Method and device for cross-warehouse material allocation

Abstract

This application discloses a method and apparatus for cross-warehouse material transfer, applied to warehouse nodes in a cross-warehouse material transfer system, which also includes a management node. The method includes: receiving model parameters from a global model at the management node; generating a local model based on the global model parameters; acquiring production plan data, historical consumption data, and real-time inventory data, and calling the local model to determine the local inventory gap probability distribution based on the production plan data, historical consumption data, and real-time inventory data; sending the local inventory gap probability distribution and real-time inventory data to the management node, so that the management node determines an optimal transfer quantity matrix based on the local inventory gap probability distribution and real-time inventory data of each warehouse node; the optimal transfer quantity matrix is ​​used to indicate the quantity and path allocation of cross-warehouse material transfers. This improves the efficiency and accuracy of cross-warehouse material transfers.

Description

Technical Field

[0001] This application relates to the field of intelligent warehousing technology, and in particular to a method and apparatus for cross-warehouse material transfer. Background Technology

[0002] In modern supply chain management, inventory coordination and optimization within multi-level warehousing systems has always been a key aspect of improving operational efficiency. These systems typically involve the collaborative operation of a central warehouse and multiple regional warehouses, requiring a dynamic balance between material supply and production demand. However, because central and regional warehouses are managed by different entities, detailed inventory data cannot be shared due to privacy barriers. This leads to cross-warehouse transfers relying on manual communication, resulting in severely delayed decision-making and response, ultimately causing inefficient emergency transfers and low scheduling accuracy.

[0003] Therefore, improving the efficiency and accuracy of cross-warehouse material transfer has become an urgent problem to be solved. Summary of the Invention

[0004] This application provides a method and apparatus for cross-warehouse material transfer, which can improve the efficiency and accuracy of cross-warehouse material transfer.

[0005] In a first aspect, embodiments of this application provide a method for cross-warehouse material transfer, applied to warehouse nodes in a cross-warehouse material transfer system, the cross-warehouse material transfer system further including a management node; the method includes:

[0006] Receive model parameters from the global model of the management node;

[0007] Generate a local model based on the model parameters of the global model;

[0008] Acquire production plan data, historical consumption data, and real-time inventory data, and call the local model to determine the probability distribution of local inventory gap based on the production plan data, historical consumption data, and real-time inventory data;

[0009] Send the local inventory gap probability distribution and real-time inventory data to the management node so that the management node can determine the optimal allocation matrix based on the local inventory gap probability distribution and real-time inventory data of each warehouse node; the optimal allocation matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0010] In one embodiment, a local model is invoked to determine the probability distribution of local inventory gaps based on production plan data, historical consumption data, and real-time inventory data. This includes: performing time-series alignment processing on the production plan data and historical consumption data to obtain processed production plan data and processed historical consumption data; grouping the processed production plan data and processed historical consumption data based on the material codes and production batches in the processed production plan data and processed historical consumption data, and obtaining a standardized time-series input vector based on the processed production plan data and processed historical consumption data in each group; performing expiration date compliance verification processing on the real-time inventory data, and generating a compliant inventory status matrix based on the verification results; fusing the standardized time-series input vector and the compliant inventory status matrix to obtain fused features; and invoking the local model to determine the probability distribution of local inventory gaps based on the fused features.

[0011] In one embodiment, a standardized time-series input vector is obtained based on the processed production plan data and processed historical consumption data in each group, including: summarizing the processed production plan data and processed historical consumption data in each group to obtain the demand fluctuation characteristics corresponding to each group; performing numerical mapping processing on multiple demand fluctuation characteristics to obtain multiple processed demand fluctuation characteristics; and sorting the multiple processed demand fluctuation characteristics to obtain a standardized time-series input vector.

[0012] In one embodiment, the local model is invoked to determine the local inventory gap probability distribution based on the fused features, including: inputting the fused features into the local model to obtain the inventory demand forecast for future periods; determining the inventory gap forecast for future periods based on the inventory demand forecast and the compliant inventory status matrix; and determining the local inventory gap probability distribution based on the inventory gap forecasts for multiple future periods.

[0013] In one embodiment, the method further includes: performing gradient encryption processing on the local inventory gap probability distribution to obtain encrypted gradient data; sending the encrypted gradient data to the management node so that the management node updates the model parameters of the global model based on multiple encrypted gradient data, obtains the updated model parameters, and sends the updated model parameters to each warehouse node in the next round of cross-warehouse material transfer.

[0014] Secondly, this application provides a method for cross-warehouse material transfer, applied to a management node in a cross-warehouse material transfer system, which further includes multiple warehouse nodes; the method includes:

[0015] Send the model parameters of the global model to each warehouse node so that each warehouse node can generate a local model based on the model parameters;

[0016] Receive local inventory gap probability distribution and real-time inventory data sent by each warehouse node; the local inventory gap probability distribution is determined by each warehouse node based on the production plan data, historical consumption data and real-time inventory data corresponding to the warehouse node, by calling the local model;

[0017] Based on the local inventory gap probability distribution and real-time inventory data of each warehouse node, a transfer revenue function is constructed.

[0018] Determine the optimal solution of the allocation revenue function and generate the optimal allocation quantity matrix based on the optimal solution; the optimal allocation quantity matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0019] In one embodiment, a transfer revenue function is constructed based on the local inventory gap probability distribution and real-time inventory data of each warehouse node. This includes: quantifying the stockout loss of the local inventory gap probability distribution of each warehouse node to obtain a loss weight matrix that includes the stockout cost of each material; performing warehouse capacity pressure analysis on the real-time inventory data of each warehouse node and determining the warehouse capacity optimization potential vector by material category based on the analysis results; integrating the loss weight matrix and the warehouse capacity optimization potential vector, and combining them with the transportation distance cost coefficient of the transfer path to construct the transfer revenue function.

[0020] In one embodiment, determining the optimal solution of the allocation revenue function includes: acquiring inventory lock-up status data for each warehouse, and constructing a first constraint based on the inventory lock-up status data; acquiring the capacity limit data of automated guided vehicles (AGVs) for each warehouse, and constructing a second constraint based on the capacity limit data; acquiring platform capacity data for each warehouse, and constructing a third constraint based on the platform capacity data; and determining the optimal solution of the allocation revenue function based on the first constraint, the second constraint, and the third constraint.

[0021] Thirdly, this application provides a cross-warehouse material transfer device, applied to various warehouse nodes in a cross-warehouse material transfer system, which also includes a management node; the device includes:

[0022] The transceiver module is used to receive model parameters from the global model of the management node;

[0023] The generation module is used to generate a local model based on the model parameters of the global model.

[0024] The acquisition and determination module is used to acquire production plan data, historical consumption data, and real-time inventory data, and call the local model to determine the probability distribution of local inventory gap based on the production plan data, historical consumption data, and real-time inventory data.

[0025] The transceiver module is also used to send local inventory gap probability distribution and real-time inventory data to the management node, so that the management node can determine the optimal allocation matrix based on the local inventory gap probability distribution and real-time inventory data of each warehouse node; the optimal allocation matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0026] Fourthly, this application provides a cross-warehouse material transfer device, applied to a management node in a cross-warehouse material transfer system, which further includes multiple warehouse nodes; the device includes:

[0027] The transceiver module is used to send the model parameters of the global model to each warehouse node, so that each warehouse node can generate a local model based on the model parameters.

[0028] The transceiver module is also used to receive the local inventory gap probability distribution and real-time inventory data sent by each warehouse node; the local inventory gap probability distribution is determined by each warehouse node based on the production plan data, historical consumption data and real-time inventory data corresponding to the warehouse node, by calling the local model;

[0029] The module is used to construct the transfer revenue function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node.

[0030] The determination module is used to determine the optimal solution of the allocation revenue function and generate the optimal allocation quantity matrix based on the optimal solution; the optimal allocation quantity matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0031] Fifthly, this application provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the method provided in the first aspect, or to implement the steps in the method provided in the second aspect.

[0032] Sixthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0033] In a seventh aspect, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the method provided in the first aspect, or implements the steps in the method provided in the second aspect.

[0034] The aforementioned cross-warehouse material transfer method and apparatus allows each warehouse node in the cross-warehouse material transfer system to receive model parameters from the global model of the management node in the cross-warehouse material transfer system; generate a local model based on the model parameters of the global model; acquire production plan data, historical consumption data, and real-time inventory data, and call the local model to determine the local inventory gap probability distribution based on the production plan data, historical consumption data, and real-time inventory data; send the local inventory gap probability distribution and real-time inventory data to the management node, so that the management node can construct a transfer revenue function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node, and use the optimal solution of the transfer revenue function as the optimal transfer quantity matrix; the optimal transfer quantity matrix is ​​used to indicate the quantity and path allocation of cross-warehouse material transfers. This method utilizes a federated learning framework to achieve secure collaboration among distributed nodes. Without sharing raw inventory data, each warehouse node can perform localized training based on global model parameters and generate an inventory gap probability distribution. This effectively eliminates data privacy barriers between the central warehouse and regional warehouses. By sending the local inventory gap probability distribution and real-time inventory data to the management node, the management node can construct an allocation benefit function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node. The optimal solution of the allocation benefit function is used as the optimal allocation matrix to indicate the quantity and path allocation of materials across warehouses. This improves the efficiency and accuracy of cross-warehouse material allocation without relying on manual communication. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of the architecture of a cross-warehouse material transfer system provided in an embodiment of this application;

[0037] Figure 2 This is a flowchart illustrating a cross-warehouse material transfer method provided in an embodiment of this application;

[0038] Figure 3 This is a flowchart illustrating another cross-warehouse material transfer method provided in an embodiment of this application;

[0039] Figure 4 This is a flowchart illustrating another method for cross-warehouse material transfer provided in the embodiments of this application;

[0040] Figure 5This is a schematic diagram of the structure of a cross-warehouse material transfer device applied to various warehouse nodes, provided in an embodiment of this application;

[0041] Figure 6 This is a schematic diagram of a cross-warehouse material transfer device applied to a management node, provided in an embodiment of this application.

[0042] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] Please see Figure 1 , Figure 1 This is a schematic diagram of the architecture of a cross-warehouse material transfer system provided in an embodiment of this application. Figure 1 As shown, the cross-warehouse material transfer system includes management node 101 and multiple warehouse nodes ( Figure 1 (The diagram uses warehouse nodes 1021, 1022, and 1023 as an example.) The management node and each warehouse node communicate with each other via a network.

[0045] like Figure 1As shown, management node 101 can send model parameters of the global model to each warehouse node, and correspondingly, each warehouse node can receive model parameters of the global model from management node 101. Each warehouse node can generate a local model based on the model parameters of the global model; obtain production plan data, historical consumption data, and real-time inventory data, and call the local model to determine the local inventory gap probability distribution based on the production plan data, historical consumption data, and real-time inventory data; then, each warehouse node can send the local inventory gap probability distribution and real-time inventory data to management node 101, and correspondingly, management node 101 can receive the local inventory gap probability distribution and real-time inventory data from each warehouse node; finally, management node 101 can construct an allocation revenue function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node, and use the optimal solution of the allocation revenue function as the optimal allocation quantity matrix; the optimal allocation quantity matrix is ​​used to indicate the quantity and path allocation of materials across warehouses. This method utilizes a federated learning framework to achieve secure collaboration among distributed nodes. Without sharing raw inventory data, each warehouse node can perform localized training based on global model parameters and generate an inventory gap probability distribution. This effectively eliminates data privacy barriers between the central warehouse and regional warehouses. By sending the local inventory gap probability distribution and real-time inventory data to the management node, the management node can construct an allocation benefit function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node. The optimal solution of the allocation benefit function is used as the optimal allocation matrix to indicate the quantity and path allocation of materials across warehouses. This improves the efficiency and accuracy of cross-warehouse material allocation without relying on manual communication.

[0046] Optionally, management node 101 can be a terminal device or a server. Each warehouse node can be a terminal or a server. Terminal devices mentioned here can include, but are not limited to, various personal computers, laptops, smartphones, tablets, and IoT devices. IoT devices can include smart vehicle devices, projection devices, etc. Servers mentioned here can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing cloud computing services, etc., without limitation.

[0047] Please see Figure 2 , Figure 2 This is a flowchart illustrating a cross-warehouse material transfer method provided in an embodiment of this application. This method can be implemented by various warehouse nodes in the cross-warehouse material transfer system (e.g., ...). Figure 1 The cross-warehouse material transfer system also includes a management node, which executes the transfer at each warehouse node within the system. Figure 2 As shown, this method for cross-warehouse material transfer may include, but is not limited to, the following steps:

[0048] S201, Receive model parameters from the global model of the management node.

[0049] In some embodiments, the management node may be a coordination server.

[0050] The coordination server can be an industrial-grade edge server. The coordination server has a built-in parameter management unit and stores the basic parameters of the global model. The basic parameters include hyperparameters such as weight matrix, bias vector, activation function type, batch size, and learning rate.

[0051] In some embodiments, the communication interface between the coordination server and each warehouse node is based on the Message Queuing Telemetry Transport (MQTT) protocol under the Transmission Control Protocol (TCP) / Internet Protocol (IP). This interface has low-latency transmission capabilities and supports breakpoint resume and data verification functions. Data verification is implemented through a Secure Hash Algorithm 256-bit (SHA-256) hash value to verify the integrity of parameters during transmission. Furthermore, this communication interface can adapt to different versions of the warehouse management system running on each warehouse node.

[0052] S202. Generate a local model based on the model parameters of the global model.

[0053] In some embodiments, the warehouse node generates a local model based on the model parameters of the global model. This can be achieved by: performing parameter injection processing on the initialized local model based on the model parameters of the global model, loading model weights onto the neuron nodes of each layer of the initialized local model, and generating a local model capable of making predictions.

[0054] In some embodiments, the local model can be a Long Short-Term Memory (LSTM) network, wherein the LSTM network includes an input layer, a hidden layer, and an output layer. The input layer dimension corresponds to multiple features, including historical weekly average consumption, demand fluctuation coefficient, remaining production order quantity, inventory turnover rate, remaining shelf life of materials in days, and supplier delivery delay rate. The hidden layer is set to two layers, each containing a certain number of neurons. The output layer dimension corresponds to the predicted daily average demand within a fixed future time period.

[0055] In this embodiment, the warehouse node can read the weight matrix and bias vector stored in the initialized local model. The weight matrix contains the connection weights between the input layer and hidden layer, between hidden layers, and between hidden layers and the output layer of the corresponding LSTM network. The bias vector contains the bias parameters of the neurons in each layer. The warehouse node can perform weight loading according to the hierarchical order of the LSTM network architecture. It loads the weight matrix connecting the input layer and the first hidden layer to the input weight nodes of all neurons in that layer, loads the weight matrix connecting the first hidden layer and the second hidden layer to the recurrent weight nodes of the corresponding neurons, loads the weight matrix connecting the hidden layer and the output layer to the connection weight nodes of the output layer neurons, and simultaneously loads the bias vectors corresponding to each layer to the bias nodes of the neurons in that layer.

[0056] During weight loading, the repository node can verify the dimensionality matching between the weight parameters and the neuron nodes in real time. If the number of neurons in the input layer is inconsistent with the number of columns in the input weight matrix, or if the number of neurons in the hidden layer does not match the dimensionality of the recurrent weight matrix, the parameter reloading process is triggered, rereading the parameters of the initialized local model and performing the loading operation. After weight loading is complete, the repository node can confirm the initialization state of the LSTM network, setting the initial state vector of the hidden layer and the cell state vector to zero vectors, and activating the forward computation function of the network. The repository node verifies whether the network can output prediction results that meet the dimensionality requirements by inputting randomly generated test vectors. If the network output dimension after inputting the test vector matches the preset required prediction dimension, an executable local LSTM model for prediction is generated; if the output dimension does not match, the dimensionality matching record during the weight loading process is checked, the problem is located, and the parameter injection operation is re-executed until an executable local LSTM model for prediction is generated.

[0057] S203. Obtain production plan data, historical consumption data, and real-time inventory data, and call the local model to determine the probability distribution of local inventory gap based on the production plan data, historical consumption data, and real-time inventory data.

[0058] The local inventory gap probability distribution refers to the difference between the actual inventory needed by the warehouse and the real-time inventory, as well as the probability corresponding to that difference.

[0059] In some embodiments, each warehouse node may obtain production plan data, historical consumption data, and real-time inventory data by: obtaining production plan data from the manufacturing execution system through a preset interface; obtaining historical consumption data from the report statistics system through a preset interface; and obtaining real-time inventory data by parsing the smallest packaging barcode.

[0060] The production planning data may include, but is not limited to, material codes, production order quantities, delivery cycles, and production line priorities. Optionally, each warehouse node can also retrieve new production planning data from the Manufacturing Execution System in real time or periodically to update the current production planning data.

[0061] Optionally, after obtaining historical consumption data, each warehouse node can also aggregate the historical consumption data according to the combination granularity of the material code and the time period covered by the historical consumption data to obtain aggregated historical consumption data; and perform outlier removal processing on the aggregated historical consumption data to obtain outlier-removed historical consumption data. In this case, each warehouse node calls its local model to determine the local inventory gap probability distribution based on production plan data, historical consumption data, and real-time inventory data. This can be done by calling the local model to determine the local inventory gap probability distribution based on production plan data, outlier-removed historical consumption data, and real-time inventory data.

[0062] Each warehouse node obtains real-time inventory data by parsing the smallest packaging barcode. This can be done by parsing the smallest packaging barcode using a Radio Frequency Identification (RFID) scanning device and determining the real-time inventory data based on the parsing results. Optionally, the parsing results may include, but are not limited to, information such as the material's inventory location, batch number, expiration date, and actual inventory quantity.

[0063] In one optional implementation, the local model is the LSTM model mentioned in step S202; each warehouse node calls the local model and determines the local inventory gap probability distribution based on production plan data, historical consumption data, and real-time inventory data. This can be achieved by: using an LSTM network to determine the hidden state of material demand time series characteristics based on production plan data, historical consumption data, and real-time inventory data; normalizing the hidden state; and determining the local inventory gap probability distribution based on the normalization result.

[0064] In this implementation, each warehouse node uses an LSTM network to determine the hidden state of the material demand time series characteristics based on production plan data, historical consumption data, and real-time inventory data. This can be achieved by: inputting the production plan data, historical consumption data, and real-time inventory data into the LSTM network; and calculating the hidden state of the material demand time series characteristics using the LSTM's forget gate, input gate, cell state update, and output gate. Specifically, the forget gate calculation is based on a combination of a specific function, a weight matrix, and a bias vector; the input gate calculation uses the same type of function and corresponding weight matrix and bias vector; the cell state update combines the forget gate output, the input gate output, and related calculations; and the output gate calculation yields the hidden state of the material demand time series characteristics.

[0065] In this implementation, each warehouse node normalizes the hidden state, which can be done by using a Min-Max method. This maps the hidden state to a fixed interval, eliminating differences in the units of measurement between different features.

[0066] In this implementation, each warehouse node can determine the inventory gap value based on the predicted daily inventory demand and real-time inventory quantity output by the LSTM network within a fixed future time period. The inventory gap value is the difference between the predicted inventory demand and the real-time inventory quantity. Then, each warehouse node can use a normal distribution probability density function to fit the gap value, and after fitting, a local inventory gap probability distribution with gap interval and probability as the corresponding dimensions is generated.

[0067] S204. Send the local inventory gap probability distribution and real-time inventory data to the management node so that the management node can determine the optimal allocation matrix based on the local inventory gap probability distribution and real-time inventory data of each warehouse node; the optimal allocation matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0068] In this embodiment, a federated learning framework is used to achieve secure collaboration among distributed nodes. Without sharing original inventory data, each warehouse node can perform localized training based on global model parameters and generate an inventory gap probability distribution. This effectively eliminates data privacy barriers between the central warehouse and regional warehouses. By sending the local inventory gap probability distribution and real-time inventory data to the management node, the management node can construct an allocation benefit function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node. The optimal solution of the allocation benefit function is used as the optimal allocation matrix to indicate the quantity and path allocation of materials across warehouses. This improves the efficiency and accuracy of cross-warehouse material allocation without relying on manual communication.

[0069] In one alternative implementation, Figure 2In the cross-warehouse material transfer method shown, each warehouse node calls its local model to determine the probability distribution of local inventory gaps based on production plan data, historical consumption data, and real-time inventory data. This can be achieved by: performing time-series alignment on the production plan data and historical consumption data to obtain processed production plan data and processed historical consumption data; grouping the processed production plan data and processed historical consumption data based on the material codes and production batches in the processed production plan data and processed historical consumption data, and obtaining a standardized time-series input vector based on the processed production plan data and processed historical consumption data in each group; performing expiration date compliance verification on the real-time inventory data, and generating a compliant inventory status matrix based on the verification results; fusing the standardized time-series input vector and the compliant inventory status matrix to obtain fused features; and calling the local model to determine the probability distribution of local inventory gaps based on the fused features.

[0070] In some embodiments, each warehouse node performs time-series alignment processing on production plan data and historical consumption data to obtain processed production plan data and processed historical consumption data. This can be achieved by: extracting the timestamp field (denoted as the first timestamp field) from the production plan data, which records the start and end times of the production plan execution; extracting the timestamp field (denoted as the second timestamp field) from the historical consumption data, which records the time of material consumption; and using the field information of the first timestamp field and the field information of the second timestamp field as a basis, performing time-series alignment processing on the production plan data and historical consumption data to obtain processed production plan data and processed historical consumption data.

[0071] Optionally, each warehouse node uses the field information of the first timestamp field and the field information of the second timestamp field as a basis to perform time-series alignment processing on the production plan data and historical consumption data. This can be done by associating the production plan demand data and historical consumption data within the same time interval. If the time interval of the production plan data and the time interval of the historical consumption data overlap, the overlapping interval is taken as the alignment time range. If there is no overlap, the time interval of the production plan data is used as the standard, and the part of the historical consumption data that is closest to that interval is extracted for alignment.

[0072] In some embodiments, each warehouse node obtains a standardized time-series input vector based on the processed production plan data and processed historical consumption data in each group. This can be achieved by: summarizing the processed production plan data and processed historical consumption data in each group to obtain the demand fluctuation characteristics corresponding to each group; performing numerical mapping processing on multiple demand fluctuation characteristics to obtain multiple processed demand fluctuation characteristics; and sorting the multiple processed demand fluctuation characteristics to obtain a standardized time-series input vector.

[0073] Among these, demand fluctuation characteristics may include maximum demand, minimum demand, average demand, and demand variance.

[0074] Optionally, the processed production plan data and processed historical consumption data in each group can be summarized to obtain the demand fluctuation characteristics corresponding to each group. This can be done by summarizing the production plan demand quantity and historical consumption quantity under the same group, and determining the demand fluctuation characteristics corresponding to the group based on the summary results.

[0075] In this system, each element of the standardized temporal input vector corresponds to a feature value, and the vector dimension is consistent with the number of neurons in the input layer of the local LSTM model. This ensures that the vector can be directly input into the model for computation.

[0076] Optionally, the processed demand fluctuation features are sorted to obtain a standardized time-series input vector. This can be achieved by: arranging the processed demand fluctuation features sequentially according to a preset input dimension order to form a vector corresponding to the "material code-production batch" combination; and integrating the vectors corresponding to each combination to obtain the standardized time-series input vector.

[0077] In some embodiments, each warehouse node performs validity period compliance verification on real-time inventory data, which may involve: determining the validity period field and current time of the materials in the real-time inventory data; and performing validity period compliance verification based on the validity period field and current time of the materials.

[0078] Optionally, each warehouse node performs validity compliance verification based on the material's validity period field and the current time. This can be done by comparing the time corresponding to the validity period field of each material with the current system time. If the time corresponding to the material's validity period is earlier than the current time, the material is determined to be expired; if the time corresponding to the material's validity period is later than or equal to the current time, the material is determined to be compliant.

[0079] Optionally, each warehouse node generates a compliant inventory status matrix based on the verification results. This can involve filtering the verification results, removing all data marked as expired materials, and retaining data for compliant materials. Based on the retained compliant material data, the available inventory quantity is determined, where the available inventory quantity is the unused portion of the real-time inventory quantity, excluding materials allocated to production orders but not yet shipped. A two-dimensional data structure is constructed with "Material Code-Inventory Location-Batch" as the row dimension and available inventory quantity as the column dimension. The information in this two-dimensional data structure is checked for completeness. If a row lacks any of the fields for Material Code, Inventory Location, or Batch, or if the Available Inventory Quantity field is empty, that row is marked as abnormal data and removed. If all rows have complete fields, the two-dimensional data structure is converted into a matrix form. The row index of the matrix corresponds to the unique identifier of "Material Code-Inventory Location-Batch," the column index corresponds to the attribute name of the Available Inventory Quantity, and the elements in the matrix are the specific values ​​of the corresponding Available Inventory Quantity, generating a compliant inventory status matrix. Optionally, each warehouse node can also extract the Material Code field, Inventory Location field, and Batch field.

[0080] In some embodiments, each warehouse node performs feature fusion on the standardized time-series input vector and the compliant inventory status matrix to obtain fused features. This can be achieved by: performing dimension matching processing on the standardized time-series input vector and the compliant inventory status matrix; after dimension matching, concatenating the demand fluctuation feature vector corresponding to each "material code-batch" in the standardized time-series input vector with the available inventory value corresponding to the same "material code-batch" in the compliant inventory status matrix to form fused features. The dimension of the fused features is the sum of the dimension of the demand fluctuation feature vector and the dimension of the available inventory value, and the dimension after fusion is consistent with the receiving dimension of the input layer of the local LSTM model.

[0081] Optionally, each warehouse node performs dimensional matching processing on the standardized time-series input vector and the compliant inventory status matrix. This can be done by: extracting the "material code-production batch" information from the standardized time-series input vector, extracting the "material code-batch" information from the compliant inventory status matrix, and associating the two according to "material code-batch" to ensure that the standardized time-series input vector elements of the same material and the same batch correspond one-to-one with the elements of the compliant inventory status matrix.

[0082] In some embodiments, each warehouse node invokes a local model to determine the local inventory gap probability distribution based on fused features. This can be achieved by: inputting fused features into the local model to obtain the inventory demand forecast for future periods; determining the inventory gap forecast for future periods based on the inventory demand forecast and the compliant inventory status matrix; and determining the local inventory gap probability distribution based on the inventory gap forecasts for multiple future periods.

[0083] Optionally, each warehouse node inputs the fused features into its local model to obtain the predicted inventory demand for future periods. This can be achieved by: inputting the fused features into the local model to obtain the predicted daily average inventory demand for future periods; or by accumulating the predicted daily average inventory demand for future periods to obtain the predicted inventory demand for future periods. The local model is an LSTM network. The process of each warehouse node inputting the fused features into its local model to obtain the predicted daily average inventory demand for future periods can be achieved by: first passing the features through the input layer to the first hidden layer; then, the neurons in the first hidden layer generate the first hidden layer state through operations involving a forget gate, an input gate, cell state updates, and an output gate; the first hidden layer state is passed to the second hidden layer, where the same neuronal operations are performed to generate the second hidden layer state; and finally, the second hidden layer state is passed to the output layer, where an activation function maps the hidden layer state to the predicted daily average demand for future periods.

[0084] Optionally, each warehouse node determines the inventory gap forecast for the future period based on the inventory demand forecast and the compliant inventory status matrix. This can be achieved by subtracting the total inventory demand forecast for the future period from the corresponding available inventory value in the compliant inventory status matrix to obtain the inventory gap forecast for the future period.

[0085] Optionally, each warehouse node determines its local inventory gap probability distribution based on the inventory gap forecasts for multiple future periods. This can be achieved by: fitting these forecasts with a probability density function based on the inventory gap forecasts for multiple future periods to determine the probability values ​​corresponding to different gap intervals; and arranging the gap intervals and their corresponding probability values ​​in a fixed format to generate a local inventory gap probability distribution.

[0086] Using this implementation method, the probability distribution of local inventory gaps in each warehouse can be accurately determined.

[0087] In one alternative implementation, Figure 2 In the cross-warehouse material transfer method shown, each warehouse node can also perform gradient encryption processing on the probability distribution of local inventory shortages to obtain encrypted gradient data; send the encrypted gradient data to the management node so that the management node can update the model parameters of the global model based on multiple encrypted gradient data, obtain the updated model parameters, and send the updated model parameters to each warehouse node in the next round of cross-warehouse material transfer.

[0088] In some embodiments, each warehouse node performs gradient encryption processing on the local inventory gap probability distribution to obtain encrypted gradient data. This method may include, but is not limited to, the following steps:

[0089] Step 1: Perform gradient backpropagation on the local inventory gap probability distribution, determine the partial derivatives of each weight parameter based on the local LSTM, and generate the original gradient tensor.

[0090] Optionally, each warehouse node performs gradient backpropagation on the local inventory shortage probability distribution, determines the partial derivatives of each weight parameter based on the local LSTM, and generates the original gradient tensor. This can be achieved by determining the influence of each weight parameter on the loss value, i.e., the partial derivative, according to the chain rule. The partial derivative reflects the direction and magnitude of the weight parameter adjustment. For the gating structure of the local LSTM model, the partial derivatives of the weight parameters corresponding to the input gate, forget gate, and output gate are determined separately to ensure that the gradients of the parameters of each gating structure are accurately captured. After the partial derivatives of all layer weight parameters are calculated, these partial derivatives are integrated according to the network layer and parameter position to form the original gradient tensor in the form of a multi-dimensional array. This tensor contains the gradient information of all trainable weight parameters of the local LSTM model, providing raw gradient data for subsequent encryption processing.

[0091] The loss value can be obtained by comparing the local inventory gap probability distribution data obtained by each warehouse node with the actual inventory gap probability distribution. Each warehouse node can calculate the loss value between the local inventory gap probability distribution and the actual inventory gap probability distribution using the mean squared error loss function, as shown below:

[0092]

[0093] Where L is the loss value, m is the number of samples, and y i Let y be a sample value from the probability distribution of the actual inventory gap. i These are the predicted sample values ​​in the local inventory gap probability distribution output by the local LSTM model.

[0094] Step 2: Perform homomorphic encryption transformation on the original gradient tensor. The floating-point gradient is mapped to the ciphertext integer through the modular exponentiation operation of the Paillier algorithm, generating encrypted gradient components.

[0095] Optionally, each repository node can locally generate Paillier public and private keys. The generated public key contains two key parameters, and the private key contains two more key parameters. The first parameter of the public key is obtained by multiplying two different large prime numbers, and the second parameter is the first parameter plus one. The first parameter of the private key is the least common multiple of the two large prime numbers minus one, and the second parameter is the modular inverse of the result of a specific function calculation. Afterward, each repository node can upload the generated public key to the management node, while the private key is stored in the local hardware security unit of each repository node. After the local model training process is completed, each repository node can extract the raw gradient components from the local model. The extracted raw gradient components include weight gradients and bias gradients. The weight gradients cover the relevant weight gradients of the LSTM hidden layers, and the bias gradients include the bias gradients of the corresponding hidden layers. The extracted gradient components are stored in a fixed-bit floating-point format. Afterwards, each repository node can perform Paillier encryption on each extracted gradient component. Random numbers are introduced during the encryption process, with the value of the random number ranging from 1 to the first parameter of the public key. The encryption operation is implemented through modular exponentiation, which is optimized using the Montgomery algorithm to improve computational efficiency. The encrypted gradient data generated after encryption is stored in a fixed-length integer format. The system manages the encrypted gradient data according to the correspondence between the unique identifier of the gradient component and the encrypted value.

[0096] Optionally, after obtaining the original gradient tensor, each repository node can first perform type normalization on the floating-point gradients in the original gradient tensor to ensure that all gradient data formats are consistent and meet the input requirements of the Paillier algorithm. Then, it initiates homomorphic encryption conversion, calls the pre-configured Paillier algorithm module, loads the public key parameters distributed by the management node (e.g., the coordinating server), and encrypts each gradient element in the original gradient tensor according to the modular exponentiation rules of the Paillier algorithm. The calculation formula for the encrypted gradient component is:

[0097]

[0098] in, These are the encrypted gradient components generated after encryption. Let be a single gradient element in the original gradient tensor, g be the public key generator of the Paillier algorithm, r be a random number representing a generated random number that is coprime to n, n be a product of large prime numbers, and mod represent the modulo operation. Represents modular exponentiation.

[0099] Each repository node can perform the aforementioned modular exponentiation operation on each gradient element in the original gradient tensor, ensuring that each gradient element is converted into a corresponding ciphertext integer. After all gradient elements are encrypted, these ciphertext integers can be arranged in the order of their positions in the original gradient tensor to generate encrypted gradient components. These encrypted gradient components prevent privacy leaks of the original gradient data during transmission and also support aggregation operations by subsequent management nodes.

[0100] Step 3: Perform data encapsulation processing on the encrypted gradient components, add node identifiers and timestamps to generate authentication information, and generate encrypted gradient data with metadata.

[0101] The node identifier is generated by combining the unique code of the repository node with the hardware device identifier of that node. The unique code is used to distinguish different repository nodes, and the hardware device identifier is used to locate the specific device that generates the encrypted gradient components. The node identifier can be used to ensure that the management node (e.g., the coordination server) can accurately identify the source of the encrypted gradient components. Optionally, the node identifier can be generated by each repository node after performing data encapsulation processing on the encrypted gradient components.

[0102] Optionally, each repository node can also obtain the current time as a timestamp. The timestamp is accurate to the second and is used to record the generation time of the encrypted gradient components, providing a basis for subsequent management nodes (such as the coordination server) to verify the timeliness of the data.

[0103] Optionally, each repository node can convert the node identifier and timestamp format to be compatible with the data format of the encrypted gradient component. The converted node identifier and timestamp are then appended as metadata to the encrypted gradient component to form a complete data unit. Integrity verification is performed on this data unit by calculating its hash value to ensure that the data has not been tampered with or missing during the encapsulation process. After successful verification, encrypted gradient data with metadata is generated. This encrypted gradient data can be directly sent to the management node (coordination server) via the transmission channel, providing complete and traceable encrypted data for subsequent gradient aggregation and decryption processing.

[0104] The specific process of the management node updating the model parameters of the global model based on multiple encrypted gradient data to obtain the updated model parameters is described below.

[0105] By employing this implementation method, data security can be improved and data leakage can be minimized by performing gradient encryption on the probability distribution of local inventory gaps.

[0106] Please see Figure 3 , Figure 3This is a flowchart illustrating another cross-warehouse material transfer method provided in an embodiment of this application. This method can be implemented by a management node in the cross-warehouse material transfer system (e.g., a...). Figure 1 The management node 101 in the system executes the cross-warehouse material transfer system, which also includes multiple warehouse nodes. For example... Figure 3 As shown, this method for cross-warehouse material transfer may include, but is not limited to, the following steps:

[0107] S301. Send the model parameters of the global model to each warehouse node so that each warehouse node can generate a local model based on the model parameters.

[0108] When the management node sends model parameters of the global model to each repository node, it can push parameter packets sequentially according to the unique identifier of the repository node. The parameter packets are stored in a compressed format. After receiving the parameter packets, the repository nodes parse them through the communication interface. The parsed parameters are automatically written into the parameter register of the local model. The inference engine of the local model is built on the TensorFlow Lite framework. After the parameters are written, an initial local model is generated. After the initial local model is generated, each repository node can automatically perform an empty data test on the model. During the test, a zero matrix is ​​input into the model to verify the correctness of the model's output dimensions. After the test is completed, the repository node feeds back the verification result to the management node. If the verification result shows that it has failed, the management node can trigger the parameter resend process and resend the parameter packets to the repository node.

[0109] S302. Receive the local inventory gap probability distribution and real-time inventory data sent by each warehouse node; the local inventory gap probability distribution is determined by each warehouse node based on the production plan data, historical consumption data and real-time inventory data corresponding to the warehouse node, by calling the local model.

[0110] S303. Based on the local inventory gap probability distribution and real-time inventory data of each warehouse node, construct the transfer revenue function.

[0111] The allocation revenue function aims to maximize allocation revenue. The calculation of allocation revenue covers three parts: the revenue from filling the inventory gap at the receiving node, the allocation transportation cost, and the risk cost of insufficient inventory at the sending node. The revenue from filling the inventory gap at the receiving node is the product of the gap filling amount and the profit per unit material. The allocation transportation cost is the product of the allocation amount, the transportation distance, and the freight cost per unit distance. The risk cost of insufficient inventory at the sending node is the product of the remaining inventory at the sending node and the risk coefficient. The risk coefficient is determined based on the local inventory gap probability distribution. When the gap probability reaches a fixed threshold, the corresponding risk coefficient is used.

[0112] In some embodiments, the management node can set constraints in the allocation revenue function. The first constraint is the handling equipment capacity constraint, that is, the total amount transferred out of a single warehouse does not exceed the maximum daily handling equipment capacity of that warehouse. The second constraint is the platform capacity constraint, that is, the total daily loading and unloading volume of a single warehouse does not exceed the product of the number of platforms in that warehouse and the average daily turnover of a single platform. The third constraint is the inventory lock constraint, that is, the amount transferred out of a single warehouse does not exceed the available inventory of that warehouse. Available inventory is the difference between real-time inventory and locked inventory. Locked inventory is the quantity of materials allocated to the production process.

[0113] S304. Determine the optimal solution of the allocation revenue function and generate the optimal allocation quantity matrix based on the optimal solution; the optimal allocation quantity matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0114] In some embodiments, the management node may use a genetic algorithm to find the optimal solution for the constrained allocation benefit function to obtain the optimal allocation scheme; and generate the optimal allocation amount matrix based on the optimal allocation scheme.

[0115] In the process of solving the allocation benefit function, the management node can set parameters such as population size, number of iterations, crossover probability, and mutation probability. It eliminates solutions that violate the constraints through feasibility checks and selects solutions with higher fitness through fitness sorting. The fitness is evaluated based on the calculation result of the allocation benefit function. The iteration continues until the fitness converges. The convergence criterion is that the fitness fluctuation of a fixed number of generations does not exceed a fixed proportion. After the iteration converges, the optimal allocation scheme is obtained.

[0116] In the process of generating the optimal allocation quantity matrix based on the optimal allocation plan, the dimension of the optimal allocation quantity matrix is ​​a combination of the outgoing warehouse identifier and the incoming warehouse identifier, and the elements in the optimal allocation quantity matrix correspond to the allocation quantity. The management node can add additional columns to the optimal allocation quantity matrix. The additional columns contain route allocation information, which includes transportation route identifiers, carrier information, estimated transportation time, and loading / unloading platform identifiers. Among them, the transportation route identifier corresponds to the specific transportation route, the carrier information is the entity responsible for transportation, the estimated transportation time is the time required to complete the transportation, and the loading / unloading platform identifiers are the loading / unloading platforms corresponding to the outgoing and incoming warehouses.

[0117] The management node can also synchronize the optimal allocation matrix to the transportation management system and the warehouse management system through an interface after the optimal allocation matrix is ​​generated. After the synchronization is completed, the automatic generation and execution process of the allocation order will be triggered.

[0118] In this embodiment of the application, by fusing the probability distribution of inventory gaps with real-time inventory data, a transfer revenue function is constructed, and an optimal transfer quantity matrix is ​​generated through distributed solution. In this way, the quantity allocation and path planning of cross-warehouse material transfers can be accurately coordinated, thereby solving the defect of traditional optimization models that are divorced from the actual execution conditions of warehousing.

[0119] In one alternative implementation, Figure 3 Step S303 of the cross-warehouse material transfer method shown, namely, the way the management node constructs the transfer revenue function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node, can be as follows: Quantify the stockout loss of the local inventory gap probability distribution of each warehouse node to obtain a loss weight matrix that includes the stockout cost of each material; perform warehouse capacity pressure analysis on the real-time inventory data of each warehouse node, and determine the warehouse capacity optimization potential vector by material category based on the analysis results; integrate the loss weight matrix and the warehouse capacity optimization potential vector, and combine them with the transportation distance cost coefficient of the transfer path to construct the transfer revenue function.

[0120] In some embodiments, the management node performs stockout loss quantification on the local inventory gap probability distribution of each warehouse node to obtain a loss weight matrix that includes the stockout cost of each material. This can be achieved by performing stockout loss quantification on the local inventory gap probability distribution of each warehouse node, mapping the probability values ​​to a preset stockout level penalty cost lookup table, and generating a loss and weight matrix that includes the picking cost of each material.

[0121] Optionally, after obtaining the local inventory shortage probability distribution, the management node can extract the shortage probability value corresponding to each material in the distribution. This probability value reflects the possibility of each material being out of stock in the future period. The pre-stored shortage level penalty cost lookup table is called, and the probability value is matched to the corresponding shortage level and penalty cost through a mapping formula.

[0122] The out-of-stock level penalty cost comparison table divides out-of-stock levels according to the probability range of out-of-stock, with each level corresponding to a fixed unit out-of-stock penalty cost. The range division is determined based on historical out-of-stock loss data, and the cost value is calculated based on factors such as material production interruption losses and emergency procurement premiums.

[0123] The mapping formula can be expressed as follows:

[0124]

[0125] Among them, C s (p) represents the unit stockout penalty cost of a certain material, p is the stockout probability of that material, and C t Let P be the unit penalty cost corresponding to the t-th stockout level. tLet I() be the probability interval corresponding to the t-th out-of-stock level, and let I() be the indicator function, when p∈P t If I() is true, then I() = 1; otherwise, I() = 0.

[0126] The management node can organize the unit stockout penalty cost in a two-dimensional "material-warehouse" format. The row dimension corresponds to each warehouse node, and the column dimension corresponds to each material type. The matrix element is the unit stockout penalty cost of a certain material in a certain warehouse, generating a loss weight matrix that includes the stockout cost of each material. This loss weight matrix that includes the stockout cost of each material can be stored in the cost data area of ​​the management node.

[0127] In some embodiments, the management node performs capacity pressure analysis on the real-time inventory data of each warehouse node and determines the capacity optimization potential vector by material category based on the analysis results. This can be achieved by: performing capacity pressure analysis on the real-time inventory data of each warehouse node to obtain the storage cost savings space for the portion of the current inventory exceeding the safety stock threshold; and generating a capacity optimization potential vector by material category based on the storage cost savings space.

[0128] The safety stock threshold is set based on factors such as historical material consumption fluctuations and procurement cycles, and is the minimum inventory level used to ensure the continuity of material supply.

[0129] Optionally, the management node can calculate the difference between the current inventory level of each material and the safety stock threshold. A positive difference indicates that the material's inventory exceeds the safety level, resulting in redundant storage capacity. A negative or zero difference indicates that the material's inventory does not exceed the safety level, and there is no room for storage capacity optimization. The management node can calculate the space savings in terms of warehousing costs based on the excess amount, using the following formula:

[0130]

[0131] Where, ΔS m Q represents the storage capacity optimization potential of the m-th material. m Let S be the current inventory level of the m-th material. m Let c be the safety stock threshold for the m-th material. h This is the cost coefficient per unit of storage space.

[0132] Optionally, the management node can organize the storage capacity optimization potential of all materials according to the material classification order and store it in the form of a one-dimensional array. The array length is equal to the total number of material types, and the array element is the storage capacity optimization potential of the corresponding material. This generates a storage capacity optimization potential vector by material classification, which is transmitted to the optimization parameter area of ​​the coordination server for multi-objective benefit integration.

[0133] In some embodiments, during the process of integrating the loss weight matrix and the warehouse capacity optimization potential vector, and combining them with the transportation distance cost coefficient of the allocation path to construct the allocation revenue function, the management node can first retrieve the loss weight matrix ΔR from the cost data area. ij Where i represents the warehouse being transferred out and j represents the warehouse being transferred in, the storage capacity optimization potential vector ΔH is retrieved from the optimization parameter region. ij ΔH ij This relates to the potential for warehouse capacity optimization when transferring materials from warehouse i to warehouse j. Simultaneously, it obtains the transfer path distance dij between warehouses from the geographic information system and retrieves the pre-stored unit distance transportation cost c. ij Then, the management node can set weight adjustment coefficients α, β, and γ, where α is used to balance the weight of reduced revenue due to stockout losses, β is used to balance the weight of revenue from warehouse capacity optimization, and γ is used to balance the weight of transportation costs. The values ​​of the three are determined based on the goal of maximizing the overall revenue of the supply chain and satisfy α+β+γ=1.

[0134] Subsequently, the management node can construct a benefit function based on multi-objective integration logic, first calculating the comprehensive benefit term (αΔR) of reducing stockout losses and optimizing inventory capacity. ij +βΔH ij )x ij Then calculate the transportation cost item γd ij c ij x ij The single-path transfer revenue is obtained by subtracting the cost item from the revenue item. Finally, the revenue of all transfer-out and transfer-in warehouse combinations is summed to form a multi-level transfer revenue function (or transfer revenue function). The expression for the multi-level transfer revenue function is as follows:

[0135]

[0136] Where, f(x) ij Let be the multi-level transfer revenue function, where K is the total quantity of materials transferred out of the warehouse (i.e., the number of warehouse nodes that need to transfer materials out), M is the total quantity of materials transferred into the warehouse (i.e., the number of warehouse nodes that need to receive materials), α, β, and γ are weighting adjustment coefficients, and ∆R is the total quantity of materials transferred into the warehouse. ij Here is the loss weight matrix, ∆H ij To optimize the storage capacity potential vector, x ij For the transfer quantity from warehouse i to j, d ij Let c be the distance from warehouse i to j. ij The cost per unit distance for transportation.

[0137] By adopting this implementation method, by constructing a transfer revenue function that includes physical constraints of warehousing such as inventory lock status, equipment capacity limit, and platform capacity, it is beneficial to accurately coordinate the quantity allocation and route planning of cross-warehouse material transfers.

[0138] In one alternative implementation, Figure 3 In step S304 of the cross-warehouse material transfer method shown, determining the optimal solution of the transfer revenue function includes: obtaining inventory lock status data for each warehouse and constructing a first constraint based on the inventory lock status data; obtaining the capacity limit data of the Automated Guided Vehicles (AGVs) in each warehouse and constructing a second constraint based on the capacity limit data; obtaining platform capacity data for each warehouse and constructing a third constraint based on the platform capacity data; and determining the optimal solution of the transfer revenue function based on the first, second, and third constraints.

[0139] The inventory lock status data records the quantity of each material that has been locked for production or orders; locked inventory cannot be transferred. The AGV capacity limit data is the maximum transport volume that the warehouse can schedule for AGVs within a single cycle. The platform capacity data is the maximum amount of materials that the platform can receive within a single cycle.

[0140] In some embodiments, the management node may construct the following first constraint based on inventory lock status data:

[0141]

[0142] Among them, Q i,m Let L be the current inventory level of material m in warehouse i. i,m x represents the locked inventory quantity of material m in warehouse i. i,j,m Let m be the quantity of material i transferred from warehouse i to warehouse j. The first constraint is used to ensure that the remaining inventory after the transfer is not less than the locked quantity.

[0143] In some embodiments, the management node can construct the following second constraint based on the capacity deployment data:

[0144]

[0145] Where N is the total number of material types, x i,j,m C is the quantity of material m transferred from warehouse i to warehouse j. i,agv This represents the upper limit of AGV transport capacity for warehouse i. The second constraint ensures that the transfer volume does not exceed the AGV transport capacity.

[0146] In some embodiments, the third constraint constructed by the management node based on station capacity data may be as follows:

[0147]

[0148] Among them, C j,plat Let be the platform capacity of warehouse j. The third constraint is used to ensure that the amount received by the platform does not exceed the platform's storage capacity.

[0149] In some embodiments, the management node determines the optimal solution of the allocation revenue function based on the first constraint, the second constraint, and the third constraint. This can be achieved by satisfying the first constraint, the second constraint, and the third constraint, and by maximizing f(x). ij The goal is to generate a constrained optimization problem; the constrained optimization problem is solved in a distributed manner, and the optimal allocation scheme is determined iteratively by the alternating direction multiplier method; the optimal allocation scheme is taken as the optimal solution of the allocation benefit function.

[0150] In this approach, the management node employs the alternating direction multiplier method for distributed solution of the constrained optimization problem. First, the constrained optimization problem is decomposed into multiple local subproblems, each corresponding to the allocation decision of a single warehouse node. The objective function of each subproblem is the combination of the allocation benefit and constraint terms for that warehouse. The management node initializes dual variables and penalty parameters. The dual variables are used to coordinate the consistency of constraints among the subproblems, and the penalty parameters are used to control the iteration convergence speed. During iteration, the management node first allows each warehouse node to independently solve its local subproblem, calculating the locally optimal allocation amount based on the current dual variables. The solution process follows the gradient descent rule to ensure the minimization of the local objective function. Each node feeds back its local optimal solution to the coordination server. The server performs consistency checks on all local solutions. If a solution violates global constraints, a correction signal is sent to the corresponding node by updating the dual variables.

[0151] Afterwards, the management node can repeat the "local solution-feedback verification-dual update" process. The iteration terminates when the difference in global benefit between two adjacent iterations is less than a preset convergence threshold, and all constraints are satisfied. After iteration convergence, the management node can collect the final local optimal allocation amount from each node, organize it according to the three-dimensional dimensions of "outbound warehouse-inbound warehouse-material", and convert the three-dimensional data into a two-dimensional matrix. The row dimension represents the outbound warehouse, the column dimension represents the inbound warehouse, and the matrix elements represent the total allocation amount of all materials between the corresponding warehouses, generating an optimal allocation amount matrix. This matrix is ​​transmitted to the scheduling execution module of each warehouse to guide cross-warehouse material allocation operations.

[0152] By adopting this implementation method, physical constraints (inventory lock-in status, equipment capacity limit, and platform capacity, etc.) are imposed on the allocation revenue function, which can help to accurately coordinate the quantity allocation and route planning of cross-warehouse material transfers.

[0153] In one alternative implementation, Figure 3In the cross-warehouse material transfer method shown, the management node can also receive encrypted gradient data sent by each warehouse node. The encrypted gradient data is obtained by each warehouse node through gradient encryption processing of the determined local inventory gap probability distribution. The encrypted gradient data is aggregated and decrypted, and the model parameters of the global model are updated based on the decrypted gradient components to obtain the updated model parameters. The updated model parameters are used to distribute to each warehouse node in the next round of cross-warehouse material transfer.

[0154] The management node performs aggregated decryption processing on each encrypted gradient data and updates the global model parameters based on the decrypted gradient components. Upon obtaining the updated model parameters, it first verifies the validity of the encrypted gradient data using the previously acquired public keys of each repository node. This verification process is implemented through specific modular arithmetic. If the verification result indicates that the encrypted gradient data is invalid, the management node can send a retransmission command to the corresponding repository node, triggering the re-upload of the encrypted gradient data. During the aggregated decryption operation, the management node calls the private key stored in the hardware security unit to perform decryption operations on the encrypted gradient data uploaded by each repository node. The decryption operation is implemented through a combination of a specific function and private key parameters, yielding the gradient components corresponding to each repository node after decryption.

[0155] The management node calculates weights based on the material throughput ratio of each warehouse node. The weight is calculated as the ratio of a single warehouse node's monthly throughput to the sum of all warehouse nodes' monthly throughput. The management node then performs a weighted average calculation based on the calculated weights and the gradient components of the corresponding warehouse nodes to obtain the aggregated gradient. The sum of all weights is 1. The management node uses the Adam optimizer to update the global model parameters. The update formula is: new parameters equal old parameters minus the product of the learning rate and the aggregated gradient. The learning rate is dynamically adjusted at fixed intervals, with an initial fixed value that decays by a fixed percentage after each fixed training iteration. After the global model parameters are updated, the management node uses mean squared error (MSE) as the loss function to validate the updated model parameters. The MSE is calculated based on the average of the squared differences between predicted and actual demand. If the loss function result decreases by a fixed percentage compared to the previous training iteration, the management node saves the updated global model parameters. If this percentage is not reached, the management node triggers a re-aggregation process, re-executing gradient aggregation and parameter updates.

[0156] In some embodiments, the process of the management node aggregating and decrypting each encrypted gradient data, and updating the model parameters of the global model based on the decrypted gradient components to obtain the updated model parameters may include, but is not limited to, the following steps:

[0157] Step 1: Perform consistency verification on the encrypted gradient data from each warehouse node, compare the matching relationship between the node identifier and the management node registration list, and generate a valid gradient set.

[0158] In some embodiments, after receiving encrypted gradient data with metadata transmitted by each repository node, the management node can extract node identifiers from the metadata of the encrypted gradient data, and call the repository node registration list pre-stored in the management node. This list records the legitimate identifiers of all repository nodes that have completed authentication and are authorized to participate in the federated learning process. Each extracted node identifier is then matched against the legitimate identifiers in the repository node registration list one by one. To clarify the verification logic, the verification judgment formula is set as follows:

[0159]

[0160] Where V(i) represents the verification result of the i-th node, ID(i) represents the identifier of the i-th node, and L represents the repository node registration list. When V(i)=1, it indicates that the node identifier exists in the repository node registration list, and the management node determines that the encrypted gradient data of the corresponding repository node meets the consistency requirements; when V(i)=0, it indicates that the node identifier is not registered in the repository node registration list, and the management node determines that the encrypted gradient data of the corresponding repository node does not meet the requirements and marks it as invalid data to be removed. The management node collects all encrypted gradient data with a verification result of V(i)=1, organizes them in the order of receipt, and generates a valid gradient set, which is stored in the gradient cache area of ​​the coordination server.

[0161] Step 2: Perform parallel decryption processing on the valid gradient set, and use the private key to perform modular exponentiation inverse operation on each gradient component to restore the floating-point value, generating the decrypted gradient tensor.

[0162] In some embodiments, the management node can retrieve the private key corresponding to the Paillier homomorphic encryption algorithm from the encryption hardware. This private key contains two key parameters, λ and μ, where λ is the Euler totient function calculated based on large prime numbers, and μ is the parameter obtained from the modular inverse operation. Then, the management node can start a parallel computing engine, assigning each encrypted gradient component in the effective gradient set to an independent computing thread, improving decryption efficiency through multi-threaded parallel processing. For each encrypted gradient component, the management node can perform a modular exponentiation operation to achieve decryption, with the decryption formula as follows:

[0163]

[0164] in, This is the original gradient value obtained after decryption. To encrypt the gradient components, n is the product of large prime numbers in the Paillier algorithm, and L(x) is a specific linear function with L(x) = (x-1) / n. The gradient values ​​calculated using this formula are in integer form. The management node can perform inverse quantization according to the quantization rules before encryption to restore the integer gradient values ​​to floating-point values, ensuring that the gradient data is restored to the original data type used during local training. After all computation threads have completed decryption and inverse quantization, the management node can also integrate all floating-point gradient values ​​into a multi-dimensional array structure according to the parameter dimension order of the encrypted gradient components, generating a decrypted gradient tensor. This tensor fully preserves the dimensional characteristics and numerical information of the gradient parameters of each node.

[0165] Step 3: Perform federated aggregation on the decrypted gradient tensor, update the global weight parameters by weighted average according to the amount of data in each node, and generate the updated global model parameters.

[0166] In some embodiments, the management node can extract the amount of training data used by each warehouse node participating in this aggregation during its local training process. This amount of data is the total number of valid training samples after filtering by each node. The management node can calculate the weight coefficient corresponding to each warehouse node. The weight coefficient is determined based on the proportion of the node's data volume to the total data volume of all participating aggregation nodes. The calculation formula is as follows:

[0167]

[0168] Among them, w i Let D be the weight coefficient of the i-th node. i Let m be the amount of training data for the i-th node, and m be the total number of nodes participating in this aggregation. This represents the total amount of training data for all participating aggregation nodes. The management node can multiply the gradient components corresponding to each repository node in the decryption gradient tensor with the weight coefficients of that repository node to obtain the weighted gradient component of each repository node; then, it sums the weighted gradient components of all repository nodes to obtain the aggregated global gradient, calculated using the following formula:

[0169]

[0170] in, For global gradient, Let be the decryption gradient component of the i-th node.

[0171] The management node can invoke a preset model optimization algorithm to update the weight parameters of the original global model based on the global gradient. During the update process, the weight matrix and bias vector in the original global model are replaced according to the parameter adjustment rules of the optimization algorithm. After the update is completed, the management node can perform an integrity check on the newly generated global model parameters to confirm that all parameter dimensions and values ​​meet the model structure requirements. Finally, the updated global model parameters are generated and stored in the management node's model parameter database for subsequent allocation and optimization processing or model parameter distribution in the next round of federated learning.

[0172] By adopting this implementation method, a privacy protection mechanism can be achieved by completing secure aggregation and global model updates at the management node, ensuring that the data remains unchanged while the model is linked, thus fundamentally avoiding the risk of leakage of sensitive commercial information.

[0173] Please participate Figure 4 , Figure 4 This is a flowchart illustrating another cross-warehouse material transfer method provided in this application embodiment. This method is applied to, for example... Figure 1 The cross-warehouse material transfer system shown is an example. Figure 4 As shown, the method may include, but is not limited to, the following steps:

[0174] S401. The management node sends the model parameters of the global model to each warehouse node, and correspondingly, each warehouse node receives the model parameters of the global model from the management node.

[0175] In an optional implementation, the relevant description of step S401 can be found in the descriptions of steps S201 and S301 above, and will not be repeated here.

[0176] S402. Each warehouse node generates a local model based on the model parameters of the global model.

[0177] In an optional implementation, the relevant description of step S402 can be found in the description of step S202 above, and will not be repeated here.

[0178] S403. Each warehouse node acquires production plan data, historical consumption data, and real-time inventory data, and calls the local model to determine the probability distribution of local inventory gaps based on the production plan data, historical consumption data, and real-time inventory data.

[0179] In an optional implementation, the relevant description of step S403 can be found in the description of step S203 above, as well as the description of step S203 above, and will not be repeated here.

[0180] S404. Each warehouse node sends its local inventory gap probability distribution and real-time inventory data to the management node. Correspondingly, the management node receives the local inventory gap probability distribution and real-time inventory data from each warehouse node.

[0181] S405. The management node constructs an allocation revenue function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node.

[0182] In an optional implementation, the relevant description of step S405 can be found in the description of step S303 above, and will not be repeated here.

[0183] S406. The management node determines the optimal solution of the allocation revenue function and generates the optimal allocation quantity matrix based on the optimal solution; the optimal allocation quantity matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0184] In an optional implementation, the relevant description of step S406 can be found in the description of step S304 above, and will not be repeated here.

[0185] In one optional implementation, the management node can also receive encrypted gradient data sent by each warehouse node. The encrypted gradient data is obtained by each warehouse node through gradient encryption processing of the determined local inventory gap probability distribution. The encrypted gradient data is aggregated and decrypted, and the model parameters of the global model are updated based on the decrypted gradient components to obtain the updated model parameters. The updated model parameters are used to distribute to each warehouse node in the next round of cross-warehouse material transfer.

[0186] In this embodiment, a federated learning framework is used to achieve secure collaboration among distributed nodes. Without sharing original inventory data, each warehouse node can perform localized training based on global model parameters and generate an inventory gap probability distribution. This effectively eliminates data privacy barriers between the central warehouse and regional warehouses. By fusing the inventory gap probability distribution with real-time inventory data, a transfer benefit function is constructed. The optimal transfer amount matrix is ​​generated through distributed solution. This not only eliminates the need for manual communication but also solves the problem that traditional optimization models are detached from actual warehousing execution conditions. As a result, the efficiency and accuracy of cross-warehouse material transfer can be improved (it can accurately coordinate the quantity allocation and path planning of cross-warehouse material transfer).

[0187] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0188] Based on the same inventive concept, this application also provides a cross-warehouse material transfer device for implementing the cross-warehouse material transfer method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more cross-warehouse material transfer device embodiments provided below can be found in the limitations of the cross-warehouse material transfer method described above, and will not be repeated here.

[0189] Please see Figure 5 , Figure 5 This is a schematic diagram of a cross-warehouse material transfer device applied to various warehouse nodes, provided in an embodiment of this application. Figure 5 As shown, the inter-warehouse material transfer device may include, but is not limited to:

[0190] The transceiver module 501 is used to receive model parameters from the global model of the management node;

[0191] The generation module 502 is used to generate a local model based on the model parameters of the global model.

[0192] The acquisition and determination module 503 is used to acquire production plan data, historical consumption data and real-time inventory data, and call the local model to determine the probability distribution of local inventory gap based on the production plan data, historical consumption data and real-time inventory data;

[0193] The transceiver module 501 is also used to send the local inventory gap probability distribution and real-time inventory data to the management node, so that the management node can determine the optimal allocation matrix based on the local inventory gap probability distribution and real-time inventory data of each warehouse node; the optimal allocation matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0194] In one embodiment, when the acquisition and determination module 503 calls the local model to determine the probability distribution of local inventory gaps based on production plan data, historical consumption data, and real-time inventory data, it specifically performs the following steps: Time-series alignment processing on the production plan data and historical consumption data to obtain processed production plan data and processed historical consumption data; grouping the processed production plan data and processed historical consumption data based on the material codes and production batches in the processed production plan data and processed historical consumption data, and obtaining a standardized time-series input vector based on the processed production plan data and processed historical consumption data in each group; performing validity period compliance verification processing on the real-time inventory data, and generating a compliant inventory status matrix based on the verification results; fusing the standardized time-series input vector and the compliant inventory status matrix to obtain fused features; and calling the local model to determine the probability distribution of local inventory gaps based on the fused features.

[0195] In one embodiment, when the acquisition and determination module 503 obtains a standardized time-series input vector based on the processed production plan data and processed historical consumption data in each group, it specifically performs the following steps: summarizing the processed production plan data and processed historical consumption data in each group to obtain the demand fluctuation characteristics corresponding to each group; performing numerical mapping processing on multiple demand fluctuation characteristics to obtain multiple processed demand fluctuation characteristics; and sorting the multiple processed demand fluctuation characteristics to obtain a standardized time-series input vector.

[0196] In one embodiment, when the acquisition and determination module 503 calls the local model to determine the local inventory gap probability distribution based on the fused features, it is specifically used to: input the fused features into the local model to obtain the inventory demand forecast for the future period; determine the inventory gap forecast for the future period based on the inventory demand forecast and the compliant inventory status matrix; and determine the local inventory gap probability distribution based on the inventory gap forecasts for multiple future periods.

[0197] In one embodiment, the device may further include an encryption module; the encryption module is used to perform gradient encryption processing on the local inventory gap probability distribution to obtain encrypted gradient data; the transceiver module 501 is also used to send the encrypted gradient data to the management node so that the management node updates the model parameters of the global model based on multiple encrypted gradient data, obtains the updated model parameters, and sends the updated model parameters to each warehouse node in the next round of cross-warehouse material transfer.

[0198] Please see Figure 6 , Figure 6 This is a schematic diagram of a cross-warehouse material transfer device applied to a management node, as provided in an embodiment of this application. Figure 6As shown, the inter-warehouse material transfer device may include, but is not limited to:

[0199] The transceiver module 601 is used to send the model parameters of the global model to each warehouse node, so that each warehouse node can generate a local model based on the model parameters.

[0200] The transceiver module 601 is also used to receive the local inventory gap probability distribution and real-time inventory data sent by each warehouse node; the local inventory gap probability distribution is determined by each warehouse node based on the production plan data, historical consumption data and real-time inventory data corresponding to the warehouse node, by calling the local model;

[0201] Module 602 is used to construct the transfer revenue function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node.

[0202] The determination module 603 is used to determine the optimal solution of the allocation revenue function and generate the optimal allocation quantity matrix based on the optimal solution; the optimal allocation quantity matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

[0203] In one embodiment, when constructing the transfer revenue function based on the local inventory gap probability distribution and real-time inventory data of each warehouse node, the construction module 602 specifically performs the following steps: quantifies the stockout loss of the local inventory gap probability distribution of each warehouse node to obtain a loss weight matrix containing the stockout cost of each material; performs warehouse capacity pressure analysis on the real-time inventory data of each warehouse node and determines the warehouse capacity optimization potential vector by material category based on the analysis results; integrates the loss weight matrix and the warehouse capacity optimization potential vector, and combines them with the transportation distance cost coefficient of the transfer path to construct the transfer revenue function.

[0204] In one embodiment, when determining the optimal solution of the allocation revenue function, the determining module 603 is specifically used to: acquire inventory lock status data of each warehouse, and construct a first constraint based on the inventory lock status data; acquire the capacity limit data of the automated guided vehicles of each warehouse, and construct a second constraint based on the capacity limit data; acquire platform capacity data of each warehouse, and construct a third constraint based on the platform capacity data; and determine the optimal solution of the allocation revenue function based on the first constraint, the second constraint, and the third constraint.

[0205] Each module in the aforementioned inter-warehouse material transfer device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of the terminal device in hardware form or independent of it, or stored in the memory of the terminal device in software form, so that the processor can call and execute the corresponding operations of each module.

[0206] In one exemplary embodiment, a computer device is provided, which may be the aforementioned management node or the aforementioned warehouse nodes, and its internal structure diagram may be as follows. Figure 7 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements steps performed by a management node in a cross-warehouse material transfer method, or steps performed by individual warehouse nodes in a cross-warehouse material transfer method. The display unit of the computer device forms a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0207] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0208] In one exemplary embodiment, this application provides a computer device including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps performed by the management node in the above-described cross-warehouse material transfer method, or implements the steps performed by each warehouse node in the above-described cross-warehouse material transfer method.

[0209] In one exemplary embodiment, this application provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements the steps performed by the management node in the above-described cross-warehouse material transfer method, or implements the steps performed by each warehouse node in the above-described cross-warehouse material transfer method.

[0210] In one exemplary embodiment, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps performed by the management node in the above-described cross-warehouse material transfer method, or implements the steps performed by each warehouse node in the above-described cross-warehouse material transfer method.

[0211] It should be noted that the data involved in this application (including but not limited to acquired data, data used for analysis, and stored data) are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0212] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0213] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0214] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for inter-warehouse material transfer, characterized in that, The method is applied to warehouse nodes in a cross-warehouse material transfer system, which further includes a management node; the method includes: Receive model parameters from the global model of the management node; Based on the model parameters of the global model, a local model is generated; Acquire production plan data, historical consumption data, and real-time inventory data, and call the local model to determine the local inventory gap probability distribution based on the production plan data, historical consumption data, and real-time inventory data; The local inventory gap probability distribution and the real-time inventory data are sent to the management node so that the management node can construct an allocation revenue function based on the local inventory gap probability distribution and the real-time inventory data of each warehouse node, and use the optimal solution of the allocation revenue function as the optimal allocation amount matrix; the optimal allocation amount matrix is ​​used to indicate the quantity and path allocation of materials across warehouses.

2. The method according to claim 1, characterized in that, The step of calling the local model to determine the local inventory gap probability distribution based on the production plan data, the historical consumption data, and the real-time inventory data includes: The production plan data and the historical consumption data are time-series aligned to obtain the processed production plan data and the processed historical consumption data. Based on the material codes and production batches in the processed production plan data and the processed historical consumption data, the processed production plan data and the processed historical consumption data are grouped, and a standardized time series input vector is obtained based on the processed production plan data and the processed historical consumption data in each group. The real-time inventory data is subjected to validity period compliance verification, and a compliance inventory status matrix is ​​generated based on the verification results. The standardized time-series input vector and the compliant inventory status matrix are fused to obtain fused features; The local model is invoked, and the probability distribution of the local inventory gap is determined based on the fused features.

3. The method according to claim 2, characterized in that, Based on the processed production plan data and processed historical consumption data from each group, a standardized time-series input vector is obtained, including: The processed production plan data and processed historical consumption data of each group are summarized to obtain the demand fluctuation characteristics corresponding to each group. Numerical mapping processing is performed on the multiple demand fluctuation characteristics to obtain multiple processed demand fluctuation characteristics; The processed demand fluctuation characteristics are sorted to obtain a standardized time series input vector.

4. The method according to claim 2, characterized in that, The step of calling the local model and determining the local inventory gap probability distribution based on the fused features includes: The fused features are input into the local model to obtain the inventory demand forecast for future periods; Based on the inventory demand forecast and the compliant inventory status matrix, the inventory gap forecast for the future period is determined. The probability distribution of the local inventory gap is determined based on the inventory gap forecasts for multiple future periods.

5. The method according to claim 1, characterized in that, The method further includes: The local inventory gap probability distribution is subjected to gradient encryption processing to obtain encrypted gradient data; The encrypted gradient data is sent to the management node so that the management node updates the model parameters of the global model based on multiple encrypted gradient data, obtains the updated model parameters, and sends the updated model parameters to each warehouse node in the next round of cross-warehouse material transfer.

6. A method for inter-warehouse material transfer, characterized in that, A management node applied to a cross-warehouse material transfer system, wherein the cross-warehouse material transfer system further includes multiple warehouse nodes; the method includes: The model parameters of the global model are sent to each of the warehouse nodes so that each of the warehouse nodes can generate a local model based on the model parameters; Receive local inventory gap probability distribution and real-time inventory data sent by each of the warehouse nodes; the local inventory gap probability distribution is determined by each of the warehouse nodes based on the production plan data, historical consumption data and real-time inventory data corresponding to the warehouse node, by calling the local model; Based on the local inventory gap probability distribution and real-time inventory data of each of the warehouse nodes, a transfer revenue function is constructed. The optimal solution of the allocation revenue function is determined, and the optimal solution is used as the optimal allocation quantity matrix; the optimal allocation quantity matrix is ​​used to indicate the quantity and path allocation of materials across warehouses.

7. The method according to claim 6, characterized in that, The transfer revenue function is constructed based on the local inventory gap probability distribution of each of the warehouse nodes and the real-time inventory data, including: The stockout loss is quantified by performing stockout loss quantification on the local inventory gap probability distribution of each warehouse node to obtain a loss weight matrix that includes the stockout cost of each material. The warehouse capacity pressure analysis is performed on the real-time inventory data of each warehouse node, and the warehouse capacity optimization potential vector is determined by material classification based on the analysis results. The loss weight matrix and the warehouse capacity optimization potential vector are integrated, and the transportation distance cost coefficient of the allocation path is combined to construct the allocation revenue function.

8. The method according to claim 6, characterized in that, Determining the optimal solution of the allocation revenue function includes: Obtain inventory lock status data for each warehouse, and construct a first constraint based on the inventory lock status data; Obtain the maximum capacity data of the automated guided vehicles in each of the warehouses, and construct a second constraint based on the maximum capacity data; Obtain the platform capacity data of each of the warehouses, and construct a third constraint based on the platform capacity data; Based on the first constraint, the second constraint, and the third constraint, the optimal solution of the allocation revenue function is determined.

9. A cross-warehouse material transfer device, characterized in that, The device is applied to warehouse nodes in a cross-warehouse material transfer system, which also includes a management node; the device includes: The transceiver module is used to receive model parameters from the global model of the management node; The generation module is used to generate a local model based on the model parameters of the global model; The acquisition and determination module is used to acquire production plan data, historical consumption data, and real-time inventory data, and call the local model to determine the local inventory gap probability distribution based on the production plan data, the historical consumption data, and the real-time inventory data. The transceiver module is further configured to send the local inventory gap probability distribution and the real-time inventory data to the management node, so that the management node can determine the optimal allocation matrix based on the local inventory gap probability distribution and the real-time inventory data of each of the warehouse nodes; the optimal allocation matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

10. A cross-warehouse material transfer device, characterized in that, A management node applied in a cross-warehouse material transfer system, wherein the cross-warehouse material transfer system further includes multiple warehouse nodes; the device includes: The transceiver module is used to send model parameters of the global model to each of the warehouse nodes, so that each of the warehouse nodes can generate a local model based on the model parameters; The transceiver module is also used to receive the local inventory gap probability distribution and real-time inventory data sent by each of the warehouse nodes; the local inventory gap probability distribution is determined by each of the warehouse nodes based on the production plan data, historical consumption data and real-time inventory data corresponding to the warehouse node, by calling the local model; The construction module is used to construct the allocation revenue function based on the local inventory gap probability distribution of each of the warehouse nodes and the real-time inventory data; The determination module is used to determine the optimal solution of the allocation revenue function and use the optimal solution as the optimal allocation quantity matrix; the optimal allocation quantity matrix is ​​used to indicate the quantity and path allocation of materials transferred across warehouses.

Images