Generative AI-assisted reverse logistics intelligent recovery network planning and resource recycling system

Abstract

The invention discloses a generative AI-assisted reverse logistics intelligent recovery network planning and resource reutilization system, and particularly relates to the technical field of logistics management, and the system comprises an intelligent perception and access layer which is used for obtaining multi-modal data including a recovery order, sensing data of a recovery carrier, market data and resourceful processing terminal state data in real time; and constructing a generative AI decision center and a multi-modal large language model based on field fine tuning. According to the invention, a generative AI decision center is taken as an intelligent core, integrated cooperation and dynamic global optimization of a reverse logistics full link are realized, a scheduling strategy is dynamically adjusted according to real-time data, a key point and a recycling scheme are evaluated, the disadvantage of traditional module splitting is overcome, the commodity evaluation accuracy is improved, the distribution time efficiency is improved, the battery endurance is prolonged, and the system is suitable for large-scale popularization and application. The resource utilization rate is improved, the self-iterative optimization capability is achieved, sustainable evolution can be achieved, and the overall efficiency of reverse logistics and the resource recycling economic value are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of logistics management technology, and more specifically, to a generative AI-assisted intelligent recycling network planning and resource reuse system for reverse logistics. Background Technology

[0002] With the deepening of the circular economy concept and the booming development of e-commerce, the demand for reverse logistics of post-consumer electronic and electrical equipment, packaging materials and other goods has shown explosive growth. Efficient recycling network planning and precise resource reuse have become key to reducing environmental burden. To this end, the industry has tried to apply modern information technologies such as artificial intelligence and the Internet of Things to all aspects of reverse logistics in order to improve the automation and intelligence level of the system.

[0003] However, these existing technical solutions typically employ a modular design, with each functional module, such as scheduling, evaluation, and disposal, operating relatively independently and transmitting information through simple interfaces. This results in route planning being unable to flexibly respond to changes in commodity value and downstream processing needs, value assessment failing to incorporate real-time scheduling costs and transportation capacity status, and resource-based decision-making lacking a global perspective on material supply and demand. Each part of the system is designed for local optima, thereby compromising overall efficiency. Furthermore, the decision-making logic of existing systems largely relies on static rules or historical models, making it difficult to integrate real-time dynamic information and lacking the ability to continuously self-optimize based on feedback from actual operations. Consequently, the overall level of intelligence is limited, hindering further improvements in recycling network efficiency and resource reuse rates. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a generative AI-assisted intelligent reverse logistics recycling network planning and resource reuse system to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a generative AI-assisted intelligent reverse logistics recycling network planning and resource reuse system, comprising: The intelligent sensing and access layer is used to acquire in real time multimodal data including recycling orders, sensor data of recycling vehicles, market data, and status data of resource recovery processing terminals; The generative AI decision-making hub is built on a multimodal large language model with domain fine-tuning and is connected to the intelligent perception and access layer. It is used to perform unified understanding and contextual reasoning on the acquired multi-source heterogeneous data and generate collaborative control instructions containing decision logic explanations. The dynamic scheduling and path optimization engine is connected to the generative AI decision-making center to receive scheduling instructions issued by it and to plan dynamic recovery paths for the recovery vehicle based on the integrated reinforcement learning agent and real-time environmental information. A multimodal value assessment engine, connected to the generative AI decision center, is used to receive assessment instructions issued by the center and integrate computer vision analysis results with market semantic information extracted by the decision center to generate a commodity value assessment report. The material recycling strategy recommendation engine is connected to the generative AI decision center to receive dismantling instructions issued by it and generate executable resource-based dismantling solutions for products that are determined to be irreparable. The generative AI decision-making center is further configured to dynamically generate and uniformly issue the collaborative control instructions to the corresponding engine based on the output of the multimodal value assessment engine and the status data of the resource processing terminal, so as to coordinate the recycling planning and resource reuse decisions of the entire chain.

[0006] Preferably, the generative AI decision-making center interacts with each engine through a predefined tool call interface; the collaborative control instructions include at least: path optimization target scheduling instructions, commodity evaluation dimension focusing instructions, and material recycling strategy triggering instructions.

[0007] Preferably, the generative AI decision-making center is configured to dynamically set or adjust the optimization target weight in the reward function of the reinforcement learning agent based on at least one of the real-time order value distribution, vehicle energy consumption status, or resource processing terminal requirements.

[0008] Preferably, the multimodal value assessment engine includes a privacy-preserving computational data fusion unit, which is used to combine market data from multiple independent commodity trading platforms to jointly train and optimize the engine's price assessment model without aggregating the original data. The privacy-preserving computation data fusion unit is implemented using federated learning technology.

[0009] Preferably, the resource recycling dismantling scheme generated by the material recycling strategy recommendation engine is a structured chemical bill, which includes at least a dismantling step sequence, the expected recovery rate of the target precious metal, and a recommended processing terminal.

[0010] Preferably, the generative AI decision-making center is configured to: receive inventory or demand feedback from the resource processing terminal, and generate instructions to the dynamic scheduling and path optimization engine accordingly, so as to prioritize the scheduling of specific categories of recycled goods that match the demand.

[0011] Preferably, the system further includes a digital twin simulation module, used to simulate and extrapolate the collaborative process of dynamic scheduling, value assessment and material recycling before the generative AI decision-making center executes major strategies, and to feed the results back to the decision-making center for instruction optimization.

[0012] Preferably, the generative AI decision-making center is configured to have self-iterative optimization capabilities, used to continuously fine-tune the multimodal large language model based on the historical execution feedback data of the dynamic scheduling and path optimization engine, the multimodal value assessment engine and the material recycling strategy recommendation engine, so as to optimize its decision logic and instruction generation strategy.

[0013] Preferably, the system is further configured with intelligent workflow and task assignment functions, used for: Based on the product evaluation results output by the multimodal value evaluation engine, and combined with the real-time instructions of the generative AI decision-making center; The product is automatically assigned to at least two different subsequent processing flows: one is the whole machine circulation process that directly enters the second-hand sales channel; the other is the resource recycling process that triggers the material recycling strategy recommendation engine to generate and execute the dismantling scheme.

[0014] The technical effects and advantages of this invention are as follows: By using a generative AI decision-making hub as the unified intelligent core of the system, and through predefined tool call interfaces, it performs unified understanding and contextual reasoning on multi-source heterogeneous data across the entire chain, and generates collaborative control instructions with logical explanations. This dynamically directs the three downstream engines, including scheduling, evaluation, and circulation, so that scheduling targets can be adjusted in real time according to evaluation value and dismantling needs, evaluation focus can be dynamically focused according to scheduling status and material requirements, and dismantling strategies can be accurately generated based on commodity status and processing capacity. This deep collaborative mechanism driven by a single intelligent agent breaks through the limitations of simple splicing of traditional system modules, and realizes the integration and global optimization of reverse logistics full-chain decision-making, thereby significantly improving overall operational efficiency and resource coordination capabilities. The decision-making center can dynamically set or adjust the reward function weight of the path optimization engine based on real-time order value, vehicle energy consumption, or processing terminal demand, making the scheduling strategy business-sensitive. Secondly, the system directly intervenes in the scheduling strategy by receiving real-time feedback from the processing terminal, prioritizing the recycling of specific categories, thus achieving demand-driven precise recycling. The system has self-iterative optimization capabilities and can continuously fine-tune the large language model of the decision-making center based on historical execution feedback data of each engine. This allows the system to continuously learn from actual operating results, enabling the accuracy, timeliness, and economy of core decisions such as path planning, commodity valuation, and dismantling scheme recommendation to continuously evolve, overcoming the drawbacks of static systems gradually becoming ineffective. By organically integrating technologies such as generative AI, federated learning, reinforcement learning, and computer vision, the system addresses long-standing pain points in the industry. Regarding valuation accuracy, the multimodal value assessment engine utilizes federated learning technology to integrate market data from multiple platforms, significantly enriching the training samples while protecting data privacy. This results in a valuation accuracy rate far exceeding traditional methods that rely on a single data source or human experience. In terms of operational efficiency and cost, the dynamic scheduling model, coordinated by the generative AI hub, comprehensively considers real-time road conditions, order value, and vehicle status, achieving a holistic optimization of delivery timeliness and extended battery life. Finally, in terms of resource utilization, the refined chemical orders generated by the material recycling strategy recommendation engine, combined with intelligent circulation and allocation functions, ensure that irreparable goods receive the most appropriate resource processing, thereby improving overall resource utilization. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1 This embodiment takes the recycling of consumer electronics products within an urban area as an application scenario, and elaborates in detail on the composition and workflow of the generative AI-assisted reverse logistics intelligent recycling network planning and resource reuse system.

[0018] As attached Figure 1 As shown, this system adopts a core collaborative architecture, which mainly includes: an intelligent sensing and access layer, a generative AI decision-making center, a dynamic scheduling and path optimization engine, a multimodal value assessment engine, a material recycling strategy recommendation engine, and advanced functional modules. Each module communicates through a data bus and API interface.

[0019] Intelligent Sensing and Access Layer: This layer serves as the system's data entry point, integrating multiple data interfaces and responsible for real-time acquisition and preprocessing of multi-source heterogeneous data. Recover order data: Receive text descriptions, product images, and short videos submitted by users through a client application.

[0020] Vehicle sensor data: Real-time acquisition of GPS coordinates, remaining battery power, onboard camera images, acceleration and weather sensor readings of drones or unmanned vehicles via IoT protocols.

[0021] Market data: Through the platform's authorized API, we periodically capture and analyze the transaction history, active model list, and price fluctuation trends of second-hand trading platforms.

[0022] Processing terminal status data: Interacting with the MES system of the resource recovery center to obtain its real-time inventory level, procurement needs for specific materials such as gold, palladium, and cobalt, and the status of production line equipment.

[0023] Generative AI Decision Center: This central hub is the intelligent core of the system. It is based on an open-source or basic multimodal large language model and is obtained through supervised domain fine-tuning using massive amounts of reverse logistics work order data, commodity inspection reports, disassembly manuals, and market analysis reports.

[0024] Unified understanding and reasoning: The generative AI decision center receives all data from the intelligent perception and access layer and can perform cross-modal alignment and understanding. For example, it can parse "functioning normally" in the order text, but at the same time identify the minor cracks in the screen in the appearance image, and combine it with the known vulnerability characteristics of the model to determine that there is a potential motherboard risk, thereby generating more prudent evaluation instructions.

[0025] Instruction Generation and Coordination: The central engine interacts with downstream engines through predefined function call interfaces. The structured instructions it generates are essentially API call commands encapsulated with specific parameters, primarily including: (a) Scheduling instructions: including optimization objectives, constraints, and priority parameters; (b) Evaluation instructions: Specify the evaluation dimensions that need to be focused on, such as instructing the visual model to focus on detecting specific defects such as corrosion on the motherboard or bulging of the battery, and the market reference range.

[0026] (c) Disassembly instruction: a trigger signal, with the initial evaluation conclusion as input.

[0027] Dynamic Decision Optimization: The generative AI decision-making center has the ability to adjust strategies in real time. Specifically, it has a built-in strategy weight calculation submodule. Based on the input processing terminal demand data, such as the characteristics of a certain item's inventory shortage and the current order pool, this module calculates a set of weight vectors [α, β, γ] in real time, which correspond to the "timeliness", "energy consumption" and "recycling value of specific materials" in the reward function, respectively. This vector is then sent to the reinforcement learning agent of the dynamic scheduling and path optimization engine, enabling it to immediately adjust its optimization direction.

[0028] Dynamic scheduling and path optimization engine: The engine receives scheduling instructions from the generative AI decision-making center.

[0029] At its core is an intelligent agent based on a deep reinforcement learning algorithm. The reward function of this agent consists of a weighted sum of multiple optimization objective sub-items. The weight parameters [α,β,γ] are dynamically calculated and distributed by the generative AI decision center according to the global strategy, so that the path planning can respond to different business priorities in real time.

[0030] During engine operation, the latest traffic events and weather information are continuously obtained from the intelligent perception and access layer. The fleet routes are replanned or adjusted at a frequency of minutes. In an optimal implementation, the generative AI decision center can generate multiple preliminary route plans. The reinforcement learning agent of the dynamic scheduling and route optimization engine performs rapid simulation and evaluation based on the "strategy candidate set" composed of these plans, and selects the path with the highest comprehensive reward as the final execution plan.

[0031] The planned precise trajectory and instructions are sent to the flight control or piloting systems of each recovery vehicle via a low-latency communication link.

[0032] Multimodal value assessment engine: The engine receives evaluation instructions from the generative AI decision-making center and performs commodity value assessment.

[0033] Its visual analysis module uses a convolutional neural network to extract feature vectors from more than ten dimensions of the product's appearance, such as scratches, dents, discoloration, and liquid infiltration indicators.

[0034] To achieve high-precision valuation, the engine includes a privacy-preserving computational data fusion unit that employs a federated learning framework. This unit deploys model clients locally on multiple collaborating e-commerce platforms, uses local data to compute model updates, and only uploads encrypted gradient updates to a cloud server for aggregation. After generating a global model, it is then distributed. This iterative process allows the valuation model of the multimodal value assessment engine to learn the data distribution across multiple platforms, while ensuring that the original transaction data remains within the domain, thus meeting privacy compliance requirements.

[0035] After the evaluation report is generated, the system will automatically invoke the intelligent workflow and task assignment functions.

[0036] Material recycling strategy recommendation engine: This engine is specifically designed to handle items marked as irreparable.

[0037] It receives dismantling instructions and accesses the built-in commodity structure knowledge graph to generate a structured electronic work order. This work order includes at least: a sequence of operation steps, a list of target recyclable materials and their expected recovery rate based on the Bill of Materials (BOM), such as an estimated recovery of 0.028 grams ± 0.001 grams of gold, and the optimal processing terminal number recommended based on processing capacity and logistics costs.

[0038] Work orders are automatically pushed to the production management system of the corresponding processing terminal through the enterprise service bus to guide assembly line operations.

[0039] Digital twin simulation module: To improve decision-making reliability, the system includes this module. When the instructions generated by the generative AI decision-making center exceed preset thresholds, such as scheduling more than 50% of vehicles or the total value exceeding a certain amount, the digital twin simulation module is automatically activated. Based on the current system digital twin, which includes virtual road network, vehicle model, order model, and factory model, it injects instructions to be executed and runs Monte Carlo simulation or discrete event simulation to predict system KPIs in the next few hours, such as total cost, average response time, and total resource recovery. The simulation report will highlight potential bottlenecks and risks and provide feedback to the center as correction suggestions for generating instructions.

[0040] Intelligent workflow and task assignment functions: This function is implemented by the system's built-in rule engine, which is equipped with a predefined flow rule library based on decision trees and fuzzy logic. It is used to automatically execute preliminary allocation decisions based on the product evaluation results output by the multimodal value evaluation engine.

[0041] For example, the rule base has the following pre-defined rules: Rule 1: If a product's appearance rating is ≥ B and its functions are tested and found to be normal, it will be assigned to the "second-hand sales channel".

[0042] Rule 2: If the motherboard of a product is severely damaged or is identified as containing high-value rare metals, it will be assigned to the "resource recycling channel" and the material recycling strategy recommendation engine will be triggered.

[0043] The generative AI decision-making hub is configured to send overriding instructions to the rule engine based on real-time global policies, so as to dynamically adjust the priority, judgment threshold or parameters of predefined rules, thereby achieving intelligent adaptive assignment that goes beyond fixed rules.

[0044] Self-iterative optimization capability: The generative AI decision-making hub has a continuous learning mechanism. The system has a feedback data warehouse that persistently stores the real results data of each recycling task from order placement to completion. It regularly, such as weekly, starts model fine-tuning tasks, using these feedback data and original prediction data to construct a loss function and perform lightweight fine-tuning on the hub's large language model, thereby continuously evolving its ability to judge market value, assess commodity loss, and predict path time.

[0045] Example 2 To enable those skilled in the art to better reproduce the invention, a simplified operational example is provided: Sensing: A user submits an order for an old mobile phone of model "X", uploading photos showing a cracked screen. Simultaneously, the system detects a request from dismantling factory "Center C": urgently needing a motherboard of model "X" for precious metal extraction.

[0046] Central decision-making and coordination: The generative AI decision-making center analyzed the photos and confirmed that the screen was severely damaged, resulting in low resale value for the entire device.

[0047] The generative AI decision-making center integrates information such as "model X" and "urgent demand for motherboards" to deduce a strategy of "prioritizing recycling and direct delivery for disassembly".

[0048] Generative AI decision-making center generates collaborative instruction packages: I) Send instructions to the dynamic scheduling and path optimization engine to dynamically adjust its reward function and significantly increase the weight γ of "model X order sent to center C"; Ii) sends instructions to the multimodal value assessment engine, requesting it to focus its assessment on "whether the motherboard is physically intact".

[0049] Execution and Closed Loop: The dynamic scheduling and path optimization engine plans the optimal drone path to avoid congested areas and fly directly to the center C.

[0050] The visual model of the multimodal valuation engine confirms that the motherboard is intact, and the federated learning model, combined with real-time metal prices, gives a high dismantling estimate.

[0051] The intelligent workflow function automatically marks the order as "disassembled", triggering the material recycling strategy recommendation engine to generate a detailed disassembly work order and send it to center C.

[0052] After the drone completes the recovery and transportation, and is disassembled at center C, the actual weight data of the recovered palladium metal is fed back to the system as feedback and stored in the feedback warehouse for subsequent model iterations.

[0053] Finally, it should be noted that the accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A generative AI-assisted intelligent reverse logistics network planning and resource reutilization system, characterized in that, include: The intelligent sensing and access layer is used to acquire in real time multimodal data including recycling orders, sensor data of recycling vehicles, market data, and status data of resource recovery processing terminals; The generative AI decision-making hub is built on a multimodal large language model with domain fine-tuning and is connected to the intelligent perception and access layer. It is used to perform unified understanding and contextual reasoning on the acquired multi-source heterogeneous data and generate collaborative control instructions containing decision logic explanations. The dynamic scheduling and path optimization engine is connected to the generative AI decision-making center to receive scheduling instructions issued by it and to plan dynamic recovery paths for the recovery vehicle based on the integrated reinforcement learning agent and real-time environmental information. A multimodal value assessment engine, connected to the generative AI decision center, is used to receive assessment instructions issued by the center and integrate computer vision analysis results with market semantic information extracted by the decision center to generate a commodity value assessment report. The material recycling strategy recommendation engine is connected to the generative AI decision center to receive dismantling instructions issued by it and generate executable resource-based dismantling solutions for products that are determined to be irreparable. The generative AI decision-making center is further configured to dynamically generate and uniformly issue the collaborative control instructions to the corresponding engine based on the output of the multimodal value assessment engine and the status data of the resource processing terminal, so as to coordinate the recycling planning and resource reuse decisions of the entire chain.

2. The generative AI-augmented intelligent reverse logistics network planning and resource re-purposing system of claim 1, wherein, The generative AI decision-making center interacts with each engine through a predefined tool call interface; the collaborative control instructions include at least: path optimization target scheduling instructions, commodity evaluation dimension focusing instructions, and material recycling strategy triggering instructions.

3. The generative AI-augmented intelligent reverse logistics network planning and resource re-purposing system of claim 1, wherein, The generative AI decision-making center is configured to dynamically set or adjust the optimization target weight in the reward function of the reinforcement learning agent based on at least one of the following: real-time order value distribution, vehicle energy consumption status, or resource processing terminal requirements.

4. The generative AI-augmented intelligent reverse logistics intelligent recycling network planning and resource reusing system of claim 1, wherein, The multimodal value assessment engine includes a privacy-preserving computational data fusion unit, which combines market data from multiple independent commodity trading platforms to jointly train and optimize the engine's price assessment model without aggregating the original data. The privacy-preserving computation data fusion unit is implemented using federated learning technology.

5. The generative AI-assisted reverse logistics intelligent recycling network planning and resource reuse system according to claim 1, characterized in that, The resource recycling strategy recommendation engine generates a structured chemical bill of materials, which includes at least the dismantling step sequence, the expected recovery rate of the target precious metal, and the recommended processing terminal.

6. The generative AI-assisted reverse logistics intelligent recycling network planning and resource reuse system according to claim 1, characterized in that, The generative AI decision-making center is configured to receive inventory or demand feedback from the resource processing terminal and generate instructions to the dynamic scheduling and path optimization engine accordingly, so as to prioritize the scheduling of specific categories of recycled goods that match the demand.

7. The generative AI-assisted reverse logistics intelligent recycling network planning and resource reuse system according to claim 1, characterized in that, The system also includes a digital twin simulation module, which is used to simulate and extrapolate the collaborative process of dynamic scheduling, value assessment and material recycling before the generative AI decision-making center executes major strategies, and feeds the results back to the decision-making center for instruction optimization.

8. The generative AI-assisted reverse logistics intelligent recycling network planning and resource reuse system according to claim 1, characterized in that, The generative AI decision-making center is configured to have self-iterative optimization capabilities, and is used to continuously fine-tune the multimodal large language model based on the historical execution feedback data of the dynamic scheduling and path optimization engine, the multimodal value assessment engine and the material recycling strategy recommendation engine, so as to optimize its decision logic and instruction generation strategy.

9. The generative AI-assisted reverse logistics intelligent recycling network planning and resource reuse system according to claim 1, characterized in that, The system is also configured with intelligent workflow and task assignment functions, used for: Based on the product evaluation results output by the multimodal value evaluation engine, and combined with the real-time instructions of the generative AI decision-making center; The product is automatically assigned to at least two different subsequent processing flows: one is the whole machine circulation process that directly enters the second-hand sales channel; the other is the resource recycling process that triggers the material recycling strategy recommendation engine to generate and execute the dismantling scheme.

Images