Warehousing system configuration method fusing neural network and queuing theory

Abstract

The invention relates to a storage system configuration method fusing a neural network and a queuing theory, and the method comprises the following steps: S1, constructing a complex queuing network model, and employing an AMVA algorithm for solving; s2, determining evaluation indexes and related parameters; s3, constructing and preprocessing a data set gamma; s4, constructing a full-connection feedforward type deep neural network comprising three hidden layers and an output layer; s5, training by using the training set gamma t, and recording a mean absolute error MSE and a mean square error MAE of a loss function in the training process so as to observe the difference between a predicted value and a real value; s6, performing model verification by using the test set gamma v, determining whether to increase the size of the data set or adjust the structure of the model according to the generalization ability of the model, and retraining the model; and S7, inputting the configuration possible condition into the deep neural network to obtain a corresponding configuration scheme. The working process of the robot is fully considered, the evaluation method is close to reality, and the evaluation result is more accurate.

Description

Technical Field

[0001] The invention relates to the field of management science and engineering, and in particular to a storage system configuration method integrating neural network and queuing theory. Background Art

[0002] The configuration design of a warehousing system needs to balance the storage space requirements and storage throughput, and multiple parameters need to be considered to ensure the effectiveness and efficiency of the system. The traditional design method of system configuration often modifies one of the parameters to achieve the best, ignoring the direct correlation between the various parameters. Moreover, traditionally, the comparison of different systems is often based on simulation experiments. However, a major disadvantage of simulation experiments is that they are very time-consuming. This disadvantage becomes more significant if several optimization problems need to be solved in each simulation run. Therefore, it is not feasible to test all possible configurations of a given order structure during the planning stage of a new warehouse, which will cause delays in the planning process.

[0003] In order to improve the operating efficiency of the system, it is necessary to improve the performance of the system. The performance of the system is determined by three factors: the throughput of the system; the operating cost of the system; and the energy consumption of the system. These factors directly depend on the configuration method of the system. The present invention helps warehouse decision makers to judge the selection of box turning strategies, warehouse aspect ratios, and sub-layer heights in different situations. Traditional methods and mathematical models are time-consuming and lack objectivity in logic. Therefore, it is necessary to adopt a new method to optimize the configuration design of the intelligent warehousing system. Summary of the invention

[0004] In view of the defects of the prior art, the purpose of the present invention is to provide a storage system configuration method that integrates neural networks and queuing theory, combining queuing theory with deep learning. Queuing theory is used to model the system specified by the user and generate a data set; the data set is used to train a deep neural network to compare all possible system configurations that meet the user's requirements. By constructing a queuing network model and a deep neural network, throughput and energy consumption are used as key performance indicators to find the system configuration with the maximum throughput under limited cost, compare different system configurations, and design the optimal system configuration according to user needs.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: a storage system configuration method integrating neural network and queuing theory, comprising the following steps: S1: Construct a complex queuing network model and solve it using the AMVA algorithm; S2: Determine the evaluation indicators and related parameters; S3: construct dataset γ and preprocess it; S4: Build a fully connected feedforward deep neural network with three hidden layers and one output layer; S5: Use the training set γt for training, and record the mean absolute error (MSE) and mean square error (MAE) of the loss function during the training process to observe the difference between the predicted value and the true value; S6: Use the test set γv to verify the model, and decide whether to increase the data set size or adjust the model structure and retrain the model based on the generalization ability of the model; S7: Input the possible configuration situations into the deep neural network to obtain the corresponding configuration solutions.

[0006] Furthermore, in step S1, the algorithm solution includes the following steps: S11: Initialization: For all service stations, m∈M; p m (0|0)=1 Q m (0) = 0 L m (0) = 0 _L m (0) = 0 In the formula, p m (i|r) represents the probability that there are i transport robots in node m when there are r transport robots in the system, i=1,...,r,r=1,...,R,m∈M. Q m (r) represents the probability that all transport robots in node m are busy when there are r transport robots in the system, r = 1, ..., R, m∈M. L m (r) represents the expected queue length of the transport robots in node m (including the transport robots in service) when there are r transport robots in the system, r = 1, ..., R, m∈M; _L m (0) represents the expected queue length of the transport robots in node m when there are r transport robots in the system, r = 1, ..., R, m ∈ M; S12: Iteration. The iteration includes the following steps: S121: When orders arrive at a node, they must wait for the orders being served to be completed before they can be served; the time an order stays at the node can be divided into three parts: the first part is the average remaining service time of the orders being served; the second part is the waiting time required for the orders currently in the queue (excluding the orders being served); the third part is the average time an order takes to be served at the node; when there are r transport robots in the system, the throughput time ET of node m m (r) can be expressed as: In the formula, c m Represents the number of parallel servers of node m, m∈M. ETm (r) represents the throughput time of node m when there are r transport robots in the system, m∈M; ES rem,m represents the average remaining time to leave the task currently being served at node m, m∈M; EST m Represents the first-order moment of the service time of node m, m∈M; Represents the second-order moment of the service time of node m, m∈M. S122: Calculate the throughput. The calculation formula is as follows: In the formula, v m Represents the access rate of node m, m∈M. S123: Calculate the marginal probability that service node m has one robot when the system has r robots. The calculation formula is as follows: Where Q m The value of (r) can be calculated as: Where TH(r) represents the throughput when there are r transport robots in the system, r=1,...,R. S124: Calculate the queue length. The calculation formula is as follows: S125: Calculate the energy consumption of the immediate box turning process and the delayed box turning process, and the calculation formula is as follows: In the immediate box search, In delayed box turning, E e =E m +E R2 +E t +E p +E b Where, T e It represents the expected time that the robot consumes energy in one cycle; E r represents the energy consumed by the robot moving without acquiring the bin; E′ r represents the energy consumed by the robot moving to obtain the material box; E a Indicates the energy consumed by the robot arm moving without obtaining the material box; E′ a Indicates the energy consumed by the robot arm to move the material box; E uIndicates the time when the robot arm loads / unloads goods; E d Indicates the energy consumed by the conveyor belt. E w Indicates the energy consumed by picking at the workstation; E e Indicates the energy consumption in one cycle; T m It represents the time it takes for the robot to move from its standby point to the picking position; E m It represents the energy consumption of the robot moving from the standby point to the picking position; E R1 E represents the energy consumption of immediately turning the box; R2 It represents the energy consumption of delayed box turning. t It indicates the time it takes for the robot to transport the target bin from the picking location to the delivery point; E t It represents the energy consumption of the robot in the process of transporting the target bin from the picking location to the delivery point; E e1 Indicates the energy consumption in one cycle of immediate box flipping; E e2 represents the energy consumption in one cycle of delayed box turning; T w Indicates the time needed to be spent at the workstation; h b Indicates the height of the material box. l Indicates the time for the elevator to load and unload the material box; H k Indicates the height of the kth layer of the storage stack.

[0007] Furthermore, in step S2, the process of determining the evaluation index and related parameters is as follows: Select closed-loop throughput and cost as the evaluation indicators of the system; find system indicators related to throughput and cost, including maximum storage capacity, length, width and height limits, sub-layer height, number of robots, number of charging piles, number of workbenches, annual site rent, robot price and service life, and stack price and service life.

[0008] Further, step S3 includes the following steps: S31: Generate a data set: Take the minimum storage capacity, length, width and height limit, sub-layer height, number of robots, number of charging piles, and number of workbenches as feature values. If the user-given conditions are met, the corresponding throughput is solved using a queuing network; otherwise, the throughput is 0; use the throughput as the label value, generate 60,000 data, and construct a data set γ; S32: Data set preprocessing, including the following steps: S321: Normalize the eigenvalues, that is, S322: Dataset division: The data set is divided into two parts: γt and γv, where γt accounts for 70% of the data set as the training set; γv accounts for 30% of the data set as the test set.

[0009] Further, step S4 includes the following steps: S41: Construct a fully connected feedforward neural network with three hidden layers and one output layer. The hidden layer passes data to the output layer through the weights between the hidden layer and the output layer. The neurons in the output layer also process data and pass the results to the outside world. The hidden layer uses the ReLU activation function, and the mathematical expression is f(x)=max(0,x). S42: MSE is used as the loss function, and MAE is used as the indicator to be monitored during training. The calculation formula is as follows:

[0010] In step S5, according to the training results, the number of neurons in each layer of the model, training epochs and batchsize are adjusted to minimize MSE and MAE.

[0011] Furthermore, in step S6, the process of model verification is as follows: S61: Test set evaluation: The generalization ability of the constructed neural network is tested using the test set γv, and MSE is used as the evaluation criterion; S62: Model tuning: If the model fails to demonstrate good generalization ability, repeat steps S3 to S5, increase the dataset size or adjust the model structure, and retrain the model; If the results are good, the model can be used directly to predict the system throughput and find the optimal system configuration.

[0012] The beneficial effects of this application are: 1. This application fully considers the working process of the robot, the evaluation method is close to reality, and the evaluation result is more accurate.

[0013] 2. This application develops a model that combines queuing networks and deep learning to study the throughput of smart warehousing systems with different configuration methods, and also calculates the energy consumption.

[0014] 3. This application can help warehouse decision makers choose robot box turning strategies under different floor areas, that is, immediate box turning or delayed box turning.

[0015] 4. This application helps warehouse decision makers select the best system aspect ratio given the system area.

[0016] 5. This application helps warehouse decision makers to select the optimal number of sub-layers when the total stack height is limited or unlimited. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 To immediately turn over the box system model diagram;

[0018] Figure 2This is a model diagram of the delayed box turning system;

[0019] Figure 3 Comparison of throughput and energy consumption between immediate and delayed box turning under different floor space conditions;

[0020] Figure 4 This is a comparison chart of energy consumption between delayed box turning strategy and immediate box turning strategy when the floor space changes;

[0021] Figure 5 The throughput and energy consumption diagrams of delayed box turning strategy and immediate box turning strategy under different aspect ratios;

[0022] Figure 6 The throughput and energy consumption diagram of different numbers of sublayers when the total height is not restricted;

[0023] Figure 7 The throughput and energy consumption diagram of different numbers of sub-layers when the total stack height is limited to 15m. DETAILED DESCRIPTION

[0024] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] This embodiment provides a storage system configuration method integrating neural network and queuing theory, comprising the following steps: S1: Construct a complex queuing network model and solve it using the AMVA algorithm; S2: Determine the evaluation indicators and related parameters; S3: construct dataset γ and preprocess it; S4: Build a fully connected feedforward deep neural network with three hidden layers and one output layer; S5: Use the training set γt for training, and record the loss function mean absolute error MSE (MeanSquare Error) and mean square error MAE (Mean Absolute Error) during the training process to observe the difference between the predicted value and the true value; S6: Use the test set γv to verify the model, and decide whether to increase the data set size or adjust the model structure and retrain the model based on the generalization ability of the model; S7: Input the possible configuration situations into the deep neural network to obtain the corresponding configuration solutions.

[0026] In step S1, the modeling and solving process is as follows:

[0027] 1. Scenario description:

[0028] System composition layout:

[0029] ① Shelves: The shelves include a robot motion layer and a storage layer. Each storage layer includes multiple storage sub-layers, and a robot motion layer is set above each storage layer.

[0030] ②Workbench: distributed in the four corners of the system.

[0031] ③Robot: The robot is equipped with a lifting device that can move horizontally and complete the picking of pallets in the vertical direction.

[0032] ④ Charging station: A charging station located in the middle of both sides of the long side of the stack and extending outward (located outside the stack).

[0033] 2. Establish a complex queuing network model:

[0034] The present invention chooses the complex queuing network model because it can be used to simulate the throughput under different system configurations. The present invention analyzes in detail the differences and advantages and disadvantages of immediate box turnover and delayed box turnover. The complex queuing network model is chosen because it can calculate the throughput under a given number of robots to compare the impact of the layout on system performance. After the order is assigned to the robot, the robot moves along the path at the top of the stack to the pickup position (μ m ), assuming that there is no blocking in the process, and using infinite servers to represent the process. Using shared stacks and random storage, assuming that there is no blocking in the process of obtaining the bin, and also using infinite servers to represent (IR:μ R DR:μ r ,μ b ). The process of the robot from the pickup location to the workstation (μ t ) Similarly, it is represented by an infinite server. When the robot arrives at the workstation, it queues in the buffer and waits for service. The picking process is represented by a server with infinite buffer space (μ w ). Under the immediate box flipping strategy, after the material box is completed at the workstation, the robot checks the battery level. Under the delayed box flipping strategy, after the material box is completed at the workstation, the robot returns to the pickup position to put the blocked material box back to the original stack and then checks the battery level. If the power is below the charging threshold, the robot will charge before continuing the next cycle. The charger serves one robot at a time, so it is modeled as one server.

[0035] 3. Solve using the mean value approximation algorithm (AMVA):

[0036] S11: Initialization:

[0037] For all service stations, m∈M; p m (0|0)=1 Q m (0) = 0 L m (0) = 0 _L m (0) = 0

[0038] In the formula, p m (i|r) represents the probability that there are i transport robots in node m when there are r transport robots in the system, i=1,...,r,r=1,...,R,m∈M. Q m (r) represents the probability that all transport robots in node m are busy when there are r transport robots in the system, r = 1, ..., R, m∈M. L m (r) represents the expected queue length of the transport robots in node m (including the transport robots in service) when there are r transport robots in the system, r = 1, ..., R, m∈M; _L m (0) represents the expected queue length of the transport robots in node m when there are r transport robots in the system, r = 1, ..., R, m ∈ M.

[0039] S12: Iteration, for r=1:R, for all service stations m∈M. The iteration includes the following steps:

[0040] S121: When orders arrive at a node, they must wait for the orders being served to be completed before they can be served; the time an order stays at the node can be divided into three parts: the first part is the average remaining service time of the orders being served; the second part is the waiting time required for the orders currently in the queue (excluding the orders being served); the third part is the average time an order takes to be served at the node; when there are r transport robots in the system, the throughput time ET of node m m (r) can be expressed as:

[0041] In the formula, c m Represents the number of parallel servers of node m, m∈M. ET m (r) represents the throughput time of node m when there are r transport robots in the system, m∈M; ES rem,m represents the average remaining time to leave the task currently being served at node m, m∈M; EST m Represents the first-order moment of the service time of node m, m∈M; Represents the second-order moment of the service time of node m, m∈M.

[0042] S122: Calculate the throughput. The calculation formula is as follows:

[0043]

[0044] In the formula, v m Represents the access rate of node m, m∈M.

[0045] S123: Calculate the marginal probability that service node m has one robot when the system has r robots. The calculation formula is as follows:

[0046]

[0047] Where Q m The value of (r) can be calculated as:

[0048]

[0049] Where TH(r) represents the throughput when there are r transport robots in the system, r=1,...,R.

[0050] S124: Calculate the queue length. The calculation formula is as follows:

[0051]

[0052] S125: Calculate the energy consumption of the immediate box turning process and the delayed box turning process, and the calculation formula is as follows:

[0053] In the immediate box search,

[0054] E m =T m *E r

[0055]

[0056] E t =T t *E r '

[0057] E p =T w *E w

[0058] E e1 =E m +E R1 +E t +E p

[0059] In delayed box turning,

[0060]

[0061] E e2 =E m +ER2 +E t +E p +E b

[0062] Where, T e It represents the expected time that the robot consumes energy in one cycle; E r represents the energy consumed by the robot moving without acquiring the bin; E′ r represents the energy consumed by the robot moving to obtain the material box; E a Indicates the energy consumed by the robot arm moving without obtaining the material box; E′ a Indicates the energy consumed by the robot arm to move the material box; E u Indicates the time when the robot arm loads / unloads goods; E d Indicates the energy consumed by the conveyor belt. E w Indicates the energy consumed by picking at the workstation; E e Indicates the energy consumption in one cycle; T m It represents the time it takes for the robot to move from its standby point to the picking position; E m It represents the energy consumption of the robot moving from the standby point to the picking position; E R1 E represents the energy consumption of immediately turning the box; R2 It represents the energy consumption of delayed box turning. t It indicates the time it takes for the robot to transport the target bin from the picking location to the delivery point; E t It represents the energy consumption of the robot in the process of transporting the target bin from the picking location to the delivery point; E e1 Indicates the energy consumption in one cycle of immediate box flipping; E e2 represents the energy consumption in one cycle of delayed box turning; T w Indicates the time needed to be spent at the workstation; h b Indicates the height of the material box; t l Indicates the time for the elevator to load and unload the material box; H k Indicates the height of the kth layer of the storage stack.

[0063] In step S2, the evaluation index and related parameters are determined as follows: closed-loop throughput (TH) and cost (C) are selected as the evaluation indexes of the system. System indicators related to throughput and cost are found, including maximum storage capacity, length, width and height limits, sub-layer height, number of robots, number of charging piles, number of workbenches, annual site rent, robot price and service life, and stack price and service life.

[0064] Step S3 includes the following steps: S31: Generate a data set: Take the minimum storage capacity, length, width and height limit, sub-layer height, number of robots, number of charging piles, and number of workbenches as feature values. If the user-given conditions are met, the corresponding throughput is solved using a queuing network; otherwise, the throughput is 0; use the throughput as the label value, generate 60,000 data, and construct a data set γ; S32: Data set preprocessing, including the following steps: S321: Normalize the eigenvalues, that is, S322: Dataset division: The data set is divided into two parts: γt and γv, where γt accounts for 70% of the data set as the training set; γv accounts for 30% of the data set as the test set.

[0065] Step S4 includes the following steps: S41: Construct a fully connected feedforward neural network with three hidden layers and one output layer. The hidden layer passes data to the output layer through the weights between the hidden layer and the output layer. The neurons in the output layer also process data and pass the results to the outside world. The hidden layer uses the ReLU activation function, and the mathematical expression is f(x)=max(0,x). S42: MSE is used as the loss function, and MAE is used as the indicator to be monitored during training. The calculation formula is as follows:

[0066] In step S5, according to the training results, the number of neurons in each layer of the model, training epochs and batchsize are adjusted to minimize MSE and MAE.

[0067] In step S6, the process of model verification is as follows: S61: Test set evaluation: The generalization ability of the constructed neural network is tested using the test set γv, and MSE is used as the evaluation criterion; S62: Model tuning If the model fails to demonstrate good generalization ability, repeat steps S3 to S5, increase the data set size or adjust the model structure, and retrain the model; If the results are good, the model can be used directly to predict the system throughput and find the optimal system configuration.

[0068] In step S7, the process of selecting a configuration scheme is as follows: according to user requirements (limits on length, width, height, and maximum cost), all possible configurations are generated and input into the established deep neural network. DNN will help designers select a suitable system configuration. The five schemes with the highest throughput and their corresponding costs are obtained, and decision makers can comprehensively consider and select the most suitable configuration scheme.

[0069] When the system stack occupies a small area, the system throughput of the delayed box flipping strategy is higher than that of the system using the immediate box flipping strategy under the same conditions. As the robot's movement speed increases, the performance advantage of the delayed box flipping strategy becomes more obvious. However, as the system area increases, the performance of the immediate box flipping strategy gradually outperforms the delayed box flipping strategy, and this difference becomes more significant as the area increases. Although speeding up the robot's movement speed can make up for the disadvantages of the delayed box flipping strategy to a certain extent, in general, the performance of the immediate box flipping strategy is better than the delayed box flipping strategy in a large area system. When the area is small, the energy consumption of the delayed box flipping strategy is lower than that of the immediate box flipping strategy, and the system throughput under the two strategies is not much different. As the area gradually increases, the energy consumption of the immediate box flipping strategy gradually decreases from that of the delayed box flipping strategy, and this difference becomes more obvious as the area increases. In addition, the difference in throughput also becomes more significant as the area increases.

[0070] Under different total areas, whether it is the delayed box turning strategy or the immediate box turning strategy, when r (the ratio of the number of boxes on the long side to the number of boxes on the wide side) is around 0.75, the throughput is the largest (that is, the aspect ratio is 1), the system throughput reaches the maximum value, and the energy consumption reaches the minimum value. As r deviates from 0.75, the system throughput under both strategies gradually decreases and the energy consumption increases. Therefore, r = 0.75 can be used as the optimal aspect ratio of the system to optimize system performance.

[0071] The throughput increases with the reduction of sub-layers, but the fewer layers, the better. If the number of layers is too small, the throughput will also decrease. However, energy consumption decreases with the reduction of sub-layers. Therefore, in order to increase throughput and reduce energy consumption, the number of sub-layers can be reduced appropriately, but a robot motion layer is required above each layer of material boxes. The reduction of sub-layers means an increase in the number of material box layers and robot motion layers, which also means an increase in the height of the entire storage stack. In reality, the height of the storage stack cannot be increased infinitely. In order to store a sufficient number of goods, its floor space can be increased. When the total stack height is limited, the lower the storage layer, the higher the throughput is not necessarily, and the energy consumption does not gradually decrease. The number of storage layers is low, the number of material box layers and robot motion layers increase, and the height limit results in only an increase in area, and the throughput and energy consumption also change accordingly. Therefore, the layer height should be appropriate.

[0072] Therefore, the present invention concludes that in a large-area system, the performance of the immediate box-turning strategy is better than that of the delayed box-turning strategy. When the floor area is small, the energy consumption of the delayed box-turning strategy is lower than that of the immediate box-turning strategy, and the system throughput under the two strategies is not much different. Under different total areas, whether it is the delayed box-turning strategy or the immediate box-turning strategy, r=0.75 is the optimal aspect ratio of the system. When the total stack height is limited, the lower the storage layer, the higher the throughput is not necessarily, and the energy consumption does not gradually decrease. Therefore, the layer height should be appropriate, and the recommended number of layers is 4 / 5 / 6 layers.

[0073] In summary, this paper proposes a framework to evaluate and compare the configurations of intelligent warehousing systems under two box-turning strategies to study order throughput, system performance, and energy consumption. Finally, this paper determines the optimization conditions of system size and storage layer height under different box-turning strategies for intelligent warehousing systems.

[0074] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A storage system configuration method integrating neural network and queuing theory, characterized in that: The following steps are involved: S1: Construct a complex queuing network model and solve it using the AMVA algorithm; S2: Determine the evaluation indicators and related parameters; S3: construct dataset γ and preprocess it; S4: Build a fully connected feedforward deep neural network with three hidden layers and one output layer; S5: Use the training set γt for training, and record the mean absolute error (MSE) and mean square error (MAE) of the loss function during the training process to observe the difference between the predicted value and the true value; S6: Use the test set γv to verify the model, and decide whether to increase the data set size or adjust the model structure and retrain the model based on the generalization ability of the model; S7: Input the possible configuration situations into the deep neural network to obtain the corresponding configuration solutions.

2. A storage system configuration method integrating neural network and queuing theory according to claim 1, characterized in that: In step S1, the algorithm solution includes the following steps: S11: Initialization: For all service stations, m∈M; p m (0|0)＝1 Q m (0)＝0 L m (0)＝0 _L m (0)＝0 In the formula, p m (i|r) represents the probability that there are i transport robots in node m when there are r transport robots in the system, i=1,...,r,r=1,...,R,m∈M. Q m (r) represents the probability that all transport robots in node m are busy when there are r transport robots in the system, r = 1, ..., R, m∈M. L m (r) represents the expected queue length of the transport robots in node m (including the transport robots in service) when there are r transport robots in the system, r = 1, ..., R, m∈M; _L m (0) represents the expected queue length of the transport robots in node m when there are r transport robots in the system, r = 1, ..., R, m ∈ M; S12: Iteration. The iteration includes the following steps: S121: When orders arrive at a node, they must wait for the orders being served to be completed before they can be served; the time an order stays at the node can be divided into three parts: the first part is the average remaining service time of the orders being served; the second part is the waiting time required for the orders currently in the queue (excluding the orders being served); the third part is the average time an order takes to be served at the node; When there are r transport robots in the system, the throughput time ET of node m is m (r) can be expressed as: In the formula, c m Represents the number of parallel servers of node m, m∈M. ET m (r) represents the throughput time of node m when there are r transport robots in the system, m∈M; ES rem,m represents the average remaining time to leave the task currently being served at node m, m∈M; EST m Represents the first-order moment of the service time of node m, m∈M; Represents the second-order moment of the service time of node m, m∈M. S122: Calculate the throughput. The calculation formula is as follows: In the formula, v m Represents the access rate of node m, m∈M. S123: Calculate the marginal probability that service node m has one robot when the system has r robots. The calculation formula is as follows: Where Q m The value of (r) can be calculated as: Where TH(r) represents the throughput when there are r transport robots in the system, r=1,...,R. S124: Calculate the queue length. The calculation formula is as follows: S125: Calculate the energy consumption of the immediate box turning process and the delayed box turning process, and the calculation formula is as follows: In the immediate box search, AND t =T t *AND' r AND p =T w *AND w AND e1 =And m +E R1 +E t +E p In delayed box turning, AND e2 =And m +E R2 +E t +E p +E b Where, T e It represents the expected time that the robot consumes energy in one cycle; E r represents the energy consumed by the robot moving without acquiring the bin; E′ r represents the energy consumed by the robot moving to obtain the material box; E a Indicates the energy consumed by the robot arm moving without obtaining the material box; E′ a Indicates the energy consumed by the robot arm to move the material box; E u Indicates the time when the robot arm loads / unloads goods; E d Indicates the energy consumed by the conveyor belt. E w Indicates the energy consumed by picking at the workstation; E e Indicates the energy consumption in one cycle; T m It represents the time it takes for the robot to move from its standby point to the picking position; E m It represents the energy consumption of the robot moving from the standby point to the picking position; E R1 E represents the energy consumption of immediately turning the box; R2 It represents the energy consumption of delayed box turning. t It indicates the time it takes for the robot to transport the target bin from the picking location to the delivery point; E t It represents the energy consumption of the robot in the process of transporting the target bin from the picking location to the delivery point; E e1 Indicates the energy consumption in one cycle of immediate box flipping; E e2 represents the energy consumption in one cycle of delayed box turning; T w Indicates the time needed to be spent at the workstation; h b Indicates the height of the material box. l Indicates the time for the elevator to load and unload the material box; H k Indicates the height of the kth layer of the storage stack.

3. The method for configuring a storage system integrating neural network and queuing theory according to claim 1, characterized in that: In step S2, the process of determining the evaluation index and related parameters is as follows: Select closed-loop throughput and cost as the evaluation indicators of the system; find system indicators related to throughput and cost, including maximum storage capacity, length, width and height limits, sub-layer height, number of robots, number of charging piles, number of workbenches, annual site rent, robot price and service life, and stack price and service life.

4. The method for configuring a storage system integrating neural network and queuing theory according to claim 1, characterized in that: Step S3 includes the following steps: S31: Generate a data set: Take the minimum storage capacity, length, width and height limit, sub-layer height, number of robots, number of charging piles, and number of workbenches as feature values. If the user-given conditions are met, the corresponding throughput is solved using a queuing network; otherwise, the throughput is 0; use the throughput as the label value, generate 60,000 data, and construct a data set γ; S32: Data set preprocessing, including the following steps: S321: Normalize the eigenvalues, that is, S322: Dataset division: The data set is divided into two parts: γt and γv, where γt accounts for 70% of the data set as the training set; γv accounts for 30% of the data set as the test set.

5. The method for configuring a storage system integrating neural network and queuing theory according to claim 1, characterized in that: Step S4 includes the following steps: S41: Construct a fully connected feedforward neural network with three hidden layers and one output layer. The hidden layer passes data to the output layer through the weights between the hidden layer and the output layer. The neurons in the output layer also process data and pass the results to the outside world. The hidden layer uses the ReLU activation function, and the mathematical expression is f(x)=max(0,x). S42: MSE is used as the loss function, and MAE is used as the indicator to be monitored during training. The calculation formula is as follows:

6. The method for configuring a storage system integrating neural network and queuing theory according to claim 1, characterized in that: Step S5, comprises the following steps: In step S5, according to the training results, the number of neurons in each layer of the model, the training epochs and the batch size are adjusted to minimize the MSE and MAE.

7. The method for configuring a storage system integrating neural network and queuing theory according to claim 1, characterized in that: In step S6, the process of model verification is as follows: S61: Test set evaluation: The generalization ability of the constructed neural network is tested using the test set γv, and MSE is used as the evaluation criterion; S62: Model tuning If the model fails to demonstrate good generalization ability, repeat steps S3 to S5, increase the data set size or adjust the model structure, and retrain the model; If the results are good, the model can be used directly to predict the system throughput and find the optimal system configuration.

Images