Warehouse management method based on digital twin

Abstract

The present invention is a warehouse management method based on digital twins, belonging to the field of intelligent warehousing technology. Based on model construction and data mapping, the method combines task data and warehouse basic source data, and realizes time prediction of various tasks through relevant prediction algorithms, provides data basis for task scheduling and planning, and assists management personnel in decision-making. The method of the present invention is based on the warehouse site environment, establishes a simulated virtual model of an automated warehouse space, and considers information such as the location of the automatic conveying platform, the location of the goods, the entry port, the exit port, the entry road and the exit road; the model has good reconfigurability and can be flexibly set according to specific needs. At the same time, the roads are planned and defined, and through intelligent scheduling and planning simulation, vehicle collisions and path conflicts are avoided, route optimization is achieved, and work efficiency is improved.

Description

Technical Field

[0001] The present application relates to the technical field of warehouse management and control, and in particular to a warehouse management method based on digital twins. Background Art

[0002] Today's warehouse operations and warehouse management are very complicated. Manual input and operation are not only time-consuming and labor-intensive, but also prone to errors. At the same time, there are problems such as difficulty in real-time control of warehouse cargo information, difficulty in real-time monitoring of the status of storage equipment during operation, and low warehouse management efficiency. Based on digital twin technology, the present invention constructs an intelligent warehouse management system and method, and improves warehouse operation efficiency and comprehensive management level through technologies such as visual control, transportation time prediction, intelligent scheduling and path planning. Summary of the invention

[0003] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology and provide a new warehouse management method based on digital twins. Through intelligent scheduling and planning simulation, this method can avoid vehicle collisions and path conflicts, achieve route optimization, and improve work efficiency.

[0004] The technical problem to be solved by the present invention is achieved through the following technical solutions. The present invention is a warehouse management method based on digital twins, which is characterized by:

[0005] Construct digital twins of warehouse physical entities: The construction objects are mainly automatic conveying platforms, shelves, goods and other physical equipment in the warehouse; through 3D modeling of various construction objects, combined with large-screen display, the twin display of the digital twin warehouse is completed;

[0006] Establish a mapping from the physical warehouse to the virtual warehouse: mainly including physical entities, virtual entities, services, twin data, and mapping connections between various components; the data-driven approach can use the system's historical and real-time operation data to update and modify the physical model and display the real-time operation status of the warehouse;

[0007] Based on model construction and data mapping, combined with task data and warehouse basic source data, through relevant prediction algorithms, time prediction of various tasks is achieved, providing a data basis for task scheduling and planning, and assisting managers in decision-making;

[0008] Based on the warehouse environment, a simulated virtual model of an automated warehouse space is established, taking into account the location of the automatic conveying platform, cargo location, cargo entry, cargo exit, cargo entry road, and cargo exit road information; at the same time, the roads are planned and defined, and vehicle collisions and path conflicts are avoided through intelligent scheduling and planning simulation.

[0009] The warehouse management method based on digital twin described in the present invention has a further preferred technical solution: predicting the demand time for goods to be shipped out of the warehouse. The demand forecast is at the bottom layer in the entire warehouse management system and plays a supporting role, supporting multiple decision optimization systems at the upper layer. These decision optimization systems use accurate prediction data combined with operations research technology to make the best decision and provide the results to the higher-level business execution system.

[0010] A warehouse management method based on digital twins described in the present invention has a further preferred technical solution: a method for predicting the time required for goods to be shipped out of the warehouse is: obtaining the required business data from an external data source, then processing and cleaning the basic data, and then processing and analyzing the data through time series or machine learning or other artificial intelligence technologies, and finally calculating the prediction results and pushing them to downstream modules for use through various channels;

[0011] The overall architecture of the cargo outbound demand forecasting module is from top to bottom: data source input layer, basic data processing layer, core business layer, data output layer and downstream execution module;

[0012] (1) Data source input layer: The data source of warehouse management includes most of the required business data, including cargo information, storage information, and historical outbound information. Most of the cargo transportation plan data comes from operators entering it through the storage and transportation system.

[0013] (2) Core business layer: This layer is the core part, mainly including feature construction, prediction algorithm and prediction result processing; it can also include multiple non-intersecting business lines;

[0014] Feature construction is to convert the previously cleaned basic data into feature data in a standard format through further processing, and provide it to the subsequent algorithm model for use; the core algorithm mainly uses time series analysis, machine learning or other artificial intelligence technologies to predict the demand for goods outbound; the processing of the prediction results cannot meet the downstream modules in terms of format and some special requirements, so it is also necessary to process it according to the actual situation, such as adding additional information such as standard deviation and goods outbound identification;

[0015] The core technology of cargo outbound demand forecasting is mainly divided into the basic layer, framework layer, tool layer and algorithm layer. The basic layer uses HDFS for data storage, Yarn for resource scheduling, and develops related task scheduling modules. The framework layer is mainly based on Spark RDD, Spark SQL, and Hive. The tool layer directly uses relatively mature and stable algorithms and models, which are encapsulated in third-party Python packages, which are selected from xgboost, numpy, pandas, sklearn, scipy or hyperopt. The algorithm layer is mainly divided into time series and machine learning algorithms.

[0016] The memory inside the RNN is used to process input sequences of any time series; the time series task prediction of cargo transportation mainly includes ARIMA and Holt winters; among them, ARIMA is an autoregressive integrated moving average model, which is mainly used to predict stable sequences such as the number of warehouse cargo outbound orders; Holt winters, also known as the cubic exponential smoothing algorithm, is used to predict outbound cargo with obvious seasonality and trends.

[0017] The warehouse management method based on digital twins described in the present invention has a further preferred technical solution: the algorithm layer is an iterative decision tree algorithm, GBDT, which is composed of multiple decision trees. The decision results of all trees are accumulated to make the final decision, and it is used to predict goods with high out-of-stock but no obvious historical rules.

[0018] The digital twin-based warehouse management method described in the present invention has a further preferred technical solution: the general process of cargo outbound demand forecasting based on machine learning algorithm is as follows:

[0019] (1) Feature construction: Determine the main features through data analysis and model experiments, and generate feature data in a standard format through a series of tasks;

[0020] (2) Model selection: Different goods have different characteristics, so different algorithm models will be assigned based on factors such as the historical outbound demand of the goods and the sensitivity of special training days;

[0021] (3) Feature selection: Screening a batch of features to filter out unnecessary features. Different types of goods have different features.

[0022] (4) Sample partitioning: Group the training data into multiple groups of samples. During training, a model file is generated for each group of samples. Generally, goods of the same type are grouped together.

[0023] (5) Model parameters: Select the optimal model parameters. Appropriate parameters will improve the accuracy of the model because different parameter combinations need to be used for model training and prediction.

[0024] (6) Model training: After the features, models, and samples are determined, model training is performed. After training, a model file is generated and stored in HDFS.

[0025] (7) Model prediction: read the model file to perform prediction;

[0026] (8) Multi-model selection: In order to improve the prediction accuracy, multiple algorithm models are used. After the prediction results of each model are output, the best prediction result is selected;

[0027] (9) Abnormal prediction interception: The more complex and difficult to explain algorithms are, the more likely they are to have extremely high prediction values. Such high predictions cannot be explained by combining historical cargo outbound demand data. Therefore, these abnormal values ​​can be intercepted by some rules and replaced with a more conservative value.

[0028] (10) Model evaluation: calculation of prediction accuracy;

[0029] (11) Error analysis: By analyzing the prediction accuracy, the distribution of errors in different dimensions can be obtained to provide a reference for algorithm optimization.

[0030] The digital twin-based warehouse management method described in the present invention has a further preferred technical solution: the method performs intelligent scheduling and path planning design of the automatic conveying platform for multiple goods entering and leaving the warehouse at the same time,

[0031] First, a simulation virtual model of an automated warehouse space is established. The main considerations are the parking positions of vehicles and goods, i.e., empty or occupied shelves, inlets, outlets, inlet roads, and outlet roads. The warehouse model has good reconfigurability, and the length and width of the warehouse space, the number and density of shelves and vehicles, the number and location of inlets and outlets, and the number and density of transport order tasks can be flexibly set according to specific needs.

[0032] In the warehouse space, when multiple automatic conveying platforms are running at the same time, in order to avoid collisions, simplify the operation rules of the automatic conveying platforms, and improve the safety of system operation, the horizontal and vertical roads between the shelf areas are set to be one-way streets;

[0033] For the established warehouse space model, the task is to transport a shelf from its parking position to a certain inlet / outlet position, and then transport it back to the shelf area after the inlet / outlet task is completed;

[0034] The task can be broken down into the following three steps:

[0035] Step 1: The automatic conveying platform moves to the shelf corresponding to the task;

[0036] Step 2: After the automatic conveying platform is filled, it moves to the inlet / outlet;

[0037] Step 3: The automatic conveying platform moves to a suitable empty space to wait for unloading;

[0038] For the first step of the above logistics task, all the automatic transport platforms that may perform the task are evaluated according to an evaluation function, and the most suitable carrier is selected from them. The evaluation function is defined as:

[0039] g n =w*tn1*tn2

[0040] The above formula represents the total cost of the nth automatic conveyor platform to perform this task. tn1 represents the estimated time it takes for the nth automatic conveyor platform to complete the currently running task. w represents the congestion coefficient, which is used to reflect the degree of congestion in the system. Setting w>1 can reflect that the actual time it takes to complete the current task is longer than the estimated time. tn2 represents the time it takes for the nth automatic conveyor platform to run to the shelf required by the task.

[0041] According to the evaluation function, all automatic transport platforms are divided into two types: those that are currently executing the current task and those that are in an idle state. The waiting cost and path cost are calculated respectively, and then all automatic transport platforms are evaluated based on the sum of the two. The automatic transport platform with the smallest total cost gn is selected to undertake the transportation task.

[0042] For the third step, the shelves transported by the automatic conveyor platform should be parked in the free area as soon as possible so that the automatic conveyor platform can continue to perform the next task; therefore, an idle position closest to the starting point is selected for parking; the cost is estimated using Manhattan distance:

[0043] g n =abs(cur.xn.x)+abs(cur.yn.y)

[0044] Where: g n represents the cost of parking at the nth storage location;

[0045] cur.x represents the horizontal coordinate of the current point, and cur.y represents the vertical coordinate of the current point;

[0046] nx represents the horizontal coordinate of the nth storage position, ny represents the vertical coordinate of the nth storage position; abs represents the function of finding the absolute value;

[0047] In the designed warehouse space structure, the motion planning method used for multiple automatic conveying platforms is a modified and improved A* algorithm; the basic process of the A* algorithm is to start from the starting point and selectively expand the nodes according to the estimated cost until the target point is expanded; the key is to choose a suitable evaluation function:

[0048] f(n)=g(n)+h(n)

[0049] Where: f(n) represents the estimated cost from the starting point through node n to the target point, g(n) represents the actual cost from the starting point to node n, and h(n) represents the estimated cost from node n to the target point;

[0050] If the specific positions of all automatic transport platforms at each moment are known in advance, possible road conflicts can be predicted and avoided during path planning. To this end, a "time-space operation map" is set up to record the trajectory of each robot over time, and the planned robot trajectory is recorded to form a "horizontal-vertical-time" three-dimensional map as a reference for trajectory planning. At the same time, on the basis of considering the distance to a certain location, the waiting time cost increased by preventing collisions is further added.

[0051] The path planning method is modified to promptly modify the space-time operation map after each unplanned avoidance occurs, so that the map is consistent with the actual situation.

[0052] The digital twin-based warehouse management method described in the present invention has a further preferred technical solution: in order to avoid possible collisions, an exclusive point is set; at a certain moment, a certain position can only be occupied by one automatic conveying platform alone; other automatic conveying platforms at the same moment must wait if they want to enter this position.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] Based on model construction and data mapping, the method of the present invention combines task data and warehouse basic source data, and realizes time prediction of various tasks through relevant prediction algorithms, provides data basis for task scheduling and planning, and assists management personnel in decision-making. Based on the warehouse site environment, the method of the present invention establishes a simulated virtual model of an automated warehouse space, taking into account information such as the location of the automatic conveying platform, the location of the goods, the entry port, the exit port, the entry road, and the exit road; the model has good reconfigurability and can be flexibly set according to specific needs. At the same time, the roads are planned and defined, and through intelligent scheduling and planning simulation, vehicle collisions and path conflicts are avoided, route optimization is achieved, and work efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a logic diagram of the method of the present invention;

[0056] Figure 2 Develop a demand forecasting flow chart for warehousing tasks;

[0057] Figure 3 This is the business process diagram for intelligent scheduling and route planning. DETAILED DESCRIPTION

[0058] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments, but does not constitute a limitation on the rights of the present invention.

[0059] Reference Figure 1 , a warehouse management method based on digital twins:

[0060] Construct digital twins of warehouse physical entities: The construction objects are mainly automatic conveying platforms, shelves, goods and other physical equipment in the warehouse; through 3D modeling of various construction objects, combined with large-screen display, the twin display of the digital twin warehouse is completed;

[0061] Establish a mapping from the physical warehouse to the virtual warehouse: mainly including physical entities, virtual entities, services, twin data, and mapping connections between various components; the data-driven approach can use the system's historical and real-time operation data to update and modify the physical model and display the real-time operation status of the warehouse;

[0062] Based on model construction and data mapping, combined with task data and warehouse basic source data, through relevant prediction algorithms, time prediction of various tasks is achieved, providing a data basis for task scheduling and planning, and assisting managers in decision-making;

[0063] Based on the warehouse environment, a simulated virtual model of the automated warehouse space is established, taking into account the cargo parking location, cargo entry, cargo exit, cargo entry road, and cargo exit road information of the automatic conveying platform. At the same time, the roads are planned and defined, and vehicle collisions and path conflicts are avoided through intelligent scheduling and planning simulation.

[0064] like Figure 2 As shown, this embodiment predicts the time required for goods to be shipped out of the warehouse. Demand forecasting is at the bottom layer of the entire warehouse management system and plays a supporting role, supporting multiple decision optimization systems at the upper layer. These decision optimization systems use accurate forecasting data combined with operations research technology to make the best decision and provide the results to the upper-level business execution system. The overall architecture of the goods shipping demand forecasting module is from top to bottom: data source input layer, basic data processing layer, core business layer, data output layer and downstream execution module.

[0065] First, the required business data is obtained from external data sources, and then the basic data is processed and cleaned. Finally, the data is processed and analyzed through artificial intelligence technologies such as time series or machine learning. Finally, the prediction results are calculated and pushed to downstream modules for use through various channels.

[0066] Figure 2 The core content of outbound demand forecasting technology is introduced in the paper, which mainly includes the following contents:

[0067] (1) Data source input layer: The data source of warehouse management includes most of the required business data, such as cargo information, storage information, historical outbound information, etc. Most of the cargo transportation plan data comes from operators entering it through the storage and transportation system;

[0068] (2) Core business layer: This layer is the core part, mainly including feature construction, prediction algorithm and prediction result processing; it may also include multiple non-intersecting business lines. Feature construction is to convert the previously cleaned basic data into feature data in a standard format through further processing, and provide it to the subsequent algorithm model for use; the core algorithm mainly uses artificial intelligence technologies such as time series analysis and machine learning to predict the demand for goods outbound, which is the core part of the prediction system; the prediction result processing is that the format and some special requirements cannot meet the downstream modules, so it needs to be processed according to the actual situation, such as adding additional information such as standard deviation and goods outbound identification.

[0069] The core technology of cargo outbound demand forecasting is mainly divided into the basic layer, framework layer, tool layer and algorithm layer. The basic layer uses HDFS for data storage, Yarn for resource scheduling, and develops related task scheduling modules; the framework layer is mainly based on Spark RDD, Spark SQL, and Hive; the tool layer directly uses relatively mature and stable algorithms and models, which are encapsulated in third-party Python packages. The more commonly used packages are xgboost, numpy, pandas, sklearn, scipy, and hyperopt; the algorithm layer is mainly divided into time series and machine learning algorithms, such as iterative decision tree algorithm, GBDT, which consists of multiple decision trees. The decision results of all trees are accumulated to make the final decision. It is used to predict high outbound cargo, but the historical rules are not obvious. In addition, the memory inside RNN is used to process arbitrary time series input sequences, which makes it easier to handle tasks such as time series prediction. The time series task prediction of cargo transportation mainly includes ARIMA and Holt winters. Among them, ARIMA is an autoregressive integrated moving average model, which is mainly used to predict stable series such as the number of warehouse goods shipped out. Holt winters, also known as the triple exponential smoothing algorithm, is used to predict the shipment of goods with obvious seasonality and trend.

[0070] The general process of forecasting cargo outbound demand based on machine learning algorithms is as follows:

[0071] 1) Feature construction: Determine the main features through data analysis and model testing, and generate feature data in a standard format through a series of tasks.

[0072] 2) Model selection: Different goods have different characteristics, so different algorithm models will be assigned first based on factors such as the historical outbound demand of the goods and the sensitivity of special training days.

[0073] 3) Feature selection: Screen a batch of features to filter out unnecessary features. Different types of goods have different features.

[0074] 4) Sample partitioning: Group the training data into multiple groups of samples, and generate a model file for each group of samples during actual training. Generally, the same type of goods are grouped together to consider parallelization and model accuracy.

[0075] 5) Model parameters: Select the optimal model parameters. Appropriate parameters will improve the accuracy of the model because different parameter combinations need to be trained and predicted separately.

[0076] 6) Model training: After the features, models, and samples are determined, model training can be performed. Training often takes a long time. After training, a model file will be generated and stored in HDFS.

[0077] 7) Model prediction: read the model file to perform prediction.

[0078] 8) Multi-model selection: In order to improve prediction accuracy, we may use multiple algorithm models. After the prediction results of each model are output, the system will select the best prediction result through some rules.

[0079] 9) Abnormal prediction interception: The more complex and difficult to explain the algorithm is, the more likely it is that a few abnormally high prediction values ​​will appear. Such high predictions cannot be explained by combining historical cargo outbound demand data. Therefore, some rules can be used to intercept these abnormal values ​​and replace them with a more conservative value.

[0080] 10) Model evaluation: Calculate the prediction accuracy.

[0081] 11) Error analysis: By analyzing the prediction accuracy, the distribution of errors in different dimensions can be obtained to provide a reference for algorithm optimization.

[0082] like Figure 3 As shown, this embodiment performs intelligent scheduling and path planning design of the automatic conveying platform for multiple goods entering and leaving the warehouse at the same time. First, a simulation virtual model of the automated warehouse space is established. The main considerations are the parking position of the vehicle and goods (idle or occupied by the shelf), the entry port, the exit port, the entry road and the exit road. This warehouse model has good reconfigurability. The length and width of the warehouse space, the number and density of shelves and vehicles, the number and position of the entry and exit ports, the number and density of transportation order tasks, etc. can all be flexibly set according to specific needs. In the warehouse space, when multiple automatic conveying platforms are running at the same time, in order to avoid collisions, simplify the operation rules of the automatic conveying platforms, and improve the safety of system operation, the horizontal and vertical roads between the shelf areas are set to be one-way streets. For the established warehouse space model, the form of the task is to transport a certain shelf from the parking position to a certain entry port / exit port position, and then transport it back to the shelf area position after the entry / exit task is completed. The task can be decomposed into the automatic conveying platform moving to the shelf corresponding to the task, the automatic conveying platform moving to the entry / exit port after loading, and the automatic conveying platform moving to a suitable empty space to wait for unloading.

[0083] For the first step of the above-mentioned logistics task, all possible automatic transport platforms that can perform the task are evaluated according to an evaluation function, and the most suitable carrier is selected from them. The evaluation function is defined as:

[0084] g n=w*tn1*tn2

[0085] The above formula represents the total cost of the nth automatic conveying platform to perform this task, and tn1 represents the estimated time it will take for the nth automatic conveying platform to complete the currently running task (if it is currently idle, this item is 0). w represents the congestion coefficient, which is used to reflect the degree of congestion in the system. Setting w>1 can reflect that the actual time consumed to complete the current task is longer than the estimated time. tn2 represents the time it takes for the nth automatic conveying platform to run to the shelf required by the task. According to the evaluation function, all automatic conveying platforms are divided into two types: those that are executing the current task and those that are idle. The waiting cost and path cost are calculated separately, and then all automatic conveying platforms are evaluated based on the sum of the two, and the automatic conveying platform with the smallest total cost gn is selected to undertake the transportation task.

[0086] For the third step, the shelves transported by the automatic conveyor platform should be parked in the free area as soon as possible so that the automatic conveyor platform can continue to perform the next task. Therefore, an idle position closest to the starting point is selected for parking. Use Manhattan distance to estimate the cost:

[0087] g n =abs(cur.xn.x)+abs(cur.yn.y)

[0088] Where: g n Indicates the cost of parking at the nth storage location. cur.x indicates the horizontal coordinate of the current point, and cur.y indicates the vertical coordinate of the current point. nx indicates the horizontal coordinate of the nth storage location, and ny indicates the vertical coordinate of the nth storage location. abs indicates the function for finding the absolute value.

[0089] In the designed warehouse space structure, currently, the motion planning methods used for multiple automatic conveying platforms include artificial potential field method, neural network, fuzzy logic, A* and other algorithms. Among them, the A* algorithm has been widely used and can ensure the optimal solution path. Considering the constraint of one-way operation of roads in the storage space structure, the A* algorithm is corrected and improved. The basic process of the A* algorithm is to start from the starting point and selectively expand the nodes according to the estimated cost until the target point is expanded. The key is to choose a suitable evaluation function:

[0090] f(n)=g(n)+h(n)

[0091] Where: f(n) represents the estimated cost from the starting point through node n to the target point, g(n) represents the actual cost from the starting point to node n, and h(n) represents the estimated cost from node n to the target point.

[0092] If the specific locations of all automatic conveying platforms at each moment are known in advance, possible road conflicts can be predicted and avoided during path planning. To this end, a "time-space operation map" can be set up to record the trajectory of each robot over time, and the planned robot trajectory can be recorded to form a "horizontal-vertical-time" three-dimensional map as a reference for trajectory planning. At the same time, on the basis of considering the distance to a certain location, the waiting time cost increased by preventing collisions is further added. Therefore, conflict delays are taken into account in path planning, which can reduce conflicts in the planned paths.

[0093] Due to the unidirectional nature of the road, the order of planning cannot guarantee the priority during runtime. If the task planned later is inserted in front of the road of the task planned earlier, then the task planned earlier at the back of the road can only wait. This will result in the actual trajectory not being completely consistent with the original plan. In this case, the task can still be completed according to the planned path, but the time may not be in sync. If you continue to plan new tasks based on the time-space operation map with time deviations, errors will accumulate. Although the task can still be completed, the path is not optimal. For this reason, the path planning method is corrected, and the time-space operation map is corrected in time after each unplanned avoidance occurs, so that the map is consistent with the actual situation.

[0094] In addition, in order to avoid possible collisions, exclusive points are set. At a certain time, a certain position can only be occupied by one automatic conveying platform. Other automatic conveying platforms at the same time must wait if they want to enter this position.

[0095] The above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A warehouse management method based on digital twins, characterized by: Construct digital twins of warehouse physical entities: The construction objects are automatic conveying platforms, shelves, goods, and other physical equipment in the warehouse; through 3D modeling of various construction objects, combined with large-screen display, the digital twin warehouse can be displayed in twin form; Establish a mapping from the physical warehouse to the virtual warehouse: including physical entities, virtual entities, services, twin data, and mapping connections between various components; adopt a data-driven approach to use the system's historical and real-time operation data to update and modify the physical model and display the real-time operation status of the warehouse; Based on model construction and data mapping, combined with task data and warehouse basic source data, through relevant prediction algorithms, time prediction of various tasks is achieved, providing a data basis for task scheduling and planning, and assisting managers in decision-making; Based on the warehouse environment, a simulated virtual model of the automated warehouse space is established, taking into account the location of the automatic conveying platform, the location of the goods, the entry and exit ports, the entry and exit roads; at the same time, the roads are planned and defined, and vehicle collisions and path conflicts are avoided through intelligent scheduling and planning simulation; This warehouse management method based on digital twins performs intelligent scheduling of multiple goods entering and leaving the warehouse at the same time and designs the path planning of the automatic transportation platform; First, a simulation virtual model of the automated warehouse space is established; the parking positions of vehicles and goods, i.e., empty or occupied shelves, inlet and outlet, inlet roads and outlet roads are considered; the warehouse model has good reconfigurability, the length and width of the warehouse space, the number and density of shelves and vehicles, the number and location of inlet and outlet, and the number and density of transportation order tasks are considered; In the warehouse space, when multiple automatic conveying platforms are running at the same time, in order to avoid collisions, simplify the operation rules of the automatic conveying platforms, and improve the safety of system operation, the horizontal and vertical roads between the shelf areas are set to be one-way streets; For the established warehouse space model, the task is to transport a shelf from its parking position to a certain inlet / outlet position, and then transport it back to the shelf area after the inlet / outlet task is completed; The task is broken down into the following three steps: Step 1: The automatic conveying platform moves to the shelf corresponding to the task; Step 2: After the automatic conveying platform is filled, it moves to the inlet / outlet; Step 3: The automatic conveying platform moves to a suitable empty space to wait for unloading; For the first step of the above-mentioned logistics task, all possible automatic transport platforms that can perform the task are evaluated according to an evaluation function, and the most suitable carrier is selected from them. The evaluation function is defined as: g n =w*tn1*tn2 The above formula represents the total cost of the nth automatic transport platform to perform this task, and tn1 represents the estimated time it will take for the nth automatic transport platform to complete the currently running task; w represents the congestion coefficient, which is used to reflect the degree of congestion in the system; setting w>1 reflects that the actual time spent on completing the current task is longer than the estimated time; tn2 represents the time it takes for the nth automatic conveying platform to run to the shelf required by the task; According to the evaluation function, all automatic transport platforms are divided into two types: those that are currently executing the current task and those that are in an idle state. The waiting cost and path cost are calculated respectively, and then all automatic transport platforms are evaluated based on the sum of the two. The total cost g is selected. n The smallest automatic conveying platform undertakes the transportation task; For the third step, the shelves transported by the automatic conveyor platform should be parked in the free area as soon as possible so that the automatic conveyor platform can continue to perform the next task; therefore, an idle position closest to the starting point is selected for parking; the cost is estimated using Manhattan distance: g n =abs(cur.x-n.x)+abs(cur.y-n.y) Where: g n represents the cost of parking at the nth storage location; cur.x represents the horizontal coordinate of the current point, and cur.y represents the vertical coordinate of the current point; nx represents the horizontal coordinate of the nth storage position, ny represents the vertical coordinate of the nth storage position; abs represents the function of finding the absolute value; In the designed warehouse space structure, the motion planning method applied to multiple automatic conveying platforms is a modified and improved A* algorithm; the basic process of the A* algorithm is to start from the starting point and selectively expand the nodes according to the estimated cost until the target point is expanded; The key is to choose a suitable evaluation function: f(n)=g(n)+h(n) Where: f(n) represents the estimated cost from the starting point through node n to the target point, g(n) represents the actual cost from the starting point to node n, and h(n) represents the estimated cost from node n to the target point; If the specific locations of all automatic transport platforms at each moment are known in advance, possible road conflicts can be predicted and avoided during path planning. To this end, a "space-time operation map" is set up to record the trajectory of each robot over time, and the planned robot trajectory is recorded to form a "horizontal-vertical-time" three-dimensional map as a reference for trajectory planning. At the same time, on the basis of considering the distance to a certain location, the waiting time cost increased by preventing collisions is further added. The path planning method is modified to promptly modify the space-time operation map after each unplanned avoidance occurs, so that the map is consistent with the actual situation.

2. A warehouse management method based on digital twins according to claim 1, characterized in that: Predict the time required for goods to leave the warehouse. Demand forecasting is at the bottom level of the entire warehouse management system and plays a supporting role, supporting multiple decision-making optimization systems at the upper level. These decision-making optimization systems use accurate forecasting data combined with operations research technology to make the best decision and provide the results to the higher-level business execution system.

3. A warehouse management method based on digital twins according to claim 2, characterized in that: The method for predicting the time required for goods to be shipped out of the warehouse is to obtain the required business data from external data sources, then process and clean the basic data, and then process and analyze the data through time series or machine learning or other artificial intelligence technologies, and finally calculate the prediction results and push them to downstream modules for use through various channels; The overall architecture of the cargo outbound demand forecasting module is from top to bottom: data source input layer, basic data processing layer, core business layer, data output layer and downstream execution module; (1) Data source input layer: The data source of warehouse management includes most of the required business data, including cargo information, storage information, and historical outbound information. Most of the cargo transportation plan data comes from operators entering it through the storage and transportation system. (2) Core business layer: This layer is the core part, including feature construction, prediction algorithms, and prediction result processing; it also includes multiple non-intersecting business lines; Feature construction is to convert the previously cleaned basic data into feature data in a standard format through further processing, and provide it to the subsequent algorithm model for use; the core algorithm uses time series analysis, machine learning or other artificial intelligence technologies to predict the demand for goods outbound; the processing of the prediction results cannot meet the downstream modules in terms of format and some special requirements, so it is necessary to process it according to the actual situation, and add additional information such as standard deviation and goods outbound identification; The core technology of cargo outbound demand forecasting is divided into the basic layer, framework layer, tool layer and algorithm layer. The basic layer uses HDFS for data storage, Yarn for resource scheduling, and develops related task scheduling modules. The framework layer is mainly based on Spark RDD, Spark SQL, and Hive. The tool layer directly uses relatively mature and stable algorithms and models, which are encapsulated in third-party Python packages, which are selected from xgboost, numpy, pandas, sklearn, scipy or hyperopt. The algorithm layer is divided into time series and machine learning algorithms. The memory inside the RNN is used to process input sequences of any time series; the time series task prediction of cargo transportation includes ARIMA and Holt winters; among them, ARIMA is an autoregressive integrated moving average model, which is used to predict stable sequences such as the number of warehouse cargo outbound orders; Holt winters, also known as the cubic exponential smoothing algorithm, is used to predict outbound cargo with obvious seasonality and trends.

4. A warehouse management method based on digital twins according to claim 3, characterized in that: The algorithm layer is an iterative decision tree algorithm, GBDT, which consists of multiple decision trees. The decision results of all trees are accumulated to make the final decision. It is used to predict goods with high outbound delivery but no obvious historical pattern.

5. A warehouse management method based on digital twins according to claim 3, characterized in that: The general process of forecasting cargo outbound demand based on machine learning algorithms is as follows: (1) Feature construction: Determine features through data analysis and model testing, and generate feature data in a standard format through a series of tasks; (2) Model selection: Different goods have different characteristics, so different algorithm models will be assigned based on the historical outbound demand of the goods and the sensitivity factors of special training days; (3) Feature selection: Screening a batch of features to filter out unnecessary features. Different types of goods have different features. (4) Sample partitioning: Group the training data into multiple groups of samples, and generate a model file for each group of samples during training; Generally, the same type of goods are grouped together; (5) Model parameters: Select the optimal model parameters. Appropriate parameters will improve the accuracy of the model because different parameter combinations need to be used for model training and prediction. (6) Model training: After the features, models, and samples are determined, model training is performed. After training, a model file is generated and stored in HDFS. (7) Model prediction: read the model file to perform prediction; (8) Multi-model selection: In order to improve the prediction accuracy, multiple algorithm models are used. After the prediction results of each model are output, the best prediction result is selected; (9) Abnormal prediction interception: The more complex and difficult to explain algorithms are, the more likely they are to have extremely high prediction values. Such high predictions cannot be explained by combining historical cargo outbound demand data. Therefore, these abnormal values ​​are intercepted through some rules and replaced with a more conservative value. (10) Model evaluation: calculation of prediction accuracy; (11) Error analysis: By analyzing the prediction accuracy, the distribution of errors in different dimensions can be obtained to provide a reference for algorithm optimization.

6. A warehouse management method based on digital twins according to claim 5, characterized in that: In order to avoid possible collisions, exclusive points are set; at a certain moment, a certain position can only be occupied by one automatic conveying platform; other automatic conveying platforms that want to enter this position at the same moment must wait.

Images