An intelligent inventory allocation system based on multi-warehouse cooperation

Abstract

The application discloses an intelligent inventory allocation system based on multi-warehouse cooperation, comprising: a decentralized allocation management module, a multi-source heterogeneous data fusion module, a federal learning driving module, a real-time sensing and dynamic response module, a man-machine interaction and explainability module, a profit-oriented end-to-end comprehensive cost optimization module, a digital twin simulation module and a blockchain smart contract execution module. Through the decentralized blockchain architecture, single point failure is eliminated, the system robustness and cross-subject cooperation credibility are significantly enhanced, multi-source heterogeneous data is fused, real-time sensing and edge computing are used, precise demand prediction and event-driven dynamic allocation are realized, federal learning is adopted to achieve global inventory cooperative optimization on the premise of protecting the data privacy of each participant, an end-to-end profit-oriented mathematical model and digital twin simulation framework are constructed, local optimization caused by step-by-step optimization is avoided, and the decision-making scientificity is improved.

Description

Technical Field

[0001] This invention belongs to the field of warehouse management system technology, specifically an intelligent inventory allocation system based on multi-warehouse collaboration. Background Technology

[0002] Existing inventory allocation systems generally adopt a centralized scheduling model at the architectural level, with all allocation decisions relying on a single central server. If this server fails, the entire scheduling system is paralyzed. Furthermore, centralized architectures face severe scalability bottlenecks and communication load pressure when handling high-concurrency real-time data. Collaboration between multiple warehouse nodes depends on coordination by the central node, and in multi-entity (e.g., multiple brands, multiple merchants, multiple logistics service providers) collaboration scenarios, there is a lack of effective mutual trust mechanisms and transparent transaction records, resulting in high trust costs for cross-entity allocations. At the data perception and prediction level, existing systems mainly rely on historical sales and inventory data for demand forecasting, failing to deeply integrate multi-dimensional data such as social media trends, weather and seasonal changes, macroeconomic indicators, competitor dynamics, and upstream supply chain information. In addition, inventory perception is mostly limited to timed batch collection, lacking real-time inventory perception capabilities based on IoT and RFID technologies. Allocation decisions are still mainly triggered in batches within fixed time windows, failing to achieve event-driven dynamic responses. Regarding cross-entity data collaboration, inventory and sales data among different brands, merchants, and logistics service providers are severely isolated due to data privacy and trade secret protection requirements. Existing systems lack the technical means to achieve global inventory optimization while protecting the data privacy of all parties. At the decision optimization level, existing allocation models are mostly local decision-making models with step-by-step optimization, failing to incorporate allocation costs, sales profits, dynamically changing oil prices and storage fees, predictive transportation and temporary storage losses, and stockout opportunity costs into a unified profit-oriented optimization framework. This results in decisions that are often locally optimal rather than globally optimal. Furthermore, most optimization models are "black box" models, making the decision-making process difficult to interpret. Managers cannot understand the logic behind the allocation recommendations, severely impacting their trust and willingness to adopt them. Simultaneously, existing systems lack effective human-machine collaboration mechanisms, failing to provide a closed-loop channel for human intervention and feedback optimization in abnormal scenarios. These shortcomings collectively constrain the decision-making quality and operational efficiency of inventory allocation systems in dynamic and complex supply chain environments. Summary of the Invention

[0003] The technical solution adopted in this invention is as follows: An intelligent inventory transfer system based on multi-warehouse collaboration, comprising:

[0004] The module includes a decentralized allocation management module, a multi-source heterogeneous data fusion module, a federated learning-driven module, a real-time perception and dynamic response module, a human-computer interaction and interpretability module, a profit-oriented end-to-end comprehensive cost optimization module, a digital twin simulation module, and a blockchain smart contract execution module.

[0005] The decentralized allocation management module is communicatively connected to the multi-source heterogeneous data fusion module, the federated learning-driven module, the real-time perception and dynamic response module, the profit-oriented end-to-end comprehensive cost optimization module, the digital twin simulation module, and the blockchain smart contract execution module, respectively, and is used to receive data output by each module and generate global allocation decisions.

[0006] The output of the multi-source heterogeneous data fusion module is connected to the input of the federated learning driver module, and is used to provide the federated learning driver module with fused multi-dimensional training data.

[0007] The output of the federated learning-driven module is connected to the input of the profit-oriented end-to-end integrated cost optimization module, and is used to provide the profit-oriented end-to-end integrated cost optimization module with global inventory status data after cross-entity collaborative optimization.

[0008] The output of the real-time sensing and dynamic response module is connected to the input of the decentralized allocation management module and the input of the digital twin simulation module, respectively, to provide real-time inventory sensing data to the decentralized allocation management module and real-time status data to the digital twin simulation module to drive simulation.

[0009] The output of the digital twin simulation module is connected to the input of the profit-oriented end-to-end integrated cost optimization module, and is used to provide the simulation evaluation results of the candidate allocation strategy to the profit-oriented end-to-end integrated cost optimization module.

[0010] The output of the profit-oriented end-to-end integrated cost optimization module is connected to the input of the decentralized allocation management module, and is used to provide the decentralized allocation management module with the most profitable allocation scheme.

[0011] The output of the decentralized allocation management module is connected to the input of the blockchain smart contract execution module, and is used to send the allocation plan to the blockchain smart contract execution module.

[0012] The output of the blockchain smart contract execution module is connected to the execution terminal of each warehouse node, and is used to send the allocation instructions to each warehouse node for execution.

[0013] The human-computer interaction and explainability module is bidirectionally connected to the decentralized allocation management module and the profit-oriented end-to-end integrated cost optimization module, respectively, and is used to receive users' natural language queries and display an explainable report of allocation decisions to users.

[0014] Furthermore, the decentralized allocation management module includes a blockchain consensus submodule and a smart contract execution submodule distributed across various warehouse nodes.

[0015] The blockchain consensus submodule is used to verify and confirm the allocation plan among the warehouse nodes through the Byzantine fault-tolerant consensus mechanism. Each warehouse node independently reaches an allocation agreement based on the consensus mechanism, without the need for a central scheduling server.

[0016] The smart contract execution submodule is used to automatically perform transfer condition checks, transfer priority sorting, and transfer instruction issuance operations after the transfer plan is confirmed by the blockchain consensus submodule.

[0017] Furthermore, the multi-source heterogeneous data fusion module includes a social media data collection submodule, a weather and seasonal data collection submodule, a macroeconomic indicator collection submodule, a competitor dynamics collection submodule, an upstream supply chain information collection submodule, and a multimodal fusion submodule.

[0018] The outputs of the social media data collection submodule, the weather and seasonal data collection submodule, the macroeconomic indicator collection submodule, the competitor dynamics collection submodule, and the upstream supply chain information collection submodule are respectively connected to the input of the multimodal fusion submodule.

[0019] The multimodal fusion submodule is used to uniformly model text data, image data, and structured data to generate multi-dimensional fused demand prediction input data.

[0020] Furthermore, the real-time sensing and dynamic response module includes an IoT sensing submodule, an RFID tag reading submodule, and an edge computing submodule.

[0021] The IoT sensing submodule and the RFID tag reading submodule are deployed at each warehouse node to collect inventory change data, equipment status data and personnel efficiency data within the warehouse.

[0022] The edge computing submodule is deployed on the local edge server of each warehouse node. It is used to perform millisecond-level real-time processing on the inventory change data, the equipment status data and the personnel efficiency data, generate event-driven transfer trigger signals, and send the transfer trigger signals to the decentralized transfer management module.

[0023] Furthermore, the federated learning-driven module includes a local training submodule distributed among the participants, a cryptographic gradient generation submodule, a gradient aggregation submodule, and a global model update submodule.

[0024] The local training submodules of each participant are used to train local prediction models based on local inventory data.

[0025] The encrypted gradient generation submodule is used to encrypt the training gradient of the local prediction model and generate encrypted gradient parameters.

[0026] The gradient aggregation submodule is used to aggregate the encrypted gradient parameters uploaded by each participant to generate an aggregated gradient.

[0027] The global model update submodule is used to update the global inventory prediction model based on the aggregated gradient, and distribute the updated global model parameters to each participant for the next round of local training.

[0028] Furthermore, the profit-oriented end-to-end integrated cost optimization module includes a unified profit objective function construction submodule, a dynamic cost calculation submodule, a predictive loss calculation submodule, an opportunity cost quantification submodule, a multi-objective optimization submodule, and a deep reinforcement learning submodule.

[0029] The unified profit objective function construction submodule is used to construct a mathematical model with maximizing total profit as the single core objective.

[0030] The dynamic cost calculation submodule is used to calculate dynamic cost parameters based on real-time acquired oil price data, storage fee data, and commodity price data.

[0031] The predictive loss calculation submodule is used to predict the transportation loss rate and temporary storage loss rate based on real-time ambient temperature data, real-time ambient humidity data, transportation distance data, and packaging method data.

[0032] The opportunity cost quantification submodule is used to quantify the stockout costs caused by allocation delays or allocation failures.

[0033] The multi-objective optimization submodule is used to solve for the Pareto optimal solution set using a multi-objective genetic algorithm.

[0034] The deep reinforcement learning submodule is used to learn the optimal allocation strategy through continuous interaction with the supply chain environment.

[0035] The unified profit objective function construction submodule is connected to the dynamic cost calculation submodule, the predictive loss calculation submodule, and the opportunity cost quantification submodule, respectively, and is used to receive the dynamic cost parameters, the predictive loss rate, and the stockout cost as input parameters of the objective function.

[0036] The multi-objective optimization submodule and the deep reinforcement learning submodule are respectively connected to the unified profit objective function construction submodule, and are used to solve the optimal allocation scheme based on the objective function.

[0037] Furthermore, the mathematical model constructed by the unified profit objective function construction submodule is as follows:

[0038]

[0039] Where represents total profit, N represents the total number of warehouse nodes, and T represents the total number of decision cycles.

[0040] and Indicates the warehouse node number. Indicates the decision-making cycle.

[0041] Represents a warehouse node In the decision-making cycle The unit price of goods sold.

[0042] Represents a warehouse node In the decision-making cycle Sales volume of goods.

[0043] Represents a warehouse node In the decision-making cycle The unit inventory holding cost.

[0044] This represents the ending inventory of warehouse node i during decision period t.

[0045] Represents a warehouse node In the decision-making cycle The unit cost of out-of-stock items.

[0046] Represents a warehouse node In the decision-making cycle The amount of stock shortage.

[0047] This indicates that during decision cycle t, a unit of goods is transferred from warehouse node i to warehouse node i.

[0048] The allocation cost of j.

[0049] This represents the quantity of goods transferred from warehouse node i to warehouse node j during decision cycle t.

[0050] This represents the predicted loss rate of a unit of goods shipped from warehouse node i during transportation in decision period t.

[0051] This represents the predicted loss rate of temporary storage of goods per unit at warehouse node i during decision period t.

[0052] The mathematical model satisfies the following constraints:

[0053]

[0054]

[0055]

[0056] in, This represents the upper limit of the total allocation capacity during the decision-making period t.

[0057] Furthermore, the multi-objective optimization submodule uses a non-dominated sorting genetic algorithm to solve the Pareto optimal solution set. The non-dominated sorting genetic algorithm includes a population initialization unit, a non-dominated sorting unit, a crowding calculation unit, a selection unit, a crossover unit, a mutation unit, and an iteration termination determination unit.

[0058] The population initialization unit is used to generate an initial population containing multiple allocation scheme individuals.

[0059] The non-dominated sorting unit is used to perform Pareto non-dominated hierarchical sorting of individuals in the population.

[0060] The crowding calculation unit is used to calculate the crowding distance between individuals within the same non-dominated hierarchy.

[0061] The selection unit is used to select parent individuals based on the non-dominated hierarchy and the crowding distance.

[0062] The crossover unit and the mutation unit are used to perform crossover and mutation operations on the parent individual to generate offspring individuals.

[0063] The iteration termination determination unit is used to output the Pareto optimal solution set when the preset maximum number of iterations is reached.

[0064] Furthermore, the human-computer interaction and interpretability module includes a large language model interaction submodule, an interpretable artificial intelligence submodule, and a human-computer collaborative decision-making submodule.

[0065] The large language model interaction submodule is used to receive natural language query statements input by users, perform intent recognition and semantic parsing on the natural language query statements, generate query intent vectors, and send the query intent vectors to the decentralized allocation management module and the profit-oriented end-to-end integrated cost optimization module.

[0066] The explainable artificial intelligence submodule is used to generate a list of key influencing factors and a confidence score for allocation decisions through attention visualization, counterfactual explanation, and SHAP value analysis, and sends the list of key influencing factors and the confidence score to the large language model interaction submodule.

[0067] The large language model interaction submodule is also used to generate a natural language interpretability report based on the list of key influencing factors and the confidence score, and to display it to the user.

[0068] The human-machine collaborative decision-making submodule is used to receive the user's adjustment instruction for the allocation plan, and send the adjusted allocation plan as a feedback signal to the profit-oriented end-to-end integrated cost optimization module for model optimization.

[0069] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the operation method of the intelligent inventory transfer system based on multi-warehouse collaboration as described above.

[0070] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the operation method of the intelligent inventory transfer system based on multi-warehouse collaboration as described above.

[0071] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0072] (1) This invention sets up a decentralized allocation management module combined with a blockchain smart contract execution module, and deploys a blockchain consensus sub-module and a smart contract execution sub-module among each warehouse node. This enables each warehouse node to reach an allocation agreement autonomously based on the Byzantine fault-tolerant consensus mechanism without relying on a central scheduling server. This completely eliminates the risk of single point of failure inherent in traditional centralized scheduling from the architectural level, and significantly improves the system's robustness and disaster recovery capability in complex network environments. At the same time, after the allocation plan is confirmed by consensus, the smart contract execution sub-module automatically performs allocation condition checks, priority sorting, and instruction issuance operations, solidifying business rules on the blockchain in the form of immutable code. This effectively enhances the transaction transparency, data immutability, and mutual trust among multiple parties in the cross-entity collaboration process, and provides a secure and reliable underlying infrastructure for complex supply chain collaboration with multiple warehouses and multiple ownership.

[0073] (2) This invention deeply integrates social media trend data, weather and seasonal data, macroeconomic indicators, competitor dynamics and upstream supply chain information through a multi-source heterogeneous data fusion module, and uses a multi-modal fusion sub-module to unify the modeling of text, images and structured data, which greatly improves the data dimension and accuracy of demand forecasting and overcomes the limitations of traditional forecasting models that rely only on historical sales data and inventory data. At the same time, in conjunction with the IoT sensing sub-module and RFID tag reading sub-module in the real-time perception and dynamic response module, the system collects real-time data on warehouse inventory changes, equipment status and personnel efficiency, and uses the edge computing sub-module deployed on the local edge server for millisecond-level real-time processing, enabling the system to automatically generate event-driven allocation trigger signals, realizing a fundamental shift from traditional timed batch allocation to real-time dynamic response, and significantly enhancing the system's agile perception and rapid response capabilities to sudden market fluctuations and abnormal supply chain events.

[0074] (3) By introducing a federated learning-driven module, this invention enables each participant to independently train a prediction model based on their local inventory data in the local training submodule. The training gradient is encrypted by the encrypted gradient generation submodule and then uploaded to the gradient aggregation submodule. The gradient aggregation submodule aggregates the encrypted gradient parameters of each participant and then the global model update submodule updates the global inventory prediction model. Throughout the process, the original business data of each participant does not leave the local area. Under the premise of fully protecting the data privacy and business secrets of each brand, merchant and logistics service provider, this invention effectively solves the data silo problem in multi-entity collaboration scenarios and realizes cross-entity global inventory status collaborative optimization. This not only meets the rigid requirements of data security and compliance, but also achieves global inventory allocation performance that surpasses the capabilities of a single entity.

[0075] (4) This invention constructs a unified mathematical model with the single core objective of maximizing total profit through a profit-oriented end-to-end integrated cost optimization module. It incorporates allocation costs, sales profits, dynamically changing oil prices and storage fees, predictive transportation loss rates and temporary storage loss rates, and the opportunity cost of stockouts caused by allocation delays or failures into the same optimization framework. This avoids the local optima defects caused by the fragmentation of cost elements in traditional step-by-step optimization methods, ensuring that the allocation scheme output by the system has strict global economic optimality. At the same time, by using a digital twin simulation module to construct a virtual supply chain network containing all cost elements, it is possible to perform full-link "What-if" cost simulation and profit impact assessment on candidate allocation strategies, and to trace and visualize the complete cost composition of each allocation decision. This allows managers to intuitively predict the economic consequences of different strategies before making decisions, thereby greatly improving the scientificity and reliability of complex supply chain decisions.

[0076] (5) This invention receives user natural language queries and generates interpretability reports through the human-computer interaction and the large language model interaction sub-module in the interpretability module. In conjunction with the interpretable artificial intelligence sub-module, it generates a list of key influencing factors and confidence scores for allocation decisions using attention visualization, counterfactual explanation and SHAP value analysis methods. This enables managers to clearly understand the deep decision-making logic such as "why it is recommended to allocate from a specific warehouse", effectively overcoming the shortcomings of traditional black box model decision-making processes that are difficult to explain, and significantly enhancing managers' trust in and willingness to adopt the intelligent system. At the same time, the human-computer collaborative decision-making sub-module constructs a closed-loop decision-making mechanism of automatic recommendation, manual review and feedback learning. In abnormal scenarios, managers can adjust or reject the allocation suggestions generated by the system. The adjustment results are continuously sent as feedback signals to the profit-oriented end-to-end comprehensive cost optimization module for model iteration optimization. Combined with the multi-objective optimization sub-module to solve the Pareto optimal solution set and the deep reinforcement learning sub-module to learn the optimal strategy through continuous interaction with the supply chain environment, the system has the ability to adapt and evolve in a highly dynamic market environment and still maintains a robust channel for manual intervention and decision-making assurance in extreme abnormal scenarios. Attached Figure Description

[0077] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0078] Figure 2 This is a schematic diagram of the multi-source heterogeneous data fusion module of the present invention;

[0079] Figure 3 This is a schematic diagram of the federated learning driver module of the present invention;

[0080] Figure 4 This is a schematic diagram of the real-time sensing and dynamic response module of the present invention;

[0081] Figure 5 This is a schematic diagram of the human-computer interaction and interpretability module of the present invention;

[0082] Figure 6 This is a schematic diagram of the profit-oriented end-to-end integrated cost optimization module of the present invention;

[0083] Figure 7 This is a schematic diagram of the multi-objective optimization submodule of the present invention.

[0084] In the diagram: 1. Decentralized allocation management module; 2. Multi-source heterogeneous data fusion module; 3. Federated learning driven module; 4. Real-time perception and dynamic response module; 5. Human-computer interaction and interpretability module; 6. Profit-oriented end-to-end comprehensive cost optimization module; 7. Digital twin simulation module; 8. Blockchain smart contract execution module. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0086] Example 1, please refer to Figures 1-7 A smart inventory transfer system based on multi-warehouse collaboration includes: a decentralized transfer management module 1, a multi-source heterogeneous data fusion module 2, a federated learning-driven module 3, a real-time perception and dynamic response module 4, a human-computer interaction and interpretability module 5, a profit-oriented end-to-end integrated cost optimization module 6, a digital twin simulation module 7, and a blockchain smart contract execution module 8. The decentralized transfer management module 1 is integrated with the multi-source heterogeneous data fusion module 2, the federated learning-driven module 3, the real-time perception and dynamic response module 4, the profit-oriented end-to-end integrated cost optimization module 6, and the digital twin simulation module 7. The blockchain smart contract execution module 8 is connected to receive data output from each module and generate global allocation decisions. The output of the multi-source heterogeneous data fusion module 2 is connected to the input of the federated learning-driven module 3 to provide fused multi-dimensional training data. The output of the federated learning-driven module 3 is connected to the input of the profit-oriented end-to-end integrated cost optimization module 6 to provide global inventory status data optimized through cross-entity collaboration. The output of the real-time perception and dynamic response module 4 is connected to the decentralized allocation module. The input terminals of the decentralized allocation management module 1 and the digital twin simulation module 7 are used to provide real-time inventory perception data to the decentralized allocation management module 1 and real-time status data to the digital twin simulation module 7 to drive simulation inference. The output terminal of the digital twin simulation module 7 is connected to the input terminal of the profit-oriented end-to-end integrated cost optimization module 6, and is used to provide simulation evaluation results of candidate allocation strategies to the profit-oriented end-to-end integrated cost optimization module 6. The output terminal of the profit-oriented end-to-end integrated cost optimization module 6 is connected to the input terminal of the decentralized allocation management module 1, and is used to provide simulation evaluation results of candidate allocation strategies to the profit-oriented end-to-end integrated cost optimization module 6. Block 1 provides the most profitable allocation scheme. The output of the decentralized allocation management module 1 is connected to the input of the blockchain smart contract execution module 8 to send the allocation scheme to the blockchain smart contract execution module 8. The output of the blockchain smart contract execution module 8 is connected to the execution terminal of each warehouse node to issue the allocation instruction to each warehouse node for execution. The human-computer interaction and explainability module 5 is bidirectionally connected to the decentralized allocation management module 1 and the profit-oriented end-to-end integrated cost optimization module 6 to receive users' natural language queries and display an explainability report of the allocation decision to the users.

[0087] In a specific embodiment of the present invention, the system as a whole takes the decentralized allocation management module 1 as the core hub to coordinate the data flow and decision transmission between various functional modules. The multi-source heterogeneous data fusion module 2 continuously collects multi-dimensional data such as social media sentiment, weather and seasonal data, macroeconomic indicators, competitor dynamics and upstream supply chain information from external data sources. After multimodal fusion processing, it generates demand forecast input data and transmits it to the federated learning driven module 3. The federated learning driven module 3 uses local training sub-modules distributed among various participants to train local prediction models based on local inventory data. The encrypted gradient generation sub-module encrypts the training gradients and uploads them. The gradient aggregation sub-module aggregates the encrypted gradient parameters of various participants to generate aggregate gradients. The global model update sub-module updates the global inventory prediction model based on the aggregated gradients and distributes the updated global model parameters to various participants for the next round of local training. Finally, the global inventory status data optimized by cross-entity collaboration is output to the profit-oriented end-to-end comprehensive cost optimization module 6. At the same time, the real-time perception and dynamic response module 4 collects inventory change data and equipment status data through IoT perception sub-modules and RFID tag reading sub-modules deployed at various warehouse nodes. Based on personnel efficiency data, the edge computing submodule performs millisecond-level real-time processing on the aforementioned data, generates event-driven allocation trigger signals, and sends them to the decentralized allocation management module 1. Simultaneously, it sends real-time status data to the digital twin simulation module 7. The digital twin simulation module 7 simulates and extrapolates candidate allocation strategies based on the real-time status data, evaluates the performance of each strategy in different scenarios, and feeds back the simulation evaluation results to the profit-oriented end-to-end integrated cost optimization module 6. The profit-oriented end-to-end integrated cost optimization module 6 takes maximizing total profit as its single core objective, comprehensively considering dynamic costs, predictive losses, and machine costs. The system calculates the optimal allocation plan based on multiple parameters, including cost, and outputs the results to the decentralized allocation management module 1. The decentralized allocation management module 1 then sends the allocation plan to the blockchain smart contract execution module 8. The blockchain smart contract execution module 8 issues the allocation instructions to the execution terminals of each warehouse node to execute the allocation operation. The human-computer interaction and interpretability module 5 continuously interacts with the user throughout the decision-making process, receives the user's natural language queries and displays an interpretability report of the allocation decision, and simultaneously receives the user's adjustment instructions as feedback signals to the profit-oriented end-to-end comprehensive cost optimization module 6 for model optimization.

[0088] Specifically, the decentralized allocation management module 1 includes a blockchain consensus submodule and a smart contract execution submodule distributed across various warehouse nodes. The blockchain consensus submodule is used to verify and confirm the allocation plan among various warehouse nodes through a Byzantine fault-tolerant consensus mechanism. Each warehouse node autonomously reaches an allocation agreement based on the consensus mechanism without the need for a central scheduling server. The smart contract execution submodule is used to automatically perform allocation condition checks, allocation priority sorting, and allocation instruction issuance operations after the allocation plan is confirmed by the blockchain consensus submodule.

[0089] In a specific embodiment of the present invention, the decentralized allocation management module 1 uses blockchain technology to achieve distributed allocation management without a central scheduling server. The blockchain consensus sub-module distributed across each warehouse node verifies and confirms the allocation plan through a Byzantine fault-tolerant consensus mechanism. Each warehouse node autonomously reaches an allocation agreement based on the consensus mechanism, without relying on a central scheduling server, thus avoiding the risk of single point of failure. After the allocation plan is confirmed by the blockchain consensus sub-module, the smart contract execution sub-module automatically performs allocation condition checks, allocation priority sorting, and allocation instruction issuance operations to ensure the automation, transparency, and immutability of the allocation process.

[0090] Specifically, the multi-source heterogeneous data fusion module 2 includes a social media data collection submodule, a weather and seasonal data collection submodule, a macroeconomic indicator collection submodule, a competitor dynamics collection submodule, a supply chain upstream information collection submodule, and a multimodal fusion submodule. The outputs of the social media data collection submodule, the weather and seasonal data collection submodule, the macroeconomic indicator collection submodule, the competitor dynamics collection submodule, and the supply chain upstream information collection submodule are respectively connected to the input of the multimodal fusion submodule. The multimodal fusion submodule is used to uniformly model text data, image data, and structured data to generate multi-dimensional fused demand forecast input data.

[0091] In a specific embodiment of the present invention, the multi-source heterogeneous data fusion module 2 is responsible for comprehensively collecting multi-dimensional information affecting inventory demand from various external data sources. The social media data collection submodule captures user discussions, consumption trends, and hot events on social platforms in real time through API interfaces; the weather and seasonal data collection submodule obtains data such as temperature, precipitation, and seasonal changes in the areas where each warehouse is located from meteorological service interfaces; the macroeconomic indicator collection submodule collects macroeconomic data such as GDP growth rate, consumer price index, and consumer confidence index; the competitor dynamics collection submodule monitors competitors' promotional activities, new product launches, and price changes; and the upstream supply chain information collection submodule obtains upstream information such as supplier capacity, raw material prices, and logistics status. The data collected by the above submodules are output to the multimodal fusion submodule, which performs unified modeling of text data, image data, and structured data to generate multi-dimensional fused demand forecast input data, providing a high-quality data foundation for subsequent federated learning and cost optimization.

[0092] Specifically, the real-time sensing and dynamic response module 4 includes an IoT sensing submodule, an RFID tag reading submodule, and an edge computing submodule. The IoT sensing submodule and the RFID tag reading submodule are deployed at each warehouse node to collect inventory change data, equipment status data, and personnel efficiency data within the warehouse. The edge computing submodule is deployed on the local edge server of each warehouse node to perform millisecond-level real-time processing on the inventory change data, equipment status data, and personnel efficiency data, generate event-driven transfer trigger signals, and send the transfer trigger signals to the decentralized transfer management module 1.

[0093] In a specific embodiment of the present invention, the real-time sensing and dynamic response module 4 constructs an end-to-end real-time sensing and response system. The Internet of Things (IoT) sensing submodule and the radio frequency identification (RFID) tag reading submodule are deployed at each warehouse node. The IoT sensing submodule collects environmental data and inventory physical status in real time through temperature and humidity sensors, weight sensors, and motion sensors deployed in the warehouse. The RFID tag reading submodule tracks inventory changes and product flow in real time by reading RFID tags attached to goods. The edge computing submodule is deployed on the local edge server of each warehouse node. It performs millisecond-level real-time processing on the collected inventory change data, equipment status data, and personnel efficiency data. It identifies abnormal inventory changes and emergency allocation needs through preset event-driven rules and machine learning models, generates event-driven allocation trigger signals, and sends the allocation trigger signals to the decentralized allocation management module 1 in real time, enabling the system to respond instantly to sudden changes in demand and supply interruptions.

[0094] Specifically, the federated learning-driven module 3 includes a local training submodule, an encrypted gradient generation submodule, a gradient aggregation submodule, and a global model update submodule distributed among the participants. The local training submodule of each participant is used to train a local prediction model based on local inventory data. The encrypted gradient generation submodule is used to encrypt the training gradients of the local prediction model and generate encrypted gradient parameters. The gradient aggregation submodule is used to aggregate the encrypted gradient parameters uploaded by each participant and generate aggregated gradients. The global model update submodule is used to update the global inventory prediction model based on the aggregated gradients and distribute the updated global model parameters to each participant for the next round of local training.

[0095] In a specific embodiment of the present invention, the federated learning-driven module 3 uses a federated learning framework to achieve collaborative modeling across multiple warehouse entities and supply chain participants. While protecting the data privacy of each participant, a global inventory prediction model is trained. Each participant's local training submodule trains its local prediction model on its local device based on its local inventory, sales, and demand data. During training, the original data remains local; only model gradient parameters are generated. The encrypted gradient generation submodule encrypts the training gradient of the local prediction model using homomorphic encryption or differential privacy technology. After generating the encrypted gradient parameters, it is uploaded to the gradient aggregation submodule via a secure communication channel. The gradient aggregation submodule aggregates the encrypted gradient parameters uploaded by each participant and generates an aggregated gradient using secure multi-party computation or federated averaging algorithms, ensuring that the gradient information of any single participant is not leaked during the aggregation process. The global model update submodule updates the global inventory prediction model based on the aggregated gradient and distributes the updated global model parameters to each participant for the next round of local training. This process iterates until the model converges. Finally, the globally optimized inventory status data, optimized across entities, is output to the profit-oriented end-to-end integrated cost optimization module 6, enabling the global model to integrate the demand patterns and supply characteristics of all parties, improving the prediction accuracy of allocation decisions.

[0096] Specifically, the profit-oriented end-to-end integrated cost optimization module 6 includes a unified profit objective function construction submodule, a dynamic cost calculation submodule, a predictive loss calculation submodule, an opportunity cost quantification submodule, a multi-objective optimization submodule, and a deep reinforcement learning submodule. The unified profit objective function construction submodule is used to construct a mathematical model with maximizing total profit as the single core objective. The dynamic cost calculation submodule is used to calculate dynamic cost parameters based on real-time acquired oil price data, storage fee data, and commodity price data. The predictive loss calculation submodule is used to predict transportation loss rate and temporary storage loss rate based on real-time ambient temperature data, real-time ambient humidity data, transportation distance data, and packaging method data.

[0097] The opportunity cost quantification submodule is used to quantify the stockout cost caused by allocation delays or allocation failures. The multi-objective optimization submodule is used to solve for the Pareto optimal solution set using a multi-objective genetic algorithm. The deep reinforcement learning submodule is used to learn the optimal allocation strategy through continuous interaction with the supply chain environment. The unified profit objective function construction submodule is connected to the dynamic cost calculation submodule, the predictive loss calculation submodule, and the opportunity cost quantification submodule, respectively, and is used to receive dynamic cost parameters, predictive loss rate, and stockout cost as input parameters of the objective function. The multi-objective optimization submodule and the deep reinforcement learning submodule are connected to the unified profit objective function construction submodule, respectively, and are used to solve for the optimal allocation scheme based on the objective function.

[0098] In a specific embodiment of the present invention, the profit-oriented end-to-end integrated cost optimization module 6 constructs an end-to-end optimization framework with maximizing total profit as the single core objective. The unified profit objective function construction submodule constructs a mathematical model with maximizing total profit as the single core objective. This model uses the net value of sales revenue minus inventory holding costs, stockout costs, allocation costs, and predictive losses during transportation and temporary storage as the total profit. It sets constraints such as non-negative ending inventory, non-negative allocation volume, sales volume not exceeding available inventory, and total allocation volume not exceeding the transportation capacity limit, thereby achieving profit-oriented integrated cost optimization and dynamic optimization. The cost calculation submodule acquires real-time oil price data, storage fee data, and commodity price data through API interfaces. Based on this real-time data, it calculates dynamic cost parameters, including transportation costs fluctuating with oil prices, storage costs changing with market supply and demand, and holding costs changing due to commodity price fluctuations. This ensures that cost calculations reflect the real market environment. The predictive loss calculation submodule, based on real-time ambient temperature and humidity data acquired from the real-time sensing and dynamic response module 4, combined with transportation distance and packaging method data, uses a pre-trained loss prediction model to predict transportation loss rates and temporary storage loss rates, thus quantifying loss factors. The opportunity cost quantification submodule incorporates cost parameters into the objective function, analyzing the correlation between historical stockout events and sales losses to quantify stockout costs caused by delayed or failed allocations, including lost sales revenue, customer churn, and reputational damage. The multi-objective optimization submodule employs a multi-objective genetic algorithm, prioritizing total profit maximization while considering multiple optimization objectives such as allocation timeliness and inventory turnover, to find the Pareto optimal solution set, providing decision-makers with multiple allocation options. The deep reinforcement learning submodule constructs a supply chain environment simulator and uses deep reinforcement learning algorithms to simulate the supply chain environment. Continuous interaction allows the system to learn the optimal allocation strategy through a trial-and-error mechanism and update the strategy network parameters online as the environment changes. This enables the allocation strategy to adapt dynamically to changes in the market environment and supply chain status. The unified profit objective function construction submodule is connected to the dynamic cost calculation submodule, the predictive loss calculation submodule, and the opportunity cost quantification submodule, respectively, and receives dynamic cost parameters, predictive loss rate, and stockout cost as input parameters for the objective function. The multi-objective optimization submodule and the deep reinforcement learning submodule are connected to the unified profit objective function construction submodule, respectively, and solve for the optimal allocation scheme based on the objective function.

[0099] Specifically, the mathematical model constructed by the unified profit objective function construction submodule is as follows:

[0100]

[0101] Where represents total profit, N represents the total number of warehouse nodes, and T represents the total number of decision cycles.

[0102] and Indicates the warehouse node number. Indicates the decision-making cycle.

[0103] Represents a warehouse node In the decision-making cycle The unit price of goods sold.

[0104] Represents a warehouse node In the decision-making cycle Sales volume of goods.

[0105] Represents a warehouse node In the decision-making cycle The unit inventory holding cost.

[0106] This represents the ending inventory of warehouse node i during decision period t.

[0107] Represents a warehouse node In the decision-making cycle The unit cost of out-of-stock items.

[0108] Represents a warehouse node In the decision-making cycle The amount of stock shortage.

[0109] This indicates that during decision cycle t, a unit of goods is transferred from warehouse node i to warehouse node i.

[0110] The allocation cost of j.

[0111] This represents the quantity of goods transferred from warehouse node i to warehouse node j during decision cycle t.

[0112] This represents the predicted loss rate of a unit of goods shipped from warehouse node i during transportation in decision period t.

[0113] This represents the predicted loss rate of temporary storage of goods per unit at warehouse node i during decision period t.

[0114] The mathematical model satisfies the following constraints:

[0115]

[0116]

[0117]

[0118] in, This represents the upper limit of the total allocation capacity during the decision-making period t.

[0119] In a specific embodiment of the present invention, the first constraint is an inventory balance constraint, which means that the ending inventory of the current period equals the ending inventory of the previous period plus the amount transferred in from other warehouses minus the amount transferred out minus the sales volume minus transportation losses and temporary storage losses; the second constraint is a non-negativity constraint and an availability constraint, ensuring that the inventory level is non-negative, the transfer volume is non-negative, and the sales volume does not exceed the available inventory; the third constraint is a capacity limit constraint, ensuring that the total transfer volume between all warehouses does not exceed the total transfer capacity limit for the current period.

[0120] Specifically, the multi-objective optimization submodule uses a non-dominated sorting genetic algorithm to solve for the Pareto optimal solution set. The non-dominated sorting genetic algorithm includes a population initialization unit 1, a non-dominated sorting unit 2, a crowding calculation unit 3, a selection unit 4, a crossover unit 5, a mutation unit 6, and an iteration termination determination unit 7. The population initialization unit 1 is used to generate an initial population containing individuals with multiple allocation schemes. The non-dominated sorting unit 2 is used to sort the individuals in the population according to the Pareto non-dominated hierarchy. The crowding calculation unit 3 is used to calculate the crowding distance between individuals within the same non-dominated hierarchy. The selection unit 4 is used to select parent individuals based on the non-dominated hierarchy and the crowding distance. The crossover unit 5 and the mutation unit 6 are used to perform crossover and mutation operations on the parent individuals to generate offspring individuals. The iteration termination determination unit 7 is used to output the Pareto optimal solution set when the preset maximum number of iterations is reached.

[0121] In a specific embodiment of the present invention, the multi-objective optimization submodule uses a non-dominated sorting genetic algorithm to solve for the Pareto optimal solution set. Population initialization unit 1 randomly generates an initial population containing individuals with multiple allocation schemes. Each individual is encoded as a combination of an allocation amount matrix and an allocation path. Non-dominated sorting unit 2 performs Pareto non-dominated hierarchy sorting on the individuals in the population, classifying individuals not dominated by any other individual into the first non-dominated hierarchy, then classifying the remaining individuals not dominated by any other individual into the second non-dominated hierarchy, and so on. Here, individual A dominating individual B means that individual A is not inferior to individual B in all objectives and is superior to individual B in at least one objective. Crowding calculation unit 3 calculates the same non-dominated level... The crowding distance of individuals within the same hierarchy is used to determine the sparser the individual is in the target space, and the higher its retention value. Selection unit 4 selects parent individuals based on the non-dominated hierarchy and crowding distance using a tournament selection strategy, prioritizing individuals with high non-dominated hierarchies and large crowding distances. Crossover unit 5 and mutation unit 6 perform simulated binary crossover and polynomial mutation operations on parent individuals to generate offspring individuals, introducing new gene combinations to explore unsearched regions in the solution space. Iteration termination judgment unit 7 monitors the population convergence status in real time during the iteration process. It terminates the iteration when the preset maximum number of iterations is reached or the optimal frontier of the population no longer changes over multiple generations, and outputs a Pareto optimal solution set for decision-makers to choose from.

[0122] Specifically, the Human-Computer Interaction and Explainability Module 5 includes a large language model interaction submodule, an explainable artificial intelligence submodule, and a human-computer collaborative decision-making submodule. The large language model interaction submodule receives natural language queries input by users, performs intent recognition and semantic parsing on the queries, generates a query intent vector, and sends the query intent vector to the decentralized allocation management module 1 and the profit-oriented end-to-end integrated cost optimization module 6. The explainable artificial intelligence submodule generates a list of key influencing factors and confidence scores for allocation decisions using attention visualization, counterfactual explanation, and SHAP value analysis methods, and sends the list of key influencing factors and confidence scores to the large language model interaction submodule. The large language model interaction submodule also generates a natural language explainability report based on the list of key influencing factors and confidence scores and displays it to the user. The human-computer collaborative decision-making submodule receives user instructions to adjust the allocation plan and sends the adjusted allocation plan as a feedback signal to the profit-oriented end-to-end integrated cost optimization module 6 for model optimization.

[0123] In a specific embodiment of the present invention, the human-computer interaction and explainability module 5 constructs a natural language interaction and decision-making explainability interface for business personnel. The large language model interaction submodule receives natural language query statements input by users, such as "Why is it recommended to transfer from the Shanghai warehouse to the Beijing warehouse?" or "How will the transfer plan change if oil prices rise by %?". It performs intent recognition and semantic parsing through a pre-trained large language model to generate a query intent vector, which is then sent to the decentralized transfer management module 1 and the profit-oriented end-to-end integrated cost optimization module 6, triggering the corresponding data query and explanation generation process. The explainable artificial intelligence submodule uses attention visualization methods to demonstrate the degree of attention the decision model pays to each input feature when making transfer decisions. It generates explanatory instances of "how the decision result would change if a certain condition were changed" using counterfactual explanation methods, and quantifies the direction and magnitude of the contribution of each input feature to the final transfer decision using SHAP value analysis. The above methods are used to generate a list of key influencing factors and confidence scores for allocation decisions. This list and scores are then sent to the large language model interaction submodule. Based on these factors, the large language model interaction submodule utilizes its text generation capabilities to transform the technical decision logic into a natural language interpretable report easily understood by business personnel. The report includes the reasons for the allocation plan, the ranking of the contributions of each influencing factor, the recommended allocation quantity and timeframe, and potential risk warnings, which are then displayed to the user. The human-machine collaborative decision-making submodule receives user instructions to adjust the allocation plan, such as manually adjusting the allocation quantity, locking or unlocking allocation permissions for specific warehouses, and setting priority rules. The adjusted allocation plan is then sent as feedback to the profit-oriented end-to-end integrated cost optimization module 6 for model optimization. This allows the system to incorporate the domain knowledge of business experts into the decision-making process, achieving closed-loop optimization through human-machine collaboration.

[0124] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program implementing the operation method of the intelligent inventory transfer system based on multi-warehouse collaboration as described above; a computer-readable storage medium storing a computer program, and the computer program being executed by a processor implementing the operation method of the intelligent inventory transfer system based on multi-warehouse collaboration as described above.

[0125] In a specific embodiment of the present invention, the present invention constructs a full-link intelligent inventory allocation system architecture that includes multi-source heterogeneous data acquisition and fusion, federated learning-driven cross-subject predictive modeling, real-time perception-driven dynamic response, digital twin simulation verification, profit-oriented end-to-end comprehensive cost optimization, blockchain smart contract execution, and visualized human-computer interaction. The modules form a closely coordinated closed-loop decision-making link, which together realizes efficient, accurate, interpretable, and secure multi-warehouse collaborative allocation.

[0126] Working principle:

[0127] After the system starts, the multi-source heterogeneous data fusion module 2 begins operation first. Its social media data acquisition submodule captures real-time data on user discussions, consumption trends, and trending events on social media platforms via API interfaces; the weather and seasonal data acquisition submodule obtains data on temperature, precipitation, and seasonal variations for each warehouse's location from meteorological service interfaces; the macroeconomic indicator acquisition submodule collects macroeconomic data such as GDP growth rate, consumer price index, and consumer confidence index; the competitor dynamics acquisition submodule monitors competitors' promotional activities, new product launches, and price changes; and the upstream supply chain information acquisition submodule obtains upstream information such as supplier capacity, raw material prices, and logistics status. The data collected by each of these submodules is output to the multimodal fusion submodule, which performs unified modeling of text data, image data, and structured data to generate multi-dimensional fused demand forecast input data, which is then transmitted to the federated learning-driven module 3.

[0128] After receiving the fused training data, the Federated Learning Driven Module 3 initiates a cross-entity federated learning training process. Local training submodules distributed across each participant train local prediction models on their respective local devices based on their local inventory, sales, and demand data. During training, the original data remains locally; only model gradient parameters are generated. The encrypted gradient generation submodule encrypts the training gradients of the local prediction models using homomorphic encryption technology, generates encrypted gradient parameters, and uploads them to the gradient aggregation submodule via a secure communication channel. The gradient aggregation submodule aggregates the encrypted gradient parameters uploaded by each participant and generates an aggregated gradient using a federated averaging algorithm. The global model update submodule updates the global inventory prediction model based on the aggregated gradient and distributes the updated global model parameters to each participant for the next round of local training. After multiple rounds of iterative training until the model converges, the Federated Learning Driven Module 3 outputs the cross-entity collaboratively optimized global inventory status data to the profit-oriented end-to-end integrated cost optimization module 6.

[0129] While federated learning and training are underway, the real-time perception and dynamic response module 4 continues to operate. The IoT perception submodule deployed at each warehouse node collects real-time environmental data and inventory physical status within the warehouse using temperature and humidity sensors, weight sensors, and motion sensors. The RFID tag reading submodule tracks inventory inflows and outflows and product flow in real-time by reading RFID tags attached to goods. The inventory change data, equipment status data, and personnel efficiency data collected by the IoT perception submodule and the RFID tag reading submodule are transmitted in real-time to the edge computing submodule deployed on the local edge server at each warehouse node. The edge computing submodule performs millisecond-level real-time processing on the above data, identifies abnormal inventory changes and emergency transfer needs through preset event-driven rules and machine learning models, generates event-driven transfer trigger signals, and sends these signals in real-time to the decentralized transfer management module 1, while simultaneously sending real-time status data to the digital twin simulation module 7.

[0130] After receiving real-time status data, the digital twin simulation module 7 constructs a supply chain digital twin model based on the real-time status data. It then simulates and extrapolates the candidate allocation strategies generated by the profit-oriented end-to-end integrated cost optimization module 6, evaluates the performance of each strategy under different scenarios, including the robustness of allocation strategies under abnormal scenarios such as sudden demand shocks, transportation delays, and supplier supply disruptions. The simulation evaluation results are then fed back to the profit-oriented end-to-end integrated cost optimization module 6 as an important reference for optimizing the allocation plan.

[0131] The profit-oriented end-to-end integrated cost optimization module 6 initiates the optimization calculation process after receiving global inventory status data from the federated learning-driven module 3 and simulation evaluation results from the digital twin simulation module 7. The dynamic cost calculation submodule acquires real-time oil price data, storage fee data, and commodity price data via API interfaces, and calculates dynamic cost parameters based on this real-time data. The predictive loss calculation submodule uses real-time ambient temperature and humidity data obtained from the real-time sensing and dynamic response module 4, combined with transportation distance and packaging method data, to predict transportation loss rates and temporary storage loss rates using a pre-trained loss prediction model. The opportunity cost quantification submodule quantifies the stockout costs caused by allocation delays or failures by analyzing the correlation between historical stockout events and sales losses. The unified profit objective function construction submodule receives dynamic cost parameters, predictive loss rate, and stockout cost as input parameters for the objective function. It constructs a mathematical model with maximizing total profit as the single core objective. This model uses the net value after subtracting inventory holding costs, stockout costs, allocation costs, and predictive losses during transportation and temporary storage from sales revenue as the total profit. Constraints are set such as non-negative ending inventory, non-negative allocation volume, sales volume not exceeding available inventory, and total allocation volume not exceeding the transportation capacity limit. The multi-objective optimization submodule employs a non-dominated sorting genetic algorithm, with maximizing total profit as the primary objective while also considering multiple optimization objectives such as allocation timeliness and inventory turnover rate. It solves for the Pareto optimal solution set through steps including population initialization, non-dominated sorting, congestion calculation, selection, crossover, mutation, and iteration termination determination. The deep reinforcement learning submodule constructs a supply chain environment simulator and uses a deep reinforcement learning algorithm to continuously interact with the supply chain environment, learning the optimal allocation strategy through a trial-and-error mechanism. After the multi-objective optimization submodule and the deep reinforcement learning submodule solve the allocation scheme based on the objective function, the profit-oriented end-to-end comprehensive cost optimization module 6 outputs the most profitable allocation scheme to the decentralized allocation management module 1.

[0132] Upon receiving the allocation plan, the decentralized allocation management module 1 initiates a decentralized consensus and execution process. The blockchain consensus submodule, distributed across various warehouse nodes, verifies and confirms the allocation plan using a Byzantine fault-tolerant consensus mechanism. Each warehouse node autonomously reaches an allocation agreement based on this consensus mechanism, without relying on a central scheduling server. After the allocation plan is confirmed by the blockchain consensus submodule, the smart contract execution submodule automatically performs allocation condition checks, verifying whether each warehouse's inventory meets the transfer conditions, whether the receiving warehouse has sufficient storage space, and whether transportation resources are available. Once approved, allocation instructions are automatically generated based on allocation priority. The decentralized allocation management module 1 then sends the allocation plan and allocation instructions to the blockchain smart contract execution module 8.

[0133] After receiving the allocation plan and allocation instructions, the blockchain smart contract execution module 8 sends the allocation instructions to the execution terminals of each warehouse node. Each warehouse execution terminal completes the picking, outbound, loading and shipping operations of the goods according to the instructions. After the execution is completed, the execution confirmation information is fed back to the decentralized allocation management module 1 through the blockchain smart contract execution module 8, thus completing the allocation execution closed loop.

[0134] Throughout the workflow, the Human-Computer Interaction and Explainability Module 5 continuously interacts with the user. The Large Language Model Interaction Submodule receives the user's natural language query, performs intent recognition and semantic parsing using a pre-trained large language model, generates a query intent vector, and sends it to the Decentralized Allocation Management Module 1 and the Profit-Oriented End-to-End Integrated Cost Optimization Module 6. The Explainable Artificial Intelligence Submodule generates a list of key influencing factors and confidence scores for allocation decisions using attention visualization, counterfactual explanation, and SHAP value analysis methods, and sends these to the Large Language Model Interaction Submodule. Based on the list of key influencing factors and confidence scores, the Large Language Model Interaction Submodule generates a natural language explainability report and displays it to the user. The Human-Computer Collaborative Decision-Making Submodule receives the user's adjustment instructions for the allocation plan and sends the adjusted plan as a feedback signal to the Profit-Oriented End-to-End Integrated Cost Optimization Module 6 for model optimization, enabling the system to incorporate the domain knowledge of business experts into the decision-making process. At least one processor executes a computer program stored in a non-volatile computer-readable storage medium to implement the functions of all the above modules. It achieves high-speed reading, writing and transmission of data through a data bus, and connects to the execution terminals of each warehouse node through a communication interface to send allocation instructions and receive execution confirmation information. This constructs a closed-loop workflow for intelligent inventory allocation across the entire chain, from data acquisition, cross-entity modeling, real-time perception, simulation verification, comprehensive optimization, decentralized consensus execution to visual human-computer interaction.

[0135] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent inventory transfer system based on multi-warehouse collaboration, characterized in that, include: The module includes a decentralized allocation management module (1), a multi-source heterogeneous data fusion module (2), a federated learning-driven module (3), a real-time perception and dynamic response module (4), a human-computer interaction and interpretability module (5), a profit-oriented end-to-end integrated cost optimization module (6), a digital twin simulation module (7), and a blockchain smart contract execution module (8). The decentralized allocation management module (1) is connected to the multi-source heterogeneous data fusion module (2), the federated learning-driven module (3), the real-time perception and dynamic response module (4), the profit-oriented end-to-end comprehensive cost optimization module (6), the digital twin simulation module (7), and the blockchain smart contract execution module (8) respectively, and is used to receive the data output by each module and generate global allocation decisions; The output of the multi-source heterogeneous data fusion module (2) is connected to the input of the federated learning driver module (3) to provide the federated learning driver module (3) with fused multi-dimensional training data. The output of the federated learning-driven module (3) is connected to the input of the profit-oriented end-to-end integrated cost optimization module (6) to provide the profit-oriented end-to-end integrated cost optimization module (6) with global inventory status data after cross-entity collaborative optimization. The output of the real-time perception and dynamic response module (4) is connected to the input of the decentralized allocation management module (1) and the input of the digital twin simulation module (7), respectively, to provide real-time inventory perception data to the decentralized allocation management module (1) and to provide real-time status data to the digital twin simulation module (7) to drive simulation. The output of the digital twin simulation module (7) is connected to the input of the profit-oriented end-to-end integrated cost optimization module (6) and is used to provide the simulation evaluation results of the candidate allocation strategy to the profit-oriented end-to-end integrated cost optimization module (6). The output of the profit-oriented end-to-end integrated cost optimization module (6) is connected to the input of the decentralized allocation management module (1) and is used to provide the decentralized allocation management module (1) with the most profitable allocation scheme. The output of the decentralized allocation management module (1) is connected to the input of the blockchain smart contract execution module (8) to send the allocation plan to the blockchain smart contract execution module (8). The output of the blockchain smart contract execution module (8) is connected to the execution terminal of each warehouse node, and is used to send the allocation instruction to each warehouse node for execution; The human-computer interaction and interpretability module (5) is bidirectionally connected to the decentralized allocation management module (1) and the profit-oriented end-to-end integrated cost optimization module (6), respectively, and is used to receive users' natural language queries and display the interpretability report of allocation decisions to users.

2. The intelligent inventory transfer system based on multi-warehouse collaboration according to claim 1, characterized in that, The decentralized allocation management module (1) includes a blockchain consensus submodule and a smart contract execution submodule distributed across each warehouse node; The blockchain consensus submodule is used to verify and confirm the allocation plan among the warehouse nodes through the Byzantine fault-tolerant consensus mechanism. Each warehouse node independently reaches an allocation agreement based on the consensus mechanism, without the need for a central scheduling server. The smart contract execution submodule is used to automatically perform transfer condition checks, transfer priority sorting, and transfer instruction issuance operations after the transfer plan is confirmed by the blockchain consensus submodule.

3. The intelligent inventory transfer system based on multi-warehouse collaboration according to claim 1, characterized in that, The multi-source heterogeneous data fusion module (2) includes a social media data collection submodule, a weather and seasonal data collection submodule, a macroeconomic indicator collection submodule, a competitor dynamics collection submodule, a supply chain upstream information collection submodule, and a multimodal fusion submodule; The outputs of the social media data collection submodule, the weather and seasonal data collection submodule, the macroeconomic indicator collection submodule, the competitor dynamics collection submodule, and the upstream supply chain information collection submodule are respectively connected to the input of the multimodal fusion submodule; The multimodal fusion submodule is used to uniformly model text data, image data, and structured data to generate multi-dimensional fused demand prediction input data.

4. The intelligent inventory transfer system based on multi-warehouse collaboration according to claim 1, characterized in that, The real-time sensing and dynamic response module (4) includes an IoT sensing submodule, an RFID tag reading submodule, and an edge computing submodule; The IoT sensing submodule and the RFID tag reading submodule are deployed at each warehouse node to collect inventory change data, equipment status data and personnel efficiency data in the warehouse. The edge computing submodule is deployed on the local edge server of each warehouse node. It is used to perform millisecond-level real-time processing on the inventory change data, the equipment status data and the personnel efficiency data, generate event-driven transfer trigger signals, and send the transfer trigger signals to the decentralized transfer management module (1).

5. The intelligent inventory transfer system based on multi-warehouse collaboration according to claim 1, characterized in that, The federated learning-driven module (3) includes a local training submodule distributed among the participants, an encrypted gradient generation submodule, a gradient aggregation submodule, and a global model update submodule; The local training submodules of each participant are used to train local prediction models based on local inventory data; The encrypted gradient generation submodule is used to encrypt the training gradient of the local prediction model and generate encrypted gradient parameters. The gradient aggregation submodule is used to aggregate the encrypted gradient parameters uploaded by each participant to generate an aggregated gradient; The global model update submodule is used to update the global inventory prediction model based on the aggregated gradient, and distribute the updated global model parameters to each participant for the next round of local training.

6. The intelligent inventory transfer system based on multi-warehouse collaboration according to claim 1, characterized in that, The profit-oriented end-to-end integrated cost optimization module (6) includes a unified profit objective function construction submodule, a dynamic cost calculation submodule, a predictive loss calculation submodule, an opportunity cost quantification submodule, a multi-objective optimization submodule, and a deep reinforcement learning submodule; The unified profit objective function construction submodule is used to construct a mathematical model with maximizing total profit as the single core objective; The dynamic cost calculation submodule is used to calculate dynamic cost parameters based on real-time acquired oil price data, warehousing fee data, and commodity price data; The predictive loss calculation submodule is used to predict the transportation loss rate and temporary storage loss rate based on real-time ambient temperature data, real-time ambient humidity data, transportation distance data, and packaging method data. The opportunity cost quantification submodule is used to quantify the stockout costs caused by allocation delays or allocation failures. The multi-objective optimization submodule is used to solve for the Pareto optimal solution set using a multi-objective genetic algorithm; The deep reinforcement learning submodule is used to learn the optimal allocation strategy through continuous interaction with the supply chain environment. The unified profit objective function construction submodule is connected to the dynamic cost calculation submodule, the predictive loss calculation submodule, and the opportunity cost quantification submodule, respectively, and is used to receive the dynamic cost parameters, the predictive loss rate, and the stockout cost as input parameters of the objective function; The multi-objective optimization submodule and the deep reinforcement learning submodule are respectively connected to the unified profit objective function construction submodule, and are used to solve the optimal allocation scheme based on the objective function.

7. The intelligent inventory transfer system based on multi-warehouse collaboration according to claim 6, characterized in that, The mathematical model constructed by the unified profit objective function construction submodule is as follows: Where represents total profit, N represents the total number of warehouse nodes, and T represents the total number of decision cycles. and Indicates the warehouse node number. Indicates the decision-making cycle; Represents a warehouse node In the decision-making cycle The unit sales price of goods; Represents a warehouse node In the decision-making cycle Sales volume of goods; Represents a warehouse node In the decision-making cycle The unit inventory holding cost; This represents the ending inventory of warehouse node i during decision period t. Represents a warehouse node In the decision-making cycle The unit cost of out-of-stock items; Represents a warehouse node In the decision-making cycle The amount of stock shortage; This indicates that during decision cycle t, a unit of goods is transferred from warehouse node i to warehouse node i. The allocation cost of j; This represents the quantity of goods transferred from warehouse node i to warehouse node j during decision cycle t; This represents the predicted loss rate of a unit of goods shipped from warehouse node i during transportation in decision period t. This represents the predicted loss rate per unit of goods temporarily stored at warehouse node i during decision period t. The mathematical model satisfies the following constraints: in, This represents the upper limit of the total allocation capacity during the decision-making period t.

8. The intelligent inventory transfer system based on multi-warehouse collaboration according to claim 6, characterized in that, The multi-objective optimization submodule uses a non-dominated sorting genetic algorithm to solve the Pareto optimal solution set. The non-dominated sorting genetic algorithm includes a population initialization unit, a non-dominated sorting unit, a crowding calculation unit, a selection unit, a crossover unit, a mutation unit, and an iteration termination determination unit. The population initialization unit is used to generate an initial population containing multiple allocation scheme individuals; The non-dominated sorting unit is used to perform Pareto non-dominated hierarchical sorting of individuals in the population. The crowding calculation unit is used to calculate the crowding distance between individuals within the same non-dominated hierarchy; The selection unit is used to select parent individuals based on the non-dominated hierarchy and the crowding distance; The crossover unit and the mutation unit are used to perform crossover and mutation operations on the parent individual to generate offspring individuals; The iteration termination determination unit is used to output the Pareto optimal solution set when the preset maximum number of iterations is reached.

9. The intelligent inventory transfer system based on multi-warehouse collaboration according to claim 1, characterized in that, The human-computer interaction and interpretability module (5) includes a large language model interaction submodule, an interpretable artificial intelligence submodule, and a human-computer collaborative decision-making submodule; The large language model interaction submodule is used to receive natural language query statements input by users, perform intent recognition and semantic parsing on the natural language query statements, generate query intent vectors, and send the query intent vectors to the decentralized allocation management module (1) and the profit-oriented end-to-end integrated cost optimization module (6). The explainable artificial intelligence submodule is used to generate a list of key influencing factors and a confidence score for allocation decisions through attention visualization, counterfactual explanation, and SHAP value analysis, and to send the list of key influencing factors and the confidence score to the large language model interaction submodule. The large language model interaction submodule is also used to generate a natural language interpretability report based on the list of key influencing factors and the confidence score, and to display it to the user; The human-machine collaborative decision-making submodule is used to receive the user's adjustment instruction for the allocation plan and send the adjusted allocation plan as a feedback signal to the profit-oriented end-to-end integrated cost optimization module (6) for model optimization.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the operation method of the intelligent inventory transfer system based on multi-warehouse collaboration as described in any one of claims 1 to 9; A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the operation method of the intelligent inventory transfer system based on multi-warehouse collaboration as described in any one of claims 1 to 9.

Images