cope with the environmental mutation caused by sudden events. The environmental mutation includes rapid increase of tasks and change of distribution conditions in some areas.

[0051] The L-BPNN neural network structure is based on the BPNN, and a direct connection between the input layer and the output layer is added, reflecting the linear and nonlinear relationship between the input and the output. Wherein x1, x2, …, x l-1 l represents the input information, h1, h2, … h m-1 m represents the hidden layer, y1, y2, … y n-1 n represents the output value, w ji represents the weight between the input layer and the hidden layer, w kj represents the weight between the hidden layer and the output layer, represents the weight between the input layer and the output layer. Therefore, the output calculation expression of the hidden layer of the L-BPNN is still:

[0052]

[0053] wherein g(x) represents the activation function of the hidden layer, and the sigmoid function is commonly used. m represents the number of neurons in the hidden layer, and b j represents the bias of the hidden layer.

[0054] The expression corresponding to the output layer is represented as:

[0055]

[0056] wherein n represents the number of neurons in the output layer, and b k represents the bias of the output layer.

[0057] The prediction model is used to predict the changes of the characteristic variables in a future period, and then the changes of the deployment of the delivery robots in each department in a future period are obtained as prior knowledge, so as to effectively cope with the environmental mutation caused by the sudden event. The environmental mutation includes rapid increase of tasks and change of delivery conditions in some areas.

[0058] Step three: using the L-PBNN neural network to construct the charging pile quantity optimization model, and using the momentum gradient descent method to solve the optimization problem to determine the number of charging piles; under the constraints of the deployment environment of the charging pile and the task delivery time efficiency, using the L-PBNN neural network to construct a single-layer charging pile site selection model, and using the particle swarm algorithm to solve the appropriate position of the charging pile, which can prevent power shortage and avoid the problem of redundant and idle charging piles.

[0059] Step 3.1: using the L-PBNN neural network to construct the charging pile quantity optimization model, so that the charging pile quantity optimization model of the intelligent delivery robot is more suitable for the actual working condition of the charging scheduling of the intelligent delivery robot. ​​​

[0060] The L-BPNN neural network structure is based on the BPNN, and a direct connection between the input layer and the output layer is added, reflecting the linear and nonlinear relationship between the input and the output. Among them: x1, x2, …, x l-1 l represents input information, h1, h2, … h m-1 m represents hidden layer, y1, y2, … y n-1 n represents output value, w ji represents the weight between the input layer and the hidden layer, w kj represents the weight between the hidden layer and the output layer, represents the weight between the input layer and the output layer. Therefore, the output calculation expression of the L-BPNN hidden layer is still:

[0061]

[0062] where g(x) represents the activation function of the hidden layer, and the sigmoid function is commonly used. m represents the number of neurons in the hidden layer, b j represents the bias of the hidden layer.

[0063] The expression corresponding to the output layer is represented as:

[0064]

[0065] where n represents the number of neurons in the output layer, b k represents the bias of the output layer.

[0066] The optimization problem is defined as follows:

[0067]

[0068] where k i represents the number of charging piles.

[0069] Step 3.2: Use the momentum gradient descent method to solve the charging pile number optimization model obtained in step 3.1 to determine the number of charging piles.

[0070] v i = γv i + ηΔL(θ) (14)

[0071] θ i = θ i-1 - v i (15)

[0072] where v i is the current speed, γ is the momentum parameter, and η is the learning rate.

[0073] ​​​Step 3.3: Construct a charging pile site selection model for the single layer of the intelligent distribution robot using the L-PBNN neural network, so that the single layer charging pile site selection model of the intelligent distribution robot is more in line with the actual working conditions of the charging scheduling of the intelligent distribution robot.

[0074] The L-BPNN neural network structure is based on BPNN, with the addition of direct connection between the input layer and the output layer, reflecting the linear and nonlinear relationship between the input and the output. Among them: x1, x2, …, x l-1 l represents the input information, h1, h2, … h m-1 m represents the hidden layer, y1, y2, … y n-1 n represents the output value, w ji represents the weight between the input layer and the hidden layer, w kj represents the weight between the hidden layer and the output layer, represents the weight between the input layer and the output layer. Therefore, the output calculation expression of the L-BPNN hidden layer is still:

[0075]

[0076] where g(x) represents the activation function of the hidden layer, commonly used sigmoid function. m represents the number of hidden layer neurons, b j represents the hidden layer bias.

[0077] The expression corresponding to the output layer is represented as:

[0078]

[0079] where n represents the number of output layer neurons, b k represents the output layer bias.

[0080] The optimization problem is defined as follows:

[0081]

[0082] where α i represents the number of distribution addresses covered by the i-th charging pile, β i represents the number of initial positions of the distribution robot covered by the i-th charging pile.

[0083] Step 3.4: Use the particle swarm algorithm to solve the single layer charging pile site selection model obtained in step 3.3 to determine the deployment position of the charging pile.

[0084] The specific process is as follows:

[0085] ​​​①Initialize all particles, i. e. assign their velocities and positions, and set the individual's historical best pBest as the current position, and the best individual in the group as the current gBest.

[0086] ②In each generation of evolution, the fitness function value of each particle is calculated.

[0087] ③If the current fitness function value is better than the historical optimal value, update pBest.

[0088] ④If the current fitness function value is better than the global historical optimal value, update gBest.

[0089] ⑤The velocity and position of each particle i in the dth dimension are updated according to the following formula respectively:

[0090]

[0091]

[0092] where ω is the inertia weight, c1 and c2 are acceleration coefficients, and are two random numbers in [0, 1].

[0093] It also includes step four: based on the actual demand information of each department for the distribution robot, the L-BPNN neural network parameters of the intelligent distribution robot deployment optimization model, the charging pile number optimization model and the single-layer charging pile site selection model are respectively iteratively optimized by genetic algorithm, and the prediction accuracy and efficiency of the intelligent distribution robot deployment optimization model are improved. The optimization and dynamic deployment of the charging pile number and position of the distribution robot is the prerequisite guarantee for the reliable operation of the system, and reasonable arrangement of the charging pile can not only prevent power shortage, but also avoid the problem of redundant idle of the charging pile, thereby realizing reasonable allocation and efficient utilization of resources, and making the intelligent distribution robot resources achieve optimal deployment under different working conditions.

[0094] The part optimized by the genetic algorithm is the weight w ji and the hidden layer bias b j between the input layer and the hidden layer, the weight w kj and the output layer bias b k between the hidden layer and the output layer, and the weight

[0095] The neural network is used as the main function of the genetic algorithm, the error function of the L-BPNN neural network is set as the fitness function, the population number is the effective sample number ρ, and the fitness of a single individual is τ i , and its selected probability is:

[0096]

[0097] The single individual refers to the weight w between the input layer and the hidden layer ji And the hidden layer bias b j The binary representation of the weight w between the hidden layer and the output layer kj And the output layer bias b k The cross and mutation operations are performed to obtain offspring, and the optimal individual in each iteration process is reserved by calculating the fitness of the individual. Further, the rational allocation and efficient use of resources are realized, so that the intelligent distribution robot resources reach the optimal deployment under different working conditions. The optimal individual refers to the optimal weight w between the input layer and the hidden layer of the L-BPNN neural network of the intelligent distribution robot deployment optimization model, the charging pile number optimization model and the single-layer charging pile site selection model ji And the hidden layer bias b j The weight w between the hidden layer and the output layer kj And the output layer bias b k .

[0098] Beneficial effects:

[0099] 1. The intelligent distribution robot system resource dynamic planning method based on multi-objective optimization disclosed in the application analyzes and selects sample data for the existing intelligent distribution robot deployment data, and performs data preprocessing and screening on the sample data to obtain effective sample data which has an impact on the distribution result and does not include data missing and abnormal problems. The effective sample data which has an impact on the distribution result is effectively screened out from a large amount of data, and the influence of invalid sample data on the establishment of the intelligent distribution robot deployment optimization model is reduced as much as possible.

[0100] 2. The intelligent distribution robot system resource dynamic planning method based on multi-objective optimization disclosed in the application selects an L-BPNN neural network with a linear mapping relationship to construct an intelligent distribution robot deployment optimization model, so that the intelligent distribution robot deployment optimization model is more suitable for the actual working condition of the intelligent distribution robot.

[0101] 3. The intelligent distribution robot system resource dynamic planning method based on multi-objective optimization disclosed in the application selects a particle swarm algorithm to solve the intelligent distribution robot deployment optimization model, effectively solves the multi-objective coupling optimization problem of the intelligent distribution robot deployment, and improves the deployment precision of the intelligent distribution robot.

[0102] 4. The method comprises the following steps: selecting an L-BPNN neural network with a linear mapping relationship to construct a characteristic variable change prediction model, and predicting the changes of the characteristic variables in the future period of time, thereby effectively responding to environmental mutations caused by sudden events. The environmental mutations include rapid increase of tasks and changes of distribution conditions in some areas.

[0103] 5. The method comprises the following steps: selecting an L-BPNN neural network with a linear mapping relationship to construct a charging pile number optimization model and a single-layer charging pile site selection model, so that the charging pile number optimization model and the single-layer charging pile site selection model are more suitable for the actual working conditions of the intelligent distribution robot.

[0104] 6. The method comprises the following steps: selecting a momentum gradient descent method to solve the charging pile number optimization model, and obtaining the optimal number of charging piles.

[0105] 7. The method comprises the following steps: selecting a particle swarm algorithm to solve the single-layer charging pile site selection model, and obtaining the optimal position of the single-layer charging pile.

[0106] 8. The method comprises the following steps: iterating and optimizing the parameters of the L-BPNN neural network of the intelligent distribution robot deployment optimization model, the charging pile number optimization model and the single-layer charging pile site selection model through a genetic algorithm, improving the prediction accuracy and efficiency of the above models, and making the intelligent distribution robot resources reach the optimal deployment under different working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0107] Figure 1 is a flowchart of the method.

[0108] Figure 2 is a flowchart of the L-BPNN neural network.

[0109] Figure 3 is an algorithm flowchart of the genetic algorithm.

[0110] Figure 4 is an effect diagram of resource deployment. Wherein 4(a) is an initial position distribution diagram of task points and charging piles, 4(b) is an optimal distribution route of a single robot, 4(c) is an optimal distribution time of a single robot, 4(d) is an optimization diagram of the deployment position of a single robot, and 4(e) is an optimization diagram of the position of a single charging pile. DETAILED DESCRIPTION

[0111] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0112] Example 1:

[0113] This embodiment discloses a multi-objective optimization charging scheduling method for delivery robots with spatiotemporal coupling constraints, which is applied to the modern intelligent medical delivery system of a general hospital with many departments, scarce key resources, and separate and scattered pharmacies and wards.

[0114] like Figure 1 As shown in the figure, this embodiment discloses a multi-objective optimization charging scheduling method for delivery robots with spatiotemporal coupling constraints. The specific implementation steps are as follows:

[0115] Step 1: Construct an optimization model for the deployment of intelligent delivery robots;

[0116] Step 1.1: Analyze the deployment data of existing intelligent delivery robots and select sample data. Perform data preprocessing and filtering on this sample data to obtain valid sample data that does not contain missing or anomalies. This effectively filters out valid sample data from a large amount of data.

[0117] Step 1.2: Perform correlation analysis on the valid sample data using the Pearson chi-square test to screen out the characteristic data variables that can influence the deployment results of intelligent delivery robots, and construct a dataset. The characteristic data variables include historical business volume in different periods and corresponding delivery demand, delivery volume, average delivery time, location distribution of delivery robots, and number of delivery robots.

[0118] Pearson's chi-square test was used to screen for characteristic data variables that could influence the deployment results. The "null hypothesis" was that historical traffic volume in different periods was statistically independent of the deployment results. Then, a contingency table was compiled. The contingency table has r rows and c columns, with a theoretical frequency E. i,j as follows:

[0119]

[0120] Where N is the sample size, then the statistic χ² is calculated. 2 :

[0121]

[0122] Based on the set confidence level, find the chi-square distribution critical value for degrees of freedom df = rc-1, and compare it with the statistical value χ². 2If the statistical value is large, the null hypothesis cannot be rejected, that is, the feature data variable is statistically independent of the deployment result. Similarly, the statistical independence of the deployment data of all intelligent delivery robots and the deployment result can be obtained, and the feature data variable that can affect the deployment result is screened out. The deployment data includes variables including historical business volume in different periods and corresponding period demand for each distribution, distribution volume, average distribution time, distribution robot position distribution, and distribution robot quantity.

[0123] Then, a dataset for constructing a deployment optimization model of an intelligent delivery robot is constructed according to the feature data variable that can affect the deployment result.

[0124] Step 1.3: Constructing a deployment optimization model of an intelligent delivery robot through an L-BPNN neural network. The deployment optimization model of the intelligent delivery robot is more suitable for the actual working conditions of the deployment of the intelligent delivery robot.

[0125] Figure 2 The network structure in the formula is based on BPNN, and a direct connection between the input layer and the output layer is added. The linear and nonlinear relationship between the input and the output is revealed. In the figure, x1, x2, …, x l-1 ,x l represent input information, h1, h2, … h m-1 ,h m represent hidden layers, y1, y2, … y n-1 ,y n represent output values, w ji represents the weight between the input layer and the hidden layer, w kj represents the weight between the hidden layer and the output layer, represents the weight between the input layer and the output layer. Therefore, the output calculation expression of the hidden layer of the L-BPNN is still:

[0126]

[0127] In the formula, g(x) represents the activation function of the hidden layer, and the commonly used sigmoid function. m represents the number of neurons in the hidden layer, and b j represents the bias of the hidden layer.

[0128] The expression corresponding to the output layer can be represented as:

[0129]

[0130] In the formula, n represents the number of neurons in the output layer, and b k represents the bias of the output layer.

[0131] After the deployment optimization model of the delivery robot is established using the L-BPNN neural network, a constraint model is established based on real-time information of feature variables, and optimization indexes are established based on delivery time and delivery efficiency as costs. The feature variables include historical business volume in different periods and corresponding delivery demand volume, delivery volume, average delivery time, position distribution of the delivery robot, and the number of delivery robots in the corresponding period. The optimization sub-problems are defined as follows:

[0132]

[0133] and

[0134]

[0135] wherein, J i (x i ) is the delivery efficiency cost of the i th delivery robot under deployment, T i (x i ) is the delivery time cost under deployment, M i is the current task volume, m i is the maximum load, σ j is the delivery volume of a single task; t i is the single delivery time, D i is the overall distance of single delivery, E i is the current power, T is the longest time limit of the delivery task, is the energy consumption per kilometer under full load of the robot. In this example, there are 4 intelligent delivery robots to jointly bear 64 delivery tasks.

[0136] Step two: select the particle swarm algorithm to solve the deployment model optimization problem of the intelligent delivery robot, and establish a feature variable change prediction model.

[0137] Step 2.1: select the particle swarm algorithm to solve the deployment optimization model of the intelligent delivery robot.

[0138] The specific process is as follows:

[0139] (1) initialize all particles, that is, assign values to their speed and position, and set the historical optimum pBest of the individual as the current position, and the best individual in the group as the current gBest.

[0140] (2) In each generation of evolution, the fitness function value of each particle is calculated.

[0141] (3) If the current fitness function value is better than the historical optimum value, update pBest.

[0142] (4) If the current fitness function value is better than the global historical optimum value, update gBest.

[0143] (5) The velocity and position of each particle i in the dth dimension are updated according to the following equations, respectively:

[0144]

[0145]

[0146] where ω is the inertia weight, which is initialized as 0.9 and decreased to 0.4 during the evolution process, and c1 and c2 are acceleration coefficients, which are fixed as 2.0, and are two random numbers in [0, 1]. It is worth noting that during the updating process, the PSO algorithm requires setting a v max to limit the velocity range, which is generally 10% to 20% of the corresponding dimension of v max . The optimization effect is shown in Figure 4(d). Figure 4(d) is an optimization diagram of the deployment position of one intelligent delivery robot, and the coordinates of the delivery task locations are shown in Table 1. Figures 4(a) and 4(b) show the shortest path of delivery and the shortest time of delivery, respectively, and the shortest distance is 2569.653 meters, and the shortest time is 2.5697 hours.

[0147] Table 1 Coordinates of delivery task locations of a single intelligent delivery robot

[0148] NO. XCOORD. YCOORD. NO. XCOORD. YCOORD. 1 35 35 9 10 43 2 41 49 10 55 60 3 35 17 11 30 60 4 55 45 12 20 65 5 55 20 13 50 35 6 15 30 14 30 25 7 25 30 15 15 10 8 20 50 16 30 5

[0149] Step 2.2: Construct a feature variable change prediction model and predict the changes of feature variables in the future period.

[0150] This part is similar to the previous one, and the structure of the L-BPNN neural network is shown in Figure 2 . In the figure, x1, x2, …, x l-1 , x l represent input information, h1, h2, … h m-1 , h m represent hidden layers, y1, y2, … y n-1 , y n represent output values, w ji represents the weight between the input layer and the hidden layer, w kj represents the weight between the hidden layer and the output layer, and w represents the weight between the input layer and the output layer. Therefore, the output calculation expression of the L-BPNN hidden layer is still:

[0151]

[0152] where g(x) represents the activation function of the hidden layer, commonly sigmoid function. m represents the number of neurons in the hidden layer, b j represents the bias of the hidden layer.

[0153] The expression corresponding to the output layer can be expressed as:

[0154]

[0155] where n represents the number of neurons in the output layer, b k represents the bias of the output layer.

[0156] The change prediction running framework of the feature variable based on L-BPNN can be divided into the following parts:

[0157] (1) Determination of network topology structure and initialization of related parameters. Select actual parameters as input information and feature variables as output. The actual parameters include historical task quantity and corresponding period patient quantity, current task quantity and current patient quantity. Here, the number of hidden layers is set to one. After establishing the network topology structure, the weights and biases in the network are randomly initialized.

[0158] (2) Learning and training of L-BPNN. Divide the data into training set and test set. The input information of the training set is transmitted from the input layer to the output layer through the hidden layer, and the feature variable of the network is output after the calculation of the output layer. Calculate the error between the feature variable of the network output and the corresponding actual feature variable, and propagate the error back to adjust the weights and biases between the layers.

[0159] (3) Obtain the optimal network structure parameters. Recursively iterate step (2), gradually reduce the calculation error, until the error reaches the set target error or the number of iterations reaches the set maximum number, and obtain the optimal weights and biases.

[0160] (4) Prediction of the change trend of the feature variable. Input the input information of the test set into the trained network, calculate the feature variable of the test set according to the optimal weights and biases, and obtain the change trend.

[0161] Step three: build a charging pile quantity optimization model and use the momentum gradient descent method to solve the optimization problem to determine the number of charging piles; build a single-layer charging pile site selection model and use the particle swarm algorithm to solve the appropriate location of the charging pile.

[0162] Step 3.1: Build a charging pile quantity optimization model.

[0163] This part is similar to the previous one, and the structure of the L-BPNN neural network is shown in Figure 2 Fig. 1. In the figure, x1, x2, …, x l-1 , xl represents input information, h1, h2, … h m-1 m represents hidden layer, y1, y2, … y n-1 n represents output value, w ji represents weight between input layer and hidden layer, w kj represents weight between hidden layer and output layer, represents weight between input layer and output layer. Therefore, the output calculation expression of the hidden layer of the L-BPNN is still:

[0164]

[0165] where g(x) represents the activation function of the hidden layer, and the sigmoid function is commonly used. m represents the number of neurons in the hidden layer, and b j represents the bias of the hidden layer.

[0166] The expression corresponding to the output layer can be represented as:

[0167]

[0168] where n represents the number of neurons in the output layer, and b k represents the bias of the output layer.

[0169] The optimization problem is defined as follows:

[0170]

[0171] where k i represents the number of charging piles.

[0172] Step 3.2: Use the momentum gradient descent method to solve the charging pile number optimization model to determine the number of charging piles.

[0173] v i = γv i + ηΔL(θ) (14)

[0174] k i = k i-1 - v i (15)

[0175] where v i is the current speed, γ is the momentum parameter, which is a positive number not exceeding 1, η is the learning rate, and i is the iteration number. In this example, the number of charging piles finally obtained is three.

[0176] Step 3.3: Build a charging pile site selection model for the single layer.

[0177] This part is similar to the previous one, and the structure of the L-BPNN neural network is as follows​​Figure 2 As shown in the figure. x1, x2, ..., x l-1 ,x l The input information is represented by h1, h2, ... h. m-1 ,h m Represents the hidden layers, y1, y2, ... y n-1 ,y n Indicates the output value, w ji w represents the weights between the input layer and the hidden layer. kj This represents the weights between the hidden layer and the output layer. This represents the weights between the input and output layers. Therefore, the expression for calculating the output of the L-BPNN hidden layer remains:

[0178]

[0179] In the formula, g(x) represents the activation function of the hidden layer, commonly the sigmoid function. m represents the number of neurons in the hidden layer, b j This indicates the hidden layer bias.

[0180] The expression for the corresponding output layer can be represented as:

[0181]

[0182] In the formula, n represents the number of neurons in the output layer, b k This indicates the output layer bias.

[0183] The optimization problem is defined as follows:

[0184]

[0185] In the formula α i β represents the number of delivery addresses covered by the i-th charging station. i This represents the number of initial locations of delivery robots covered by the i-th charging station.

[0186] Step 3.4: Use the particle swarm optimization algorithm to solve the single-layer charging pile site selection model and determine the deployment location of the charging piles.

[0187] The specific process is as follows:

[0188] ① Initialize all particles, that is, assign values ​​to their velocity and position, and set the historical best pBset of each individual to the current position, and the best individual in the group as the current gBest.

[0189] ② In each generation of evolution, calculate the fitness function value of each particle.

[0190] ③ If the current fitness function value is better than the historical best value, then update pBest.

[0191] iv. If the current fitness function value is better than the global historical optimal value, update gBest.

[0192] v. Update the velocity and position of each particle i in the dth dimension respectively according to the following formula:

[0193]

[0194]

[0195] where ω is the inertia weight, generally initialized as 0.9, and then decreased to 0.4 along with the evolution process, c1 and c2 are acceleration coefficients, generally taking fixed value 2.0, and are two random numbers in [0, 1]. It is worth noting that in the updating process, the particle swarm algorithm requires setting a v max to limit the velocity range, generally v max of each dimension can take 10% to 20% of the corresponding dimension. The optimization effect is shown in Figure 4(e). Figure 4(e) is an optimization diagram of a single charging pile position, while Figure 4(a) shows all the task points and charging pile positions.

[0196] Step four: Optimize the L-BPNN neural network parameters of the intelligent distribution robot deployment optimization model, the charging pile quantity optimization model and the single-layer charging pile site selection model.

[0197] The part optimized by the genetic algorithm is the weight w ji between the input layer and the hidden layer and the hidden layer bias b j , the weight w kj between the hidden layer and the output layer and the output layer bias b k , and the weight w between the input layer and the output layer. The specific algorithm process is shown in Figure Three .

[0198] Take the neural network as the main function of the genetic algorithm, set the error function of the L-BPNN neural network as the fitness function. The population number is the effective sample number ρ, the fitness of a single individual is τ i , and its selected probability is:

[0199]

[0200] After that, cross and mutation operations are performed to obtain offspring, and finally the optimal individual in each iteration process is retained by calculating the fitness of the individual.

[0201] In order to illustrate the effect of the present application, Fig. 4 shows an effect diagram of the resource deployment of the delivery robot under simulation conditions. It can be seen from the figure that the resource deployment of the present application can realize high-precision efficiency planning of the charging scheduling of the delivery robot, so that the intelligent delivery robot resource can achieve optimal deployment under different working conditions.

[0202] The above specific description further details the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A dynamic resource planning method for an intelligent delivery robot system based on multi-objective optimization, characterized in that: Includes the following steps, Step 1: Analyze, preprocess, and filter the deployment data of intelligent delivery robots to construct a dataset for optimizing the deployment of intelligent delivery robots, and then construct the deployment optimization model of intelligent delivery robots using an L-BPNN neural network; Step 2: The particle swarm optimization algorithm is used to solve the deployment model optimization problem of intelligent delivery robots. The particle swarm optimization algorithm effectively solves the multi-objective coupled optimization problem of intelligent delivery robot deployment and improves the deployment accuracy of intelligent delivery robots; a predictive model of feature variable changes is established to obtain the deployment changes of intelligent delivery robots in the future. Step 3: Construct a charging pile quantity optimization model using an L-PBNN neural network, and solve the optimization problem using the momentum gradient descent method to determine the number of charging piles. Under the constraints of the charging pile deployment environment and task delivery timeliness, with the goal of maximizing resource utilization, construct a single-layer charging pile location model using an L-PBNN neural network, and use the particle swarm optimization algorithm to predict the optimal location of charging piles. This can prevent insufficient power for intelligent delivery robots and avoid the problem of redundant and idle charging piles.

2. The resource dynamic planning method for an intelligent delivery robot system based on multi-objective optimization as described in claim 1, characterized in that: The process also includes step four, which involves using a genetic algorithm to iteratively optimize the L-BPNN neural network parameters of the intelligent delivery robot deployment optimization model, the charging pile quantity optimization model, and the single-layer charging pile location model based on the actual needs of each department for delivery robots. This improves the prediction accuracy and efficiency of the intelligent delivery robot deployment optimization model, i.e., it achieves high-precision and high-efficiency resource planning for delivery robots based on the particle swarm optimization algorithm.

3. The resource dynamic planning method for an intelligent delivery robot system based on multi-objective optimization as described in claim 1 or 2, characterized in that: The implementation method for step one is as follows: Step 1.1: Analyze the deployment data of the intelligent delivery robot and select sample data. Perform data preprocessing and screening on the sample data to obtain valid sample data that does not include missing data or anomalies. This effectively filters out valid sample data from a large amount of data and minimizes the impact of invalid sample data on the establishment of the deployment optimization model of the intelligent delivery robot. The deployment data includes historical business volume at different times and corresponding delivery demand, delivery volume, average delivery time, location distribution of delivery robots, and number of delivery robots. For data under different working conditions, the selected historical sample data is cleaned through data preprocessing to address data problems in historical order samples caused by other reasons. The data preprocessing methods include interpolation, fitting, and removal. Other reasons include data transfer and human error in recording. Data problems include missing data and data anomalies. Step 1.2: For the valid sample data obtained in Step 1.1, a correlation analysis is performed on the valid sample data using the Pearson chi-square test to screen out the feature data variables that can affect the deployment results of intelligent delivery robots, and a dataset is constructed; feature data variables that do not affect the deployment effect of intelligent delivery robots are effectively removed, accelerating the learning speed of the neural network and improving the efficiency of building an optimization model for the deployment of intelligent delivery robots; the feature data variables include historical business volume in different periods and corresponding delivery demand, delivery volume, average delivery time, location distribution of delivery robots, and number of delivery robots; Pearson's chi-square test was used to screen for characteristic data variables that could influence the deployment results; the "null hypothesis" was that historical traffic volume in different periods was statistically independent of the deployment results; then a contingency table was compiled; the contingency table has r rows and c columns, with a theoretical frequency E. i,j as follows: Where N is the sample size, then the statistic χ² is calculated. 2 : Based on the set confidence level, find the chi-square distribution critical value with degrees of freedom df = rc-1, and compare it with the statistical value χ². 2 If the statistical value is large, the null hypothesis cannot be rejected, meaning that the characteristic data variable is statistically independent of the deployment result. Similarly, the statistical independence of the deployment data of all intelligent delivery robots from the deployment result can be obtained, and then the characteristic data variables that can affect the deployment result can be screened out. The deployment data includes variables such as historical business volume in different periods and corresponding delivery demand, delivery volume, average delivery time, location distribution of delivery robots, and number of delivery robots. A dataset for building an intelligent delivery robot deployment optimization model is constructed based on the feature data variables that affect the deployment results; feature data variables that do not affect the deployment effect of intelligent delivery robots are effectively removed, the learning speed of neural networks is accelerated, and the efficiency of building an intelligent delivery robot deployment optimization model is improved. Step 1.3: For the dataset obtained in Step 1.2, construct a deployment optimization model for intelligent delivery robots using an L-BPNN neural network; to make the deployment optimization model for intelligent delivery robots more closely match the actual working conditions of intelligent delivery robot deployment; An optimization model for the deployment of intelligent delivery robots was established using deep learning methods. The BPNN is improved in terms of network topology to obtain a BPNN with linear mapping relationship, namely L-BPNN; the L-BPNN neural network with linear mapping relationship is selected to construct the deployment optimization model of the intelligent delivery robot, so that the deployment optimization model of the intelligent delivery robot is more in line with the actual working conditions of the intelligent delivery robot deployment. The L-BPNN neural network structure adds a direct connection between the input and output layers to the BPNN, reflecting the linear and nonlinear relationship between the input and output; where: x1, x2, ..., x l-1 ,x l The input information is represented by h1, h2, ... h. m-1 ,h m Represents the hidden layers, y1, y2, ... y n-1 ,y n Indicates the output value, w ji w represents the weights between the input layer and the hidden layer. kj This represents the weights between the hidden layer and the output layer. This represents the weights between the input and output layers; therefore, the expression for calculating the output of the L-BPNN hidden layer remains: In the formula, g(x) represents the activation function of the hidden layer, commonly the sigmoid function; m represents the number of neurons in the hidden layer, and b... j Indicates the hidden layer bias; The expression for the corresponding output layer is: In the formula, n represents the number of neurons in the output layer, b k Indicates the output layer bias; The deployment optimization model prediction framework for intelligent delivery robots based on L-BPNN neural networks also includes four parts: the establishment of network topology and the initialization of related parameters, the learning and training of L-BPNN, obtaining the optimal network parameters, and the prediction of the deployment optimization model for intelligent delivery robots. After establishing an optimization model for delivery robot deployment using an L-BPNN neural network, a constraint model and optimization metrics based on real-time information of feature variables and delivery time and efficiency as costs are established. The feature variables include historical business volume at different times and corresponding delivery demand at various locations, delivery volume, average delivery time, location distribution of delivery robots, and the number of delivery robots. The optimization sub-problems are defined as follows: and Among them, J i (x i Let T be the delivery efficiency cost of the i-th delivery robot deployed. i (x i M represents the delivery time cost under the deployment. i m represents the current task volume. i For maximum load; t i D represents the duration of a single delivery. i E represents the total distance of a single delivery. i This is the current battery level.

4. The resource dynamic planning method for an intelligent delivery robot system based on multi-objective optimization as described in claim 3, characterized in that: The second step is implemented as follows: Step 2.1: Use the particle swarm optimization algorithm to solve the deployment optimization model of the intelligent delivery robot obtained in Step 1, so as to improve the deployment accuracy of the intelligent delivery robot; Particle swarm optimization algorithm randomly generates a certain number of particles as effective solutions to the problem search space, then performs an iterative search, determines the fitness value of the particles through the fitness function corresponding to the problem, and obtains the optimization result; The core of the particle swarm optimization algorithm is to utilize the information sharing among individuals in the swarm so that the motion of the entire swarm evolves from disorder to order in the problem solution space, thereby obtaining the optimal solution to the problem. The specific process is as follows: ① Initialize all particles, that is, assign values ​​to their velocity and position, and set the historical best pBest of each individual to the current position, and the best individual in the group as the current gBset; ② In each generation of evolution, calculate the fitness function value of each particle; ③ If the current fitness function value is better than the historical best value, then update pBest; ④ If the current fitness function value is better than the global historical optimum, then update gBset; ⑤ Update the velocity and position of each particle i in the d-th dimension according to the following formulas: In the formula, ω is the inertia weight, and c1 and c2 are acceleration coefficients. and They are two random numbers in the range [0,1]. Step 2.2: Using deep learning modeling methods, an L-BPNN neural network with linear mapping relationship is selected to construct a feature variable change prediction model to predict the deployment changes of intelligent delivery robots in the future. The prediction results are used as prior knowledge, and the established delivery robot deployment optimization model is used to obtain the deployment changes of delivery robots in each department in the future, thereby improving the robustness of the resource dynamic planning method. The L-BPNN neural network structure adds a direct connection between the input and output layers to the BPNN, reflecting the linear and nonlinear relationship between the input and output; where x1, x2, ..., x l-1 ,x l The input information is represented by h1, h2, ... h. m-1 ,h m Represents the hidden layers, y1, y2, ... y n-1 ,y n Indicates the output value, w ji w represents the weights between the input layer and the hidden layer. kj This represents the weights between the hidden layer and the output layer. This represents the weights between the input and output layers; therefore, the expression for calculating the output of the L-BPNN hidden layer remains: In the formula, g(x) represents the activation function of the hidden layer, commonly the sigmoid function; m represents the number of neurons in the hidden layer, and b... j Indicates the hidden layer bias; The expression for the corresponding output layer is: In the formula, n represents the number of neurons in the output layer, b k Indicates the output layer bias; This predictive model is used to predict changes in characteristic variables over a future period, and then uses these changes as prior knowledge to obtain the deployment changes of delivery robots in various departments over a future period.

5. The resource dynamic planning method for an intelligent delivery robot system based on multi-objective optimization as described in claim 4, characterized in that: The method for implementing step three is as follows: Step 3.1: Construct a charging pile quantity optimization model using an L-PBNN neural network to make the charging pile quantity optimization model of the intelligent delivery robot more closely match the actual working conditions of the intelligent delivery robot's charging scheduling. The L-BPNN neural network structure adds a direct connection between the input and output layers to the BPNN, reflecting the linear and nonlinear relationship between the input and output; where: x1, x2, ..., x l-1 x l The input information is represented by h1, h2, ... h. m-1 h m Represents the hidden layers, y1, y2, ... y n-1 y n Indicates the output value, w ji w represents the weights between the input layer and the hidden layer. kj This represents the weights between the hidden layer and the output layer. This represents the weights between the input and output layers; therefore, the expression for calculating the output of the L-BPNN hidden layer remains: In the formula, g(x) represents the activation function of the hidden layer, commonly the sigmoid function; m represents the number of neurons in the hidden layer, and b... j Indicates the hidden layer bias; The expression for the corresponding output layer is: In the formula, n represents the number of neurons in the output layer, b k Indicates the output layer bias; The optimization problem is defined as follows: In the formula k i Indicates the number of charging stations; Step 3.2: Use the momentum gradient descent method to solve the charging pile number optimization model obtained in Step 3.1 to determine the number of charging piles; v i =γv i +ηΔL(θ) (14) i i =θ i-1 -v i (15) In the formula v i η is the current velocity, γ is the momentum parameter, and η is the learning rate; Step 3.3: Construct a single-layer charging pile location model using an L-PBNN neural network to make the single-layer charging pile location model of the intelligent delivery robot more closely match the actual working conditions of the intelligent delivery robot's charging scheduling. The L-BPNN neural network structure adds a direct connection between the input and output layers to the BPNN, reflecting the linear and nonlinear relationship between the input and output; where: x1, x2, ..., x l-1 x l The input information is represented by h1, h2, ... h. m-1 h m Represents the hidden layers, y1, y2, ... y n-1 y n Indicates the output value, w ji w represents the weights between the input layer and the hidden layer. kj This represents the weights between the hidden layer and the output layer. This represents the weights between the input and output layers; therefore, the expression for calculating the output of the L-BPNN hidden layer remains: In the formula, g(x) represents the activation function of the hidden layer, commonly the sigmoid function; m represents the number of neurons in the hidden layer, and b... j Indicates the hidden layer bias; The expression for the corresponding output layer is: In the formula, n represents the number of neurons in the output layer, b k Indicates the output layer bias; The optimization problem is defined as follows: In the formula α i β represents the number of delivery addresses covered by the i-th charging station. i This represents the number of initial locations of delivery robots covered by the i-th charging station; Step 3.4: Use the particle swarm optimization algorithm to solve the single-layer charging pile site selection model obtained in Step 3.3 to determine the deployment location of the charging piles; The specific process is as follows: ① Initialize all particles, that is, assign values ​​to their velocity and position, and set the historical best pBest of each individual to the current position, and the best individual in the group as the current gBest; ② In each generation of evolution, calculate the fitness function value of each particle; ③ If the current fitness function value is better than the historical best value, then update pBest; ④ If the current fitness function value is better than the global historical best value, then update gBest; ⑤ Update the velocity and position of each particle i in the d-th dimension according to the following formulas: In the formula, ω is the inertia weight, and c1 and c2 are acceleration coefficients. and They are two random numbers in the range [0, 1].

Images