Agent-driven supply chain data collaboration system

Abstract

This application discloses an agent-driven supply chain data collaboration system, relating to the supply chain field. The agent-driven supply chain data collaboration system includes a data interface module, a collaboration module, and a feedback module. The data interface module interfaces with the management systems of each participant in the supply chain to obtain local environmental information of each participant. The participants in the supply chain include demanders and suppliers. The local environmental information of demanders includes their material inventory status and material demand information, while the local environmental information of suppliers includes their material supply capacity status. The collaboration module inputs the local environmental information of each target participant associated with the material demand information into a collaborative decision-making model. The collaborative decision-making model makes decisions based on preset optimization objectives and outputs a material scheduling recommendation scheme for each target participant. The feedback module feeds back the material scheduling recommendation scheme to the management systems of the corresponding participants.

Description

Technical Field

[0001] This application relates to the field of supply chain technology, and in particular to an agent-driven supply chain data collaboration system. Background Technology

[0002] Supply chain management aims to meet the product needs of end users and reduce the overall costs of all participants in the supply chain by optimizing links such as suppliers, manufacturers, warehousing and logistics.

[0003] Currently, most participants in the supply chain adopt a "point-to-point" data exchange model. For example, the management systems of suppliers and manufacturers are independent of each other, improving efficiency only within their respective business scopes. However, in this approach, data from different management systems cannot be synchronized in real time, making it difficult for upstream and downstream partners in the supply chain to form collaborative work processes, resulting in inefficient resource allocation. Summary of the Invention

[0004] This application provides an agent-driven supply chain data collaboration system, which at least addresses the problem of inefficient supply chain resource allocation caused by the difficulty in coordinating upstream and downstream supply chains in related solutions.

[0005] In a first aspect, this application provides an agent-driven supply chain data collaboration system, the agent-driven supply chain data collaboration system comprising: The data interface module is used to connect to the management systems of each participant in the supply chain, so as to obtain the local environmental information of each participant through the management systems of each participant; wherein, the participants in the supply chain include demanders and suppliers, the local environmental information of the demanders includes their material inventory status and material demand information, and the local environmental information of the suppliers includes their material supply capacity status, distance information from the demanders, and fulfillment progress information. The collaboration module is used to input the local environmental information of each target participant associated with the material demand information into the collaborative decision model, which makes a decision based on the preset optimization objective and outputs the material scheduling recommendation scheme corresponding to each target participant. The feedback module is used to feed back the recommended material scheduling scheme to the management system of the relevant participants.

[0006] In a second aspect, this application provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the energy-driven supply chain data collaboration system as provided in the first aspect.

[0007] The agent-driven supply chain data collaboration system provided in this application has at least the following beneficial effects: By connecting to the management systems of all participating parties through a data interface module, real-time synchronization of manufacturer demand data, transportation data, and supplier capacity and inventory data is achieved. This changes the previous situation where each system operated independently and data could not be shared, enabling collaborative work processes across the supply chain and solving the inefficiency caused by data asynchrony in the traditional model. Furthermore, the collaborative decision-making model, based on global optimization objectives and considering the capacity and order fulfillment status of each supplier, outputs production resource allocation ratios and procurement and logistics scheduling plans. This avoids the blind allocation of resources in the traditional model and effectively prevents supplier capacity waste and unreasonable production resource allocation. Attached Figure Description

[0008] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0009] Figure 1 This is a structural block diagram of the agent-driven supply chain data collaboration system provided in the embodiments of this application; Figure 2 This is an architecture diagram of the collaborative decision-making model provided in the embodiments of this application; Figure 3 This is a structural block diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0010] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0011] Please see Figure 1 , Figure 1 The intelligent agent-driven supply chain data collaboration system provided in this application includes a data interface module 11, a collaboration module 12, and a feedback module 13.

[0012] Data interface module 11 is used to connect to the management systems of each participant in the supply chain, so as to obtain the local environmental information of each participant through the management systems of each participant; wherein, the participants in the supply chain include demanders and suppliers, the local environmental information of demanders includes their material inventory status and material demand information, and the local environmental information of suppliers includes their material supply capacity status.

[0013] The collaboration module 12 is used to input the local environmental information of each target participant associated with the material demand information into the collaborative decision-making model. The collaborative decision-making model makes decisions based on the preset optimization objectives and outputs the material scheduling recommendation scheme corresponding to each target participant.

[0014] Feedback module 13 is used to feed back the recommended material scheduling plan to the management system of the relevant participants.

[0015] In one alternative implementation, the demand side includes manufacturers, and the supply side includes suppliers. Suppliers provide various materials to manufacturers for product production. All manufacturers and suppliers on the system have established formal cooperative relationships.

[0016] In one optional implementation, the manufacturer's material inventory status includes the current inventory status of various materials, the amount in transit, and the safety stock threshold. The manufacturer's material demand information includes the demand for various materials based on the production plan, the demand time nodes (i.e., the time when materials need to be delivered to the production line), and priority information (the importance and urgency of production of the same material / different batches of the same material).

[0017] In an optional implementation, the supplier's local environmental information includes, but is not limited to: the current inventory level of the materials supplied by the supplier, the current production capacity status, in-production order information, the distance to the manufacturer, and the historical order fulfillment on-time rate.

[0018] The production capacity status is used to define the load factor, available capacity, and maximum output of a supplier's production line. For example, a battery production line is currently operating at 85% capacity and has additional capacity to produce 200 battery packs within the next three days.

[0019] Among them, the in-production order information refers to other customer orders that the supplier has accepted but not yet delivered. The in-production order information includes the undelivered material production quantity and the expected completion time of the order.

[0020] The distance between the supplier and the manufacturer refers to the distance between the production line of the materials supplied by the supplier and the production line of the manufacturer, which is used to assess logistics costs and time.

[0021] Historical order fulfillment on-time rate is used to evaluate the on-time delivery rate of orders by suppliers within a preset historical time period, and is used to measure the reliability of suppliers.

[0022] In an optional implementation, the collaborative decision-making model employs a multi-agent reinforcement learning framework, where each participant corresponds to an agent. Through joint training and learning, a global optimization objective is achieved while satisfying local constraints, thereby determining the recommended material scheduling scheme for each participant. Specifically, for the manufacturer, the recommended material scheduling scheme is the suggested procurement ratio of each material from different suppliers; for the supplier, the recommended material scheduling scheme is the suggested allocation ratio of production resources for different materials.

[0023] The optimization objective can be flexibly configured according to actual business needs. It can be to minimize the total logistics and inventory costs of the supply chain or to maximize the order fulfillment rate. This application does not limit this.

[0024] In an optional implementation, the feedback module 13 feeds back the recommended material scheduling plan to the management system of the corresponding participants, so that each participant can make decisions on material procurement and production based on the corresponding recommended material scheduling plan. Each participant can either execute the recommended material scheduling plan suggested by the system or make adjustments based on the corresponding recommended material scheduling plan.

[0025] For example, suppose manufacturer A plans to produce 100 electric vehicles, requiring 1000 batteries, 1000 motors, and 1000 chips. In practice, the data interface module retrieves data from manufacturer A's management system, including its material inventory. Assume the inventory of batteries is 100, motors 180, and chips 320, with a safety stock threshold of 100 for each. If 200 chips are due to arrive in the next two days, the actual required number of batteries is calculated as 1000 - 100 + 100 = 1000, and the actual required number of chips is 1000 - 180 + 100 - 200 = 720, and so on. Assuming batteries are the highest priority and must arrive on the third day of the following week, the module also collects local environmental information such as inventory and production capacity from various suppliers.

[0026] The collaboration module inputs all the above-mentioned related data into the collaborative decision-making model. The model performs decision calculations based on the optimization objective and outputs a recommended material scheduling solution.

[0027] Assume the manufacturer's recommended material scheduling plan is as follows: 80% of the chips will be purchased from supplier A (780 × 80% = 624 chips), 20% will be purchased from supplier B (780 × 20% = 156 chips), and all batteries will be supplied by supplier C (1000 chips).

[0028] The supplier's recommended material scheduling solution is as follows:

[0029] Supplier A: Total production capacity is 100 chips per day, and can also handle the production of chip-related components. This order includes 624 chips (production must be completed by December 28th, with 2 days allowed for shipping). Based on capacity utilization and order priority, 85% of production resources will be allocated to chip production, and the remaining 15% will be allocated to the production of chip-related components.

[0030] Supplier B: Has a small production capacity of 20 chips per day and no other related material production business. This order includes 156 chips (to be completed by December 28th). To ensure timely fulfillment, 100% of production resources will be allocated to chip production.

[0031] Supplier C: Its core business is battery production, and it can also produce battery protective cases. The total production capacity is 150 batteries per day. This time, it has accepted an order for 1,000 batteries (highest priority, production must be completed before December 25th, with 2 days allowed for transportation). Therefore, 90% of the production resources will be allocated to battery production, and the remaining 10% will be allocated to battery protective case production.

[0032] The feedback module will push the above material scheduling recommendations to the management systems of each supplier and manufacturer, and each participant can make adjustments according to their actual production capacity.

[0033] The aforementioned solution, through a data interface module connecting to the management systems of all participating parties, achieves real-time synchronization of manufacturer demand data, transportation data, and supplier capacity and inventory data. This completely changes the previous situation where each system operated independently and data could not be shared, enabling collaborative work processes across the supply chain and solving the inefficiency caused by data asynchrony in the traditional model. Furthermore, the collaborative decision-making model, based on global optimization objectives and considering the capacity and order fulfillment status of each supplier, outputs production resource allocation ratios and procurement and logistics scheduling plans. This avoids the blind allocation of resources in the traditional model and effectively prevents supplier capacity waste and unreasonable production resource allocation.

[0034] In an alternative implementation, please refer to Figure 2 , Figure 2 This is a structural diagram of the collaborative decision-making model provided in this application embodiment. The collaborative decision-making model includes a hierarchical encoder and a decision network. The hierarchical encoder includes a local encoding layer, an entity attention layer, and a fusion layer. The collaborative module is specifically used for:

[0035] The local environment information of each target participant is input into the hierarchical encoder. The local environment representation of each target participant is obtained through the local encoding layer. The entity attention layer calculates the interaction relationship between the target participants based on the multi-head self-attention mechanism, and obtains the context representation of each target participant that incorporates information from other target participants.

[0036] The fusion layer fuses the context representations and local environment representations of each target participant to obtain the final state representation of each target participant. The final state representation of each target participant is input into the decision network, which then makes decisions based on the optimization objective to obtain the recommended material scheduling scheme for each target participant.

[0037] like Figure 2 As shown, the local coding layer and the entity attention layer receive local environmental information O1, ..., from agent 1 (i.e., target participant 1) and local environmental information O from agent |N| (i.e., target participant 2). |N| .

[0038] Optionally, the local encoding layer includes a material attention subnetwork and a non-material feature processing subnetwork. The material attention subnetwork performs multi-head self-attention computation on the material features of the target participant (such as the demand and priority of each material) to obtain a material aggregate representation; the non-material feature processing subnetwork processes non-material features (such as inventory level, capacity, location, etc.) through a fully connected layer. The outputs of the material aggregate representation and the non-material feature processing subnetwork are fused to obtain the local environment representation of the agent.

[0039] The entity attention layer calculates the attention weights between target participants through multi-head self-attention, thereby identifying the interaction relationships between target participants and finally outputting the context representation of each target participant. The context representation of each target participant incorporates the association information of other relevant participants.

[0040] The fusion layer fuses the context representation of each target participant with its local environment representation to obtain the final state representation of each target participant: T(O1),...,T(O2). |N| ), where T(O1) represents the final state representation of agent 1, T(O |N| Let represent the final state of agent |N|. The fusion layer uses a fully connected network.

[0041] Next, the final state representations of each of the aforementioned target participants are input into the decision network. The decision network performs parallel computation and optimization based on a preset global optimization objective (such as minimizing the total supply chain cost), and finally outputs a recommended material scheduling scheme for each target participant.

[0042] In this application, the supply chain collaborative management problem is abstracted as a multi-agent sequential decision-making problem. Multiple participants (i.e., agents) make interactive decisions based on a dynamic environment in a time series, and there are complex dependencies between their decisions. Since the self-attention mechanism is good at capturing long-range dependencies and complex interactions between sequential elements, this implementation adopts a hierarchical encoder with it as the core to model the two-layer relationship between "participants" and "materials within participants".

[0043] The multi-agent sequence decision problem can be further formalized as a partially observable Markov decision process (POMDP), which is described by tuples ‹N,S,A,T,R,›, with the meanings of the symbols as follows: N: The set of agents corresponding to all participants in the supply chain in POMDP. The cardinality of this set, |N|, represents the number of these agents.

[0044] S: The state space of the global state. ∈S represents the global environment state at time step t, and =[ ],in, Let represent the local environment information of the first agent at time step t, the local environment information of the second agent at time step t, ..., the local environment information of the |N|th agent at time step t, respectively.

[0045] A: The joint action space of the agents is composed of the action spaces of each agent, i.e., A = The actions of different types of intelligent agents have different meanings. For example, the actions of the producer are the proportion of purchases from the supplier, and the actions of the supplier are the proportion of production resources allocated.

[0046] T: S×A×S→[0,1]: represents the state transition probability of the global environment state.

[0047] R: The reward that an agent receives from interacting with the environment includes two categories: independent reward (the cost of a single agent) and group reward (the total reward of all agents).

[0048] As an example, the reward for each agent is defined by inventory costs, stockout losses, and overflow losses. Inventory costs are defined as the expenses incurred in maintaining inventory, including warehousing costs, warehouse management costs, etc. At time step t, agent i's inventory cost is... Represented as: , in, This represents the proportion of material j in the inventory of agent i at time step t. This indicates the quantity of different types of materials involved in the inventory of agent i.

[0049] Stockout losses are losses incurred due to insufficient inventory, resulting in the inability to meet material demand upon expiration. This refers to the stockout losses of the intelligent agent. Represented as: ; in, This represents the proportion of material j in inventory among the overdue orders that agent i has not fulfilled at time step t.

[0050] Overflow loss is the additional storage cost incurred due to inventory overflow. At time step t, the overflow loss of agent i is... Represented as: , in, This represents the percentage of inventory that agent i has in stock when it exceeds its inventory capacity limit at time step t.

[0051] like Figure 2 As shown, the decision network adopts an Actor-Critic architecture, where the final state representations of each agent are simultaneously input into both the Actor network (policy network) and the Critic network (commentary network). The Actor network outputs recommended material scheduling schemes (actions) for each target participant, and the Critic network evaluates the long-term value of the joint state and joint action to guide the policy optimization of the Actor network.

[0052] The Actor network (policy network) consists of a policy head configured for each agent. Each policy head receives the final state representation of each agent from the hierarchical encoder and outputs the probability distribution of the agent's actions. For a manufacturer, the action is a vector of the proportion of purchases from various suppliers; for a supplier, the action is a vector of the proportion of production resources allocated among different materials. The output layer of the policy head uses the Softmax activation function to ensure that the output actions satisfy the probability normalization constraint.

[0053] Critic Network: A centralized critic that receives the final state representations of all agents from the output of the hierarchical encoder and the joint actions from the outputs of all policy networks. It fits a state-action value function or a state value function through a deep neural network to evaluate the long-term expected reward of the current global state or a specific joint action.

[0054] This application adopts an "encoder + decision network" architecture. The encoder integrates multi-source heterogeneous information to build a unified collaborative state perception, and the decision network performs parallel mathematical planning and decision-making based on the unified global information to ensure that the output action meets the overall optimality.

[0055] In an optional implementation, the collaborative decision-making model further includes a meta-learner, and the collaborative module is further configured to: Input the final state representations of all target participants into the meta-learner to generate a parameterized policy space; Sampling is performed in a parameterized policy space to obtain multiple parameterized policies; A corresponding decision network is constructed based on the strategy for each parameter, resulting in multiple decision networks; Construct a digital twin of the supply chain based on the local environmental information of each target participant; In the supply chain digital twin, each decision network is driven to interact with the digital twin based on the final state representation, execute multi-step virtual decisions, and obtain the performance evaluation results of each decision network in achieving the optimization goal in the supply chain digital twin. The strategy network that performs best in the supply chain digital twin is selected as the decision network for the final output material scheduling recommendation scheme.

[0056] Specifically, the final state representation output by the hierarchical encoder serves as the input to the meta-learner, which is a meta-trained neural network whose core capability is to learn and generate a highly adaptable policy parameter space based on the current overall state of the supply chain. The output of the meta-learner is a parameter vector, which can be parsed as the probability distribution of the decision network parameters, such as the mean and variance of a Gaussian distribution, and the policy space is parameterized through the Gaussian distribution.

[0057] In the parameterized policy space, K sets of policy parameters are obtained through random sampling or confidence interval sampling, and each policy parameter corresponds to a decision policy. A decision network is instantiated based on each policy parameter, and finally K candidate decision networks with the same network structure but different parameter configurations are obtained.

[0058] The construction of a supply chain digital twin is based on the real-time local environmental information of each target participant (including real-time inventory, demand planning, capacity load, and in-transit transportation data). It can accurately simulate the material consumption, production progress, logistics and transportation, and dynamic changes in demand in the real supply chain, and support the simulation of unexpected events (such as delivery delays, capacity interruptions, and transportation failures), providing a virtual simulation environment for the performance evaluation of candidate decision networks.

[0059] During the virtual simulation evaluation phase, each candidate decision network takes the final state representation as its initial input and makes decisions such as procurement and production resource allocation at multiple time steps in the supply chain digital twin simulation. The supply chain digital twin completes the global state transition based on the decision execution results and calculates the immediate reward for each decision based on the preset optimization objective. By recording the cumulative reward of each candidate decision network throughout the entire simulation cycle, it serves as its core performance evaluation result. Finally, the candidate decision network with the highest cumulative reward (or the lowest total supply chain cost) is selected as the final strategy network, which will be used to output a recommended material scheduling scheme.

[0060] The meta-learner is trained using meta-reinforcement learning (Meta-RL) to enable it to generate policy parameter distributions based on the current state of the supply chain. The training process consists of two loops: an inner loop (adaptation loop) and an outer loop (meta-update loop). The specific training logic is as follows: Inner loop: For a single meta-training task (i.e. a specific supply chain scenario), the meta-learner generates a policy parameter distribution based on the initial state of the scenario. After sampling to obtain the policy network, it performs a small number of policy gradient updates in the simulation environment of the task, so as to realize the rapid adaptation of the policy to the specific supply chain scenario.

[0061] Outer loop: Across multiple different meta-training tasks, the parameters of the meta-learner are updated with the goal of maximizing the expected cumulative reward of the policy adapted by the inner loop on the respective meta-training tasks.

[0062] In an optional implementation, the agent-driven supply chain data collaboration system further includes a model training module, which is used for: Obtain a training set consisting of multiple training samples. Each training sample includes the local environment information of each participant at historical time step t, the material scheduling execution plan of each participant at historical time step t (i.e. the material scheduling plan actually executed), the global reward value obtained after executing the material scheduling execution plan at historical time step t, and the local environment information of each participant at historical time step t+1, where t is an integer greater than 0. Obtain the preset optimization target, and use the training set to train the parameters of the hierarchical encoder, meta-learner, and decision network in the collaborative decision-making model.

[0063] In an alternative implementation, the global reward value is determined based on the demand side's inventory holding costs, stockout losses, and spillover losses.

[0064] The above describes how, when constructing the training set, the data of each training sample is sliced ​​into time series segments according to the corresponding dimension, forming […]. ],in, Let represent the local environment information of the first agent at time step t, the local environment information of the second agent at time step t, ..., the material scheduling execution plan (corresponding to the action) executed by the first agent at time step t, the material scheduling execution plan (corresponding to the action) executed by the second agent at time step t, ..., the action executed by the first agent. The second agent receives the global reward value after performing an action. The global reward value obtained afterward, ..., the first agent in executing the action The new state transitioned to afterward, i.e., the local environment information of the first agent at time step t+1, and the second agent's execution of actions. The new state that the agent then transitions to is the local environment information of the second agent at time step t+1, ...

[0065] Then, the training process is executed, and the training logic is as follows; Parameter initialization: The parameters of the hierarchical encoder, meta-learner and decision network are randomly initialized, where the parameters of the decision network refer to its initial basic parameters or the weights corresponding to the basic parameters output by the meta-learner.

[0066] Reward function setting: The preset optimization objective (such as minimizing the total supply chain cost) is transformed into a reward function for reinforcement learning. The calculation of this reward function can be determined by inventory costs, stockout losses, overflow losses, etc.

[0067] Using the training set, an offline reinforcement learning method was employed to initially train the hierarchical encoder and decision network: With the goal of minimizing the error between the output action of the decision network and the historical real action, such as minimizing the mean square error, the model can initially learn the strategy patterns in historical supply chain decisions. We employ methods such as Conservative Q-Learning (CQL) or Behavioral Cloning Regularization to learn the initial value function and decision-making strategy from the training set.

[0068] The meta-training process employs an optimization-based meta-reinforcement learning framework, performing two-layer optimization of the outer and inner loops within the simulation environment of the supply chain digital twin. The specific process is as follows: (1) Outer loop: Sample a batch of training tasks from a series of different simulated supply chain tasks (scenarios) in the preset; each task represents a specific supply chain configuration scenario (such as sudden changes in demand patterns, a supplier's capacity reduction, fluctuations in logistics and transportation timeliness, etc.). For each sampled task, perform rapid adaptation training in the inner loop in sequence.

[0069] (2) Inner loop: For each sampled supply chain task, perform the following steps to achieve rapid strategy adaptation: a, the task The initial state is input to the current meta-learner and hierarchical encoder to generate the task. strategy parameters Or parameter distribution.

[0070] b. In a supply chain digital twin simulation environment, utilize the generated strategies and tasks Perform the interaction for a specified number of steps to obtain decision trajectory data.

[0071] c. Calculate the strategy loss for this task based on the decision trajectory data and the preset reward function. .

[0072] d. Perform one or more steps of gradient descent on the policy parameters to obtain the adapted policy parameters. : ,in This is the learning rate for the inner loop.

[0073] Global update of meta-parameters: After all tasks in the batch have completed the adaptation of the inner loop, the gradient descent algorithm is used to synchronously update the meta-learner parameters and the hierarchical encoder parameters.

[0074] Repeat the outer loop, inner loop, and meta-parameter update steps described above until a preset iteration stopping condition is met. As an example, training stops when the average performance of the meta-learner reaches a preset threshold, such as a 20% reduction in total supply chain costs.

[0075] In an optional implementation, the collaboration module is further configured to: Based on the material demand information of the demand side, obtain the historical credit ratings of each supplier and their quotation information for various materials; Based on the local environmental information, historical credit ratings, and pricing information for various materials of each supplier, a recommended list of suppliers for various materials is generated. Send a list of recommended suppliers to the demand side's management system; After receiving the supplier's recommended list of suppliers for various materials from the demand side's management system, the target participants are determined based on the suppliers for each type of material.

[0076] The material requirements information includes the quantity of each material required by the manufacturer, the timeframe for demand for each material, and the acceptable price range for each material. Based on this information, a set of qualified suppliers can be selected for each material. Then, for each material, the qualified suppliers are sorted according to their historical credit rating, with those having higher historical credit ratings ranked first, ultimately generating a supplier recommendation list.

[0077] In an optional implementation, the collaboration module is further configured to: Receive the material scheduling execution plan returned by each participating party in response to the recommended material scheduling plan; Determine the procurement execution list based on the material scheduling execution plan; Based on the procurement execution list, electronic contracts are signed with relevant participants using blockchain technology.

[0078] Among them, the material scheduling execution plan is the final execution plan selected by the manufacturer and the supplier and has been confirmed. Based on this material scheduling execution plan, a procurement execution list is generated. For example, 624 chips are purchased from supplier A, 156 chips are purchased from supplier B, and 1000 batteries are purchased from supplier C.

[0079] Then, based on the procurement execution list, all participating nodes jointly generate an electronic contract. The electronic contract includes the quantity of materials to be purchased, the supplier, the delivery time, the price settlement, and the liability for breach of contract.

[0080] Once the electronic contract is generated, it is pushed to all participating nodes in the blockchain network. The blockchain network is a consortium blockchain network composed of manufacturers and suppliers. Each relevant participating node verifies the contract content. After successful verification, node confirmation is completed via digital signature. Once all relevant participating nodes have confirmed the contract, blockchain consensus is reached, and the electronic contract is officially signed. The signed electronic contract is then linked to the procurement execution list and material scheduling execution plan, jointly stored on the blockchain, and a unique hash value is generated.

[0081] Accordingly, see Figure 3 , Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of this application. In this embodiment, the electronic device 10 includes a memory 101, a processor 100, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, it implements the agent-driven supply chain data collaboration system provided in the above embodiment.

[0082] Electronic device 10 can be a mobile computing device (e.g., a mobile phone), a desktop computer, a laptop, a PDA, or a cloud server, etc. Electronic device 10 may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will understand that... Figure 3 This is merely an example of an electronic device and does not constitute a limitation on electronic devices. It may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, it may also include input / output devices, network access devices, etc.

[0083] The processor 100 may be a microcontroller unit (MCU) or a central processing unit (CPU). It may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0084] In some embodiments, memory 101 may be an internal storage unit of an electronic device, such as a hard disk or memory. In other embodiments, memory 101 may be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Furthermore, memory 101 may include both internal and external storage units of the electronic device. Memory 101 is used to store operating systems, applications, bootloaders, data, and other programs, such as program code for computer programs. Memory 101 can also be used to temporarily store data that has been output or will be output.

[0085] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. An agent-driven supply chain data collaboration system, characterized in that, The agent-driven supply chain data collaboration system includes: The data interface module is used to connect to the management systems of each participant in the supply chain, so as to obtain the local environmental information of each participant through the management systems of each participant; wherein, the participants in the supply chain include demanders and suppliers, the local environmental information of the demanders includes their material inventory status and material demand information, and the local environmental information of the suppliers includes their material supply capacity status. The collaboration module is used to input the local environmental information of each target participant associated with the material demand information into the collaborative decision model, which makes a decision based on the preset optimization objective and outputs the material scheduling recommendation scheme corresponding to each target participant. The feedback module is used to feed back the recommended material scheduling scheme to the management system of the relevant participants.

2. The agent-driven supply chain data collaboration system as described in claim 1, characterized in that, The collaborative decision-making model includes a hierarchical encoder and a decision network. The hierarchical encoder includes a local encoding layer, an entity attention layer, and a fusion layer. The collaborative module is specifically used for: The local environment information of each target participant is input into the hierarchical encoder. The local environment representation of each target participant is obtained through the local encoding layer. The entity attention layer calculates the interaction relationship between the target participants based on the multi-head self-attention mechanism to obtain the context representation of each target participant that incorporates information from other target participants. The fusion layer fuses the context representation and local environment representation of each target participant to obtain the final state representation of each target participant. The final state representation of each target participant is input into the decision network, which then makes a decision based on the optimization objective to obtain a material scheduling recommendation scheme for each target participant.

3. The agent-driven supply chain data collaboration system as described in claim 2, characterized in that, The collaborative decision-making model further includes a meta-learner, and the collaborative module is also used for: The final state representations of all the target participants are input into the meta-learner to generate multiple parameter policies for the decision network; Based on each of the aforementioned parameter strategies, a corresponding decision network is constructed to obtain multiple decision networks; Construct a digital twin of the supply chain based on the local environmental information of each of the target participants; In the supply chain digital twin, each decision network is driven to interact with the digital twin based on the final state representation, execute multi-step virtual decisions, and obtain the performance evaluation results of each decision network in achieving the optimization goal in the supply chain digital twin. The strategy network that performs best in the supply chain digital twin is selected as the decision network for the final output material scheduling recommendation scheme.

4. The agent-driven supply chain data collaboration system as described in claim 3, characterized in that, The collaboration module is also used for: The final state representations of all the target participants are input into the meta-learner to generate a parameterized policy space; Sampling is performed in the parameterized policy space to obtain multiple parameter policies.

5. The agent-driven supply chain data collaboration system as described in claim 3, characterized in that, The agent-driven supply chain data collaboration system also includes a model training module, which is used for: A training set consisting of multiple training samples is obtained. Each training sample includes the local environment information of each participant at historical time step t, the material scheduling execution plan of each participant at historical time step t, the global reward value obtained after executing the material scheduling execution plan at historical time step t, and the local environment information of each participant at historical time step t+1, where t is an integer greater than 0. Obtain a preset optimization target, and use the training set to train the parameters of the hierarchical encoder, the meta-learner, and the decision network in the collaborative decision-making model.

6. The agent-driven supply chain data collaboration system as described in claim 5, characterized in that, The global reward value is determined based on the target participant's inventory holding costs, stockout losses, and spillover losses.

7. The agent-driven supply chain data collaboration system as described in any one of claims 1 to 6, characterized in that, The collaboration module is also used for: Based on the material demand information of the demand side, obtain the historical credit rating of each of the suppliers and their quotation information for various types of materials; Based on the local environmental information, historical credit ratings, and pricing information for each of the aforementioned suppliers, a recommended list of suppliers for each of the aforementioned materials is generated. Send the supplier recommendation list to the management system of the demand side; After receiving the supplier's final selection for each type of material from the supplier recommendation list returned by the demand side's management system, each target participant is determined based on the final selection for each type of material.

8. The agent-driven supply chain data collaboration system as described in any one of claims 1 to 6, characterized in that, The collaboration module is also used for: Receive the material scheduling execution plan returned by each of the participating parties in response to the material scheduling recommendation plan; The procurement execution list is determined based on the aforementioned material scheduling execution plan; Based on the aforementioned procurement execution list, electronic contracts are signed with relevant participants using blockchain technology.

9. The agent-driven supply chain data collaboration system as described in any one of claims 1 to 6, characterized in that, The demand side includes manufacturers, and the supply side includes suppliers.

10. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement an agent-driven supply chain data collaboration system as described in any one of claims 1 to 9.

Images