Methods, apparatus, equipment and media for optimizing material flow in a circular supply chain

Abstract

This application relates to the field of intelligent supply chain management, providing a method, apparatus, equipment, and medium for optimizing material flow in a circular supply chain. The method includes: collecting multi-source heterogeneous data; semantically parsing recycled product quality inspection reports using a Large Language Model (LLM) and generating a quality vector characterizing the remanufacturing potential of recycled products by combining physical imaging analysis results; constructing a dynamic cost matrix including tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs by integrating real-time cross-border logistics cost information and environmental and tariff regulations using a dynamic modeling module; inputting the quality vector and dynamic cost matrix into a multi-objective optimization engine to generate a Pareto optimal solution set for remanufacturing center location selection and dynamic material flow scheduling, under the conditions of satisfying preset refurbishment rate thresholds, total cost constraints, and regional compliance; and dynamically adjusting the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set in response to external environmental changes, achieving closed-loop optimization.

Description

Technical Field

[0001] This application relates to the field of intelligent supply chain management, and in particular to a method, apparatus, equipment and medium for optimizing the material flow of a circular supply chain. Background Technology

[0002] With increasing global emphasis on sustainable development and the circular economy, circular supply chains have become a key path for the transformation and upgrading of the manufacturing industry. Their core lies in recycling used products and restoring their value through remanufacturing, thereby reducing resource consumption and environmental impact. Within this system, how to efficiently, economically, and compliantly plan the flow of recycled materials and scientifically decide on the location of remanufacturing centers constitutes a complex global optimization problem.

[0003] Currently, optimization methods in the field of circular supply chains typically rely on traditional operations research models or static cost-benefit analyses. These methods first collect historical data, such as average transportation costs, fixed tariff rates, and quality assessment reports based on sampling. They then construct linear or nonlinear programming models to seek the solution with the lowest total cost or highest efficiency while satisfying basic constraints. These methods largely treat dynamically changing real-world parameters as static or quasi-static, and their optimization results depend on the stability of historical data.

[0004] Due to the global nature of circular supply chains, their operating environment is highly dynamic and uncertain. The aforementioned existing technological methods suffer from the following main drawbacks: 1) At the data level, it is difficult to effectively integrate and process multi-source, heterogeneous data from around the world (e.g., quality reports in different languages ​​and formats, real-time fluctuating logistics costs, and frequently updated regulations from various countries), leading to incomplete, untimely, or inconsistent information on which decisions are based; 2) At the modeling level, static models cannot effectively respond to rapid changes in the external environment. For example, when a country suddenly adjusts its tariff policy, a natural disaster causes logistical disruptions, or there are significant batch differences in the quality of recycled products, the optimal solution derived from existing data may quickly become invalid, even leading to compliance risks and economic losses. Summary of the Invention

[0005] Based on this, it is necessary to address the technical problems of the existing technologies, such as the inability to dynamically respond to real-time changes, resulting in delayed decision-making information and invalid optimization results, and propose a method, device, equipment and medium for optimizing the material flow of a circular supply chain.

[0006] Firstly, a method for optimizing material flow in a circular supply chain is provided, the method comprising:

[0007] Collect multi-source heterogeneous data, including quality inspection reports of recyclables from different regions around the world, real-time cost information of cross-border logistics, and texts of environmental protection and tariff regulations of various countries;

[0008] The quality inspection report of the recycled products is semantically parsed using a large language model (LLM), and a quality vector characterizing the remanufacturing potential of the recycled products is generated by combining the results of physical imaging analysis.

[0009] By using the dynamic modeling module to integrate the real-time cost information of cross-border logistics with the text of environmental protection and tariff regulations, a dynamic cost matrix is ​​constructed that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs.

[0010] The quality vector and the dynamic cost matrix are input into the multi-objective optimization engine to generate a Pareto optimal solution set for remanufacturing center location selection and dynamic scheduling of material flow, under the conditions of satisfying the preset refurbishment rate threshold, total cost constraint and regional compliance.

[0011] In response to changes in the external environment, the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set are dynamically adjusted to achieve closed-loop optimization.

[0012] Secondly, a circular supply chain material flow optimization device is provided, the device comprising:

[0013] The acquisition module is used to collect multi-source heterogeneous data, which includes quality inspection reports of recyclables from different regions around the world, real-time cost information of cross-border logistics, and texts of environmental protection and tariff regulations of various countries.

[0014] The first generation module is used to perform semantic parsing on the quality inspection report of the recycled products using a large language model (LLM), and to generate a quality vector characterizing the remanufacturing potential of the recycled products by combining the physical imaging analysis results.

[0015] The construction module is used to integrate the real-time cost information of cross-border logistics and the text of environmental protection and tariff regulations using the dynamic modeling module to construct a dynamic cost matrix that includes tariff weights, transportation cost fluctuation factors and carbon emission compliance costs.

[0016] The second generation module is used to input the quality vector and the dynamic cost matrix into the multi-objective optimization engine, and generate a Pareto optimal solution set for remanufacturing center location selection and material flow dynamic scheduling under the conditions of satisfying the preset refurbishment rate threshold, total cost constraint and regional compliance.

[0017] The adjustment module is used to respond to external environmental changes and dynamically adjust the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set to achieve closed-loop optimization.

[0018] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described circular supply chain material flow optimization method.

[0019] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described circular supply chain material flow optimization method.

[0020] As can be seen from the technical solution provided in this application, on the one hand, by semantically parsing the quality inspection report of recycled products using a large language model and generating a quality vector characterizing the remanufacturing potential of recycled products in combination with physical imaging analysis results, the optimization decision is no longer limited to a few structured data, but is based on richer and more accurate information, providing more reliable input for subsequent optimization. The originally static cost parameters are transformed into dynamic variables that can be updated in real time with market conditions and policies and regulations, enabling the optimization model to promptly capture changes in cost structure and compliance requirements. This makes the calculated solution closer to the current actual situation, effectively avoiding the risk of cost underestimation or compliance violations due to information lag. On the other hand, by inputting the quality vector and dynamic cost matrix into the multi-objective optimization engine, while satisfying the preset refurbishment rate threshold, total cost constraints, and... Under the condition of regional compliance, the Pareto optimal solution set for the location selection of remanufacturing centers and the dynamic scheduling of material flows is generated. It provides not a single solution, but a series of optimal trade-offs (Pareto solution set), enabling decision-makers to choose according to strategic preferences. This allows for finding a globally optimal solution in complex multi-objective decision-making, rather than a locally optimal solution. Thirdly, by responding to changes in the external environment, the logistics paths and remanufacturing resource allocation strategies corresponding to the Pareto optimal solution set are dynamically adjusted to achieve closed-loop optimization. This method transforms the supply chain system from a static, one-off planning system into a dynamic, continuously learning, adaptive system. When external disturbances occur, the system can proactively trigger re-optimization and adjust resource allocation, thereby significantly improving the resilience and robustness of the supply chain and ensuring that it maintains a near-optimal operating state in a dynamic and uncertain environment. In summary, the technical solution of this application, by integrating multi-source data and constructing a dynamic model, can achieve real-time optimization and dynamic adjustment under multiple objectives, improving decision-making accuracy and supply chain resilience. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0022] Figure 1 This is a diagram illustrating an application scenario of the circular supply chain material flow optimization method in one embodiment;

[0023] Figure 2 This is a flowchart of a circular supply chain material flow optimization method in one embodiment;

[0024] Figure 3 This is a structural block diagram of a circular supply chain material flow optimization device in one embodiment;

[0025] Figure 4 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0026] 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, not all, of the embodiments of the present invention. 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.

[0027] Currently, optimization methods in the circular supply chain field typically rely on traditional operations research models or static cost-benefit analyses. These methods first collect historical data, such as average transportation costs, fixed tariff rates, and quality assessment reports based on sampling. They then construct linear or nonlinear programming models to seek the lowest total cost or most efficient solution while satisfying basic constraints. These methods largely treat dynamically changing real-world parameters as static or quasi-static, and their optimization results depend on the stability of historical data. Due to the global nature of circular supply chains, their operating environment is highly dynamic and uncertain. The aforementioned existing methods have the following main drawbacks: 1) At the data level, it is difficult to effectively integrate and process multi-source, heterogeneous data from around the world (e.g., quality reports in different languages ​​and formats, real-time fluctuating logistics costs, and frequently updated regulations from various countries), leading to incomplete, untimely, or inconsistent information on which decisions are based; 2) At the modeling level, static models cannot effectively respond to rapid changes in the external environment. For example, when a country suddenly adjusts its tariff policy, a natural disaster causes logistical disruptions, or there are significant batch differences in the quality of recycled goods, the optimal solution derived from old data may quickly become invalid, even leading to compliance risks and economic losses.

[0028] To address the aforementioned problems in the prior art, this application proposes a method for optimizing material flow in a circular supply chain, and the corresponding circular supply chain material flow optimization system can be applied to... Figure 1 In the example application scenario, this circular supply chain material flow optimization system mainly involves entities such as a large language model processing module 101, an image recognition service 102, a model fusion module 103, a dynamic modeling module 104, a multi-objective optimization engine 105, and an event listener 106. Figure 1In the example application scenario, suppose global electronics manufacturer R deploys a supply chain intelligence hub to optimize its global recycling network for used laptops. Its workflow is as follows: The system first continuously collects multi-source heterogeneous data from partners worldwide through its data acquisition interface. This includes German quality inspection reports submitted by recycling points in region A1, real-time updated shipping costs in region A2, and the latest electronic waste import regulations from countries in region A3. This raw data is then integrated into a central data platform. Next, the system's large language model processing module 101 performs semantic parsing on the textual quality inspection reports of the recyclables. It accurately understands localized terms such as "Geratezustandsbewertung: Gut" and converts them into standard quality codes. Simultaneously, the image recognition service 102 analyzes X-ray scans of the same batch of equipment. A dedicated model fusion module 103 deeply fuses the aforementioned textual semantic features with the imaging analysis results, generating a fused quality vector that accurately characterizes the remanufacturability potential of each device. On another parallel track, the system's dynamic modeling module 104 begins operation. It integrates real-time fluctuating cross-border logistics cost information with complex environmental and tariff regulations. For example, it calculates that due to the new carbon border adjustment mechanism, the carbon emission cost of a certain route has increased by 15%. Based on this rapidly changing information, the dynamic modeling module 104 constructs a dynamic cost matrix that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs. The multi-objective optimization engine 105 inputs the quality vector generated in the preceding steps along with the dynamic cost matrix. Under strict conditions that satisfy the company's preset refurbishment rate threshold, total cost constraints, and regional compliance, it performs large-scale computational simulations to generate a Pareto optimal solution set for remanufacturing center location and global material flow dynamic scheduling. This provides decision-makers with multiple balanced solutions that combine high refurbishment rates, low total costs, and high compliance. When the system's event listener 106 detects an external environmental change event that may affect the existing scheduling strategy, such as the closure of a major port due to sudden weather, the system will respond immediately. The system will call the dynamic adjustment algorithm to quickly evaluate alternative solutions and dynamically adjust the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set, such as temporarily diverting goods to a backup center, thereby continuously achieving closed-loop optimization in the face of uncertainty.

[0029] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the circular supply chain material flow optimization method provided in this embodiment of the invention mainly includes steps S201 to S206, detailed below:

[0030] Step S201: Collect multi-source heterogeneous data, which includes quality inspection reports of recyclables from different regions around the world, real-time cost information of cross-border logistics, and texts of environmental protection and tariff regulations of various countries.

[0031] Because remanufacturable resources such as electronic waste and discarded household appliances are widely distributed and have complex origins globally, their associated data formats are also highly diverse. Therefore, a distributed data acquisition engine can be deployed to connect to the following three types of data sources:

[0032] 1) Recyclable Product Quality Inspection Reports: These reports may originate from recycling centers, third-party testing agencies, or internal company systems in various countries. They come in diverse formats, including PDF documents, Excel spreadsheets, or API interface data. For example, reports may originate from professional recycling organizations or third-party inspection platforms in B1, B2, B3, or B4 countries or regions. They are typically in PDF, scanned, or structured database format and cover aspects such as equipment appearance ratings, functional test records, and maintenance history. They may also include German terms like "Geratezustandsbewertung" (equipment condition assessment) or Japanese terms like "condition assessment," etc.

[0033] 2) Real-time cost information for cross-border logistics: By accessing international logistics platforms or public databases, real-time information can be obtained from international shipping company API interfaces, oil price monitoring platforms, specific regional dry bulk cargo indices, and land transport fuel price fluctuation curves.

[0034] 3) Environmental and tariff regulations of various countries: Unstructured text data can be obtained from government official websites or professional legal databases. These text data cover official documents such as waste transportation regulations in country D, Harmonized Tariff Schedule (HTS) code tax rate announcements in country E1, revised provisions of the Solid Waste Pollution Prevention and Control Law in country C, and implementation rules of the Household Appliance Recycling Law in country E2.

[0035] These data exhibit significant heterogeneity—containing both natural language text and numerical time series; encompassing both static legal provisions and dynamic market parameters. Directly using them for subsequent modeling would lead to semantic gaps and dimensionality imbalances. Therefore, to ensure data integrity and consistency, this implementation adopts a data lake architecture for storage, with all raw data undergoing verification and cleaning before being stored. For example, a scheduled task automatically retrieves data hourly to prevent information lag.

[0036] The solutions described above enable the establishment of a data access network covering major global economies, ensuring that the system has the ability to acquire information across regions and modes, providing comprehensive and real-time basic data support for subsequent analysis, thereby reducing decision-making biases caused by missing or outdated data.

[0037] Step S202: Semantically analyze the quality inspection report of recycled products using Large Language Models (LLM), and generate a quality vector characterizing the remanufacturing potential of recycled products by combining the results of physical imaging analysis.

[0038] Traditional methods for recycling quality inspection reports cannot effectively utilize unstructured, multilingual recycling quality information, leading to inaccurate and inconsistent assessments of the recycling remanufacturing potential, which in turn affects the accuracy of subsequent resource allocation decisions. For example, having professionals read reports and manually input standardized quality grade codes is not only extremely inefficient and costly, but also susceptible to subjective factors and struggles to handle real-time data streams from large-scale global supply chains. Even with pre-defined keywords (e.g., "damaged," "good") and simple string matching for classification—a keyword-based rule engine approach—it lacks flexibility and cannot understand contextual semantics, thus failing to gain industry acceptance. For instance, a keyword-based rule engine cannot distinguish between "minor scratches on the exterior, but fully functional" and "completely unusable," both of which may contain the keywords "scratched" and "unusable" but have completely different meanings. Furthermore, it cannot handle localized terminology from languages ​​such as German and Japanese. Therefore, in order to transform unstructured text information that was originally useless or difficult to utilize into high-quality, quantifiable quality vectors, this application uses a large language model to perform semantic parsing on the quality inspection report of recycled products, and combines the results of physical imaging analysis to generate quality vectors that characterize the remanufacturing potential of recycled products.

[0039] As one embodiment of this application, semantic parsing of the recycled product quality inspection report using a Large Language Model (LLM) and generating a quality vector characterizing the remanufacturing potential of the recycled product by combining the physical imaging analysis results can be achieved through steps S2021 to S2014, as detailed below:

[0040] Step S2021: Use the multilingual pre-trained bidirectional encoder representation model (BERT) architecture to identify and standardize non-English quality rating terms, and map the non-English quality rating terms to a unified quality level coding system.

[0041] For non-standardized and multilingual mixed recycling product quality inspection reports, this application can use a large language model based on the BERT architecture for semantic parsing. Specifically, the system loads a pre-trained multilingual BERT model (e.g., mBERT or XLM-R) and fine-tunes it for the domain, enabling it to accurately identify and map localized terms used in different countries. For example, "Gerätezustandsbewertung" (device status assessment) in German is automatically classified as "Appearance integrity level 3", while "修理履歴あり" in Japanese is marked as "Having a repair history". This process achieves unified encoding transformation of multilingual terms, solving the problems of low efficiency and high error rate in manual translation.

[0042] Step S2022: Synchronously extract the structured fields regarding the service life of the device, repair history, and functional test results in the recycling product quality inspection report as one of the input dimensions of the quality vector.

[0043] In the embodiment of this application, structured fields such as the service life of the device, repair history, and functional test results may exist in the form of key-value pairs. For example, "Service life: 5 years", which can be directly extracted and normalized into numerical features after extraction.

[0044] The dual-track parsing strategy of the above steps S2021 and S2022 effectively improves the accuracy and context sensitivity of semantic understanding.

[0045] Step S2023: Perform cross-modal alignment on the quality level encoding obtained by parsing and the electronic component integrity image output by the physical imaging system to obtain aligned text semantic features and image spatial features.

[0046] The system can synchronously access physical imaging systems such as X-ray or infrared thermal imaging devices to perform non-destructive testing on typical electronic devices (e.g., smartphones, laptops) and obtain physical damage images of internal components. Among them, X-ray imaging is used to obtain the structural integrity image of internal components. For example, the crack density of the printed circuit board (PCB), and infrared thermal imaging is used to detect abnormal thermal distribution. For example, battery swelling or connector oxidation degree. The two imaging technologies work together. X-ray provides structural details, and infrared provides functional status information, jointly constituting a comprehensive physical damage assessment. By processing the images with a convolutional neural network, key indicators such as the crack density of the printed circuit board, battery swelling degree, and connector oxidation area are extracted. For example, for a recycled laptop, X-ray imaging can show the corrosion degree of the internal circuit board, while the text report may describe the appearance damage. Through a cross-modal alignment model (e.g., ViLBERT based on the attention mechanism), perform cross-modal alignment on the quality level encoding obtained by parsing and the electronic component integrity image output by the physical imaging system to obtain aligned text semantic features and image spatial features.

[0047] Specifically, the quality level code obtained from parsing is aligned across modally with the electronic component integrity image output by the physical imaging system to obtain aligned text semantic features and image spatial features. This can be achieved by: using a pre-trained text encoder (e.g., based on the BERT model in step S2021) to convert the quality level code into a text semantic feature vector; using an image encoder to convert the physical imaging image fused from X-ray and infrared light into an image spatial feature vector; calculating the association weights between text features and image features through a cross-modal attention mechanism (e.g., based on the vision-language pre-trained model ViLBERT) to achieve feature-level alignment. For example, the model will pay attention to the visual features of the corresponding region in the image and the “screen scratch” mentioned in the text to ensure the consistency of semantic and visual information; and outputting the aligned text semantic features and image spatial features.

[0048] Step S2024: Based on the attention mechanism, the aligned text semantic features and image spatial features are fused to generate a fused quality vector. The high-renewal-probability samples in the fused quality vector have a higher clustering density in the vector space.

[0049] In this embodiment, a cross-modal fusion module can be constructed using an attention mechanism to weightedly integrate textual semantic features and image spatial features. This allows the system to not only determine whether a single attribute is qualified but also comprehensively evaluate the synergistic influence between multiple dimensions, thereby avoiding biased decision-making. For example, when a mobile phone reports "minor screen scratches" but the X-ray image shows severe motherboard corrosion, the system will lower its overall refurbishment probability score. The final output is a high-dimensional vector—the quality vector—where each dimension represents: overall appearance score, core component integrity rate, software compatibility index, historical failure frequency correction factor, etc. High refurbishment probability samples in the fused quality vector have a higher clustering density in the vector space, meaning that recyclables with high refurbishment potential will cluster together in the vector space, facilitating subsequent classification.

[0050] Specifically, generating a fused quality vector based on the attention mechanism by fusing aligned text semantic features and image spatial features can be achieved by: inputting the aligned text semantic features and image spatial features into a multi-head attention layer, which calculates the interaction weights between different features. For example, for the text feature "equipment age" and the image feature "component corrosion level", the attention mechanism learns their joint influence weights on refurbishment potential; generating a unified fused feature vector through weighted summation or concatenation; and outputting a fused quality vector, where samples with high refurbishment probability show higher density in the vector space through clustering algorithms (such as K-means), facilitating subsequent classification and optimization.

[0051] As can be seen from steps S2021 to S2024 of the above embodiments, compared with the traditional method of relying on manual sampling or single sensor judgment, this application can significantly reduce the prediction error of the whole machine refurbishment rate through the dual verification mechanism of LLM+AI vision, thereby reducing the waste of resources caused by misjudgment. For example, for an old mobile phone, the value may be underestimated based on a text report alone, but after combining imaging analysis, it may be found that the screen is intact, so it can be given priority to be sent to the whole machine refurbishment line.

[0052] Furthermore, in order to avoid resource waste and improve overall efficiency, after generating the fused quality vector, the success rate of whole machine refurbishment can be predicted based on the fused quality vector, and the priority queue of recyclables can be divided based on the prediction result; when the predicted refurbishment rate of a batch of recyclables is lower than the set threshold, a dismantling resource redirection instruction is triggered to guide it to the material-level recycling process instead of the whole machine remanufacturing line.

[0053] Step S203: Utilize the dynamic modeling module to integrate real-time cross-border logistics cost information with environmental and tariff regulations to construct a dynamic cost matrix that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs.

[0054] If the supply chain cost model is static, it cannot reflect real-time fluctuations in logistics costs, frequently updated tariff policies, and emerging carbon emission costs, causing optimization solutions to fail due to "time lag" during actual implementation. For example, a static cost database solution, which uses historical average costs or periodically (e.g., quarterly) updated cost parameters for calculation, may instantly transform the calculated "optimal" path into a "high-cost" or "illegal" path when faced with soaring fuel prices, sudden tariff increases by a country, or the implementation of a new carbon tax policy. To address the shortcomings of the static cost database solution, a dynamic modeling module is used to integrate real-time cross-border logistics cost information with the aforementioned environmental and tariff regulations, constructing a dynamic cost matrix that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs. This dynamic modeling module includes a tariff weight calculation unit, a transportation cost fluctuation factor generation unit, and a carbon emission compliance cost estimation unit. By integrating real-time cost information and regulatory texts, and explicitly introducing multiple dynamic factors such as tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs, the cost model can truly reflect the instantaneous state of the external environment.

[0055] As one embodiment of this application, the dynamic modeling module integrates real-time cross-border logistics cost information with environmental and tariff regulations to construct a dynamic cost matrix that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs. This can be achieved by: real-time capture of updated HTS tariff rates for customs commodities, combined with fuel price indices and dry bulk cargo indices for specific regions, to calculate a comprehensive cost function for cross-border transportation per unit volume; and the introduction of a carbon footprint tracking module to estimate carbon dioxide equivalent emissions per ton-kilometer based on transportation mode and distance, converting this into an equivalent carbon tax cost item and incorporating it into the dynamic cost matrix. This approach allows the cost model to move beyond static averages and reflect real-time market fluctuations. Specifically, the core of the dynamic modeling module is to construct a three-dimensional tensor-based dynamic cost matrix. Each of these items has the ability to be updated in real time:

[0056] Tariff weight The system crawls HTS code tariff change announcements issued by customs authorities of various countries on a daily schedule and combines them with the rules of origin database to determine whether the current transportation route meets the preferential conditions of the free trade agreement. If it does not meet the conditions, it automatically calculates the most-favored-nation tariff rate or even punitive surcharges.

[0057] Transportation cost volatility factor The system integrates variables such as the Baltic Dry Index (BDI), fuel surcharge rate, and port congestion index to establish a nonlinear regression model for estimating the transportation cost per unit of cargo volume. For example, when route G is affected by a conflict, the system immediately increases the cost coefficient for detouring via route H.

[0058] Carbon emission compliance costs The CO2 equivalent emissions per ton-kilometer are calculated based on the ISO 14067 standard, and converted into equivalent financial costs with reference to the current tax rates of the carbon border adjustment mechanism in specific regions.

[0059] During the update process, the dynamic cost matrix in the above embodiments can also perform the following actions: monitor changes in the scope of application of free trade agreements, determine whether the current transportation route complies with the rules of origin; if not, automatically retrieve alternative routes to activate low-tariff channels, and recalculate the components of the dynamic cost matrix under that route. For example, if country J1 and country J2 sign a new agreement, the system will automatically check whether the origin of the goods meets the requirements. After recalculating the components of the dynamic cost matrix, the cost differences and time delay costs between the old and new routes can be compared. If the cost savings exceed a preset proportion and the delivery cycle extension does not exceed the tolerance threshold, a route switching suggestion is initiated and pushed to the scheduling and control system. This decision-making is based on multi-criteria analysis to ensure that cost optimization does not affect timeliness.

[0060] Step S204: Input the quality vector and dynamic cost matrix into the multi-objective optimization engine, and generate the Pareto optimal solution set for remanufacturing center location selection and material flow dynamic scheduling under the conditions of satisfying the preset refurbishment rate threshold, total cost constraint and regional compliance.

[0061] Existing single-objective optimization or simple weighted summation methods cannot effectively balance multiple conflicting objectives such as high rework rate, low cost, and high compliance. Decisions often suffer from trade-offs, making it difficult to obtain a globally satisfactory solution. This is because the results of single-objective optimization are heavily dependent on the constraints; slightly tighter or looser constraints can lead to drastically different results, and they fail to reveal the trade-offs between objectives. Furthermore, the weighting in weighted summation methods is highly subjective, and linear weighting cannot effectively handle nonlinear conflicts between objectives; it can only provide a "compromise" solution rather than a series of optimal trade-offs. Therefore, this application proposes a solution that inputs the quality vector and dynamic cost matrix into a multi-objective optimization engine. Under the conditions of satisfying a preset rework rate threshold, total cost constraints, and regional compliance, it generates a Pareto optimal solution set for remanufacturing center location selection and dynamic material flow scheduling. The reason for using a multi-objective optimization engine to generate Pareto optimal solutions is to acknowledge the inherent conflict between multiple objectives. Instead of forcibly seeking a single best answer, it scientifically reveals the set of all possible optimal trade-offs, providing decision-makers with a comprehensive perspective and flexible options. The aforementioned scheme uses a quality vector representing technical feasibility and a dynamic cost matrix representing economic efficiency and compliance as inputs precisely to consider these dimensions simultaneously during the optimization process.

[0062] In order to retain more valuable compromise solutions for decision-makers to choose from flexibly, as an embodiment of this application, the generation of the Pareto optimal solution set for remanufacturing center location and dynamic material flow scheduling, under the conditions of satisfying the preset refurbishment rate threshold, total cost constraints, and regional compliance, can be achieved through steps S2041 and S2042, as detailed below:

[0063] Step S2041: The improved Non-dominated Sorting Genetic Algorithm III (NSGA-III) algorithm is used to simultaneously optimize the objective function in the three-dimensional decision space: maximize the average renovation rate, minimize the weighted total expenditure, and maximize the regulatory compliance score.

[0064] NSGA-III is a multi-objective evolutionary algorithm that guides the search direction through reference points. The improved NSGA-III algorithm simultaneously optimizes three mutually constraining objective functions: maximizing the average renovation rate, minimizing the weighted total expenditure, and maximizing the regulatory compliance score. For example, maximizing the average renovation rate (≥80%) and minimizing the weighted total expenditure (not exceeding the budget limit) are objectives. In this embodiment, the regulatory compliance score is calculated as follows: a large language model extracts key compliance clauses from unstructured legal text, such as cross-border transfer permit conditions in waste transportation regulations for region K; the extracted clauses are converted into a Boolean logic rule base and used to evaluate the compliance risk index of each candidate site selection scheme, which is then quantified into objective components that can participate in optimization. For example, the large language model parses the regulatory text, extracts rules such as "prohibition of transporting hazardous waste," and encodes them into logical expressions, such as IF Waste Type = Hazardous THEN Compliance Score = 0. Each candidate site selection scheme corresponds to a set of remanufacturing center site combinations and corresponding material allocation ratios. For example, Option A might recommend sending high-quality mobile phones from country L1 to factories in country M1 for refurbishment, while low-quality batches from country L2 would be directed to dismantling centers in country M2 for precious metal extraction. After quantifying these into target components that can participate in optimization, they are maximized through a preset function, i.e., maximizing the regulatory compliance score.

[0065] Step S2042: Dynamically adjust the population evolution direction through the reference point adaptive mechanism to make the solution set converge towards the high-quality, low-cost, and high-compliance cooperative region.

[0066] Specifically, the implementation of step S2042 includes the following steps S1 to S5: S1: In the three-dimensional decision space composed of renovation rate, total expenditure, and regulatory compliance, based on the ideal and worst values ​​of each optimization objective, a set of initial reference points are uniformly set along each coordinate axis to guide the population to cover the entire Pareto front; S2: For each individual solution in the current evolutionary population, the vertical distance or angle between its standardized objective function value vector and all reference points is calculated, and each individual is associated with its nearest reference point; S3: The number of individuals associated with each reference point is counted and defined as the association density of that reference point, based on the... The association density distribution with reference points dynamically reduces the selection pressure of new candidate solutions associated with high-density reference points, while increasing the selection preference for candidate solutions associated with low-density or unassociated reference points; S4: In the evolutionary selection process, individuals associated with low-density reference points are preferentially retained, thereby driving the population to spread to the decision region where the current solution set is sparse; S5: Repeat steps S2 to S4, and through the adaptive guidance mechanism of reference points, enable the population to continuously converge towards the Pareto optimal region where it is difficult to improve high quality, low cost, and high compliance simultaneously during the evolutionary process, thereby obtaining a set of optimal solutions that are evenly distributed among the three objectives.

[0067] Although the system can identify and quantify compliance risks by calculating a compliance risk index, it remains merely a risk warning system. In other words, while it identifies which options are high-risk, it doesn't provide automated solutions. In actual operation, when the system indicates that a candidate site selection option is too risky, operators may need to spend a significant amount of time studying regulations and consulting experts to develop a mitigation strategy. This is not only inefficient but also highly dependent on human experience, prone to errors and delays. Therefore, in the complex global compliance environment, it is essential to transform the static, theoretically optimal solution output by the multi-objective optimization engine into an executable, interference-resistant practical operational solution, ensuring that compliance drift does not occur during execution. Figure 2The example method may further include: when the compliance risk index of a candidate site selection scheme exceeds the warning threshold, activating a compliance avoidance strategy generation subroutine driven by the Large Language Model (LLM); generating multilingual compliance execution guidelines from the compliance avoidance strategy generation subroutine to guide operators in adjusting the dismantling process sequence or changing transport agents to meet specific national regulatory requirements; embedding the multilingual compliance execution guidelines into a smart contract template and deploying it on a blockchain platform to achieve automated verification of the material traceability chain; automatically comparing the consistency between the actual operation log and the contract terms whenever a material handover occurs, and freezing subsequent transfer nodes when discrepancies are found. Specifically, embedding multilingual compliance guidelines into smart contract templates and deploying them on a blockchain platform to automate the verification of the material traceability chain can be achieved as follows: The smart contract template is designed in a modular fashion, comprising a data structure layer, a logic judgment layer, and an execution action layer. The data structure layer defines key entities (e.g., material batches, participant identities, timestamps). The logic judgment layer converts the natural language rules in the compliance guidelines (e.g., "If the transported goods are electronic waste, an import permit from the destination country is required") into machine-executable logical statements (e.g., if-then conditional statements based on Solidity). The execution action layer pre-sets automated responses (e.g., triggering flow or freezing after verifying external data via a blockchain oracle). The multilingual compliance guidelines are parsed using natural language processing technology to extract key parameters (e.g., compliance thresholds, verification conditions) and mapped to corresponding variables and functions in the smart contract template, thus achieving the "embedding" of the guidelines. The compiled smart contract is then deployed to a permissioned blockchain platform (e.g., Hyperledger). Fabric utilizes its distributed ledger characteristics to record the entire material flow process, forming an immutable traceability chain; it achieves automated verification through an event listening mechanism: when IoT sensors or ERP systems detect material handover behavior, the verification function in the smart contract is automatically triggered to compare the consistency of the actual operation logs (e.g., shipping documents, inspection reports) with the contract terms; if the verification passes, the flow permission for the next stage is released; if a deviation is detected, subsequent nodes are automatically frozen and an alarm is sent to the regulator.

[0068] The above solution addresses two main issues. First, by introducing an LLM-driven compliance avoidance strategy generation subroutine, the system moves beyond simple warnings to proactively generating customized and actionable solutions, such as adjusting the dismantling sequence or changing transport agents. This not only enables real-time, automatic comparison of material handover activities but also eliminates the need for manual auditing, improving efficiency and reliability. Second, by embedding guidelines into smart contracts and deploying them on the blockchain, the flexible operational suggestions are transformed into rigid, automatically verifiable program logic, providing a credible foundation for the entire material traceability chain. This is particularly suitable for scenarios requiring compliance verification from regulatory agencies.

[0069] Step S205: In response to changes in the external environment, dynamically adjust the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set to achieve closed-loop optimization.

[0070] The drawback of open-loop optimization models is that they cannot automatically detect and adjust to changes in the external environment, lacking resilience and adaptability. For example, although the optimization model can be rerun at fixed intervals, periodic recalculations cannot cope with unforeseen events, and the system may remain in a suboptimal or even infeasible state for an extended period during the recalculation intervals. To achieve closed-loop feedback and control, this application adopts a solution that dynamically adjusts the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set in response to changes in the external environment, thus achieving closed-loop optimization. This approach transforms optimization from a one-time static planning process into a continuous dynamic adaptive process. This is also the essential characteristic that distinguishes this application from traditional static planning tools, endowing the supply chain system with resilience and adaptability.

[0071] Specifically, as an embodiment of this application, responding to external environmental change events can be achieved through steps S1051 to S2053, as detailed below:

[0072] Step S2051: Continuously monitor policy announcements, natural disaster alerts, or port congestion signals to identify contingencies that may affect existing scheduling strategies.

[0073] Step S2051 can be implemented by subscribing to a news API or IoT sensor data. For example, the system will trigger an assessment when a typhoon warning is detected.

[0074] Step S2052: If an emergency is detected, the reinforcement learning model is invoked to evaluate the expected effectiveness of various emergency response strategies.

[0075] It should be noted that the reinforcement learning model used is trained in a simulated environment and is designed for long-term gains. The training process can be summarized as follows: A simulated environment is constructed to reproduce historical events and their cascading effects; for example, country A imposing tariffs on electronic components in country C leads to a surge in transit warehouses in region Q. Monte Carlo tree search is then run in this simulated environment to explore the long-term gains of different response sequences, continuously optimizing the policy network parameters. Training the reinforcement learning model ensures that it can learn the optimal policy in complex scenarios.

[0076] Step S2053: Select the strategy with the highest cumulative reward recommended by the reinforcement learning model as the response action. The response action includes switching the transportation mode, temporarily transferring the production load to the backup center, and releasing the safety stock buffer, etc.

[0077] After the reinforcement learning model recommends the strategy with the highest cumulative reward, the recommended strategy can be decomposed into an executable sequence of operations and injected into the task queue of the Enterprise Resource Planning (ERP) system. The entire process of the strategy implementation is simulated through a digital twin platform to confirm that there are no major conflicts before a formal scheduling instruction is issued.

[0078] To ensure that the static scheduling scheme, which is based on historical data and the optimal solution at a certain moment (Pareto solution set), can maintain long-term stable and efficient operation when faced with unavoidable fluctuations in the actual production environment, and avoid the overall system efficiency from declining due to local resource overload. Figure 2 The example method could also include: establishing a cross-regional remanufacturing center load balancing feedback loop to periodically collect the actual processing capacity utilization of each center; when a center is overloaded for several consecutive cycles, dynamically redistributing the upstream recyclable flow based on the Pareto optimal solution set to avoid the formation of local bottlenecks. This scheme, on the one hand, continuously monitors the system's operating status through the load balancing feedback loop. Once a deviation between reality and expectations is detected (a center is continuously overloaded), it immediately triggers re-optimization, enabling the system to self-correct and ensure it always operates efficiently and smoothly. On the other hand, the scheme does not rely on perfect predictions from the initial model but acknowledges the uncertainties of the real world. By dynamically redistributing the upstream recyclable flow, the system can adaptively absorb the impact of these uncertainties, preventing local problems from spreading into global paralysis.

[0079] Specifically, as an embodiment of this application, the above example of dynamically redistributing the upstream recycled material flow direction based on the Pareto optimization results of the Pareto optimal solution set when a certain center is overloaded for multiple consecutive cycles, to avoid the formation of local bottlenecks, can be achieved by: continuously collecting the actual processing capacity utilization data of each remanufacturing center through a load balancing feedback loop; when the utilization rate of a target center exceeds the safety threshold for multiple consecutive preset cycles, determining that the target center is in a state of continuous overload operation; in response to the determination of the overload state, generating a global re-optimization trigger signal, which at least includes the identifier of the overloaded target center and its current load information; injecting the re-optimization trigger signal as a new constraint into the improved NSGA-III algorithm to construct a... A new decision space is established, which excludes schemes that allocate new recyclables to the overloaded target center and / or includes redistribution schemes that alleviate its existing load. In the new decision space with load constraints, the optimization algorithm is re-run to generate a new set of Pareto optimal solutions. This set of solutions must satisfy the constraint of eliminating the overloaded state of the target center while maximizing the refurbishment rate, minimizing total expenditure, and maximizing regulatory compliance. A preferred solution is selected from the new Pareto optimal solution set, and the allocation direction of upstream recyclables to each remanufacturing center is dynamically adjusted according to the scheduling scheme of the solution. Some or all of the recyclables originally planned to flow to the overloaded target center are redirected to other centers with surplus processing capacity, thereby eliminating local bottlenecks and restoring the overall load balance of the system.

[0080] From the above appendix Figure 2The example of the circular supply chain material flow optimization method demonstrates two key advantages. First, by semantically parsing the quality inspection reports of recycled products using a large language model and combining this with physical imaging analysis results to generate a quality vector representing the remanufacturing potential of recycled products, optimization decisions are no longer limited to a few structured data points but are based on richer and more accurate information. This provides more reliable input for subsequent optimization, transforming the originally static cost parameters into dynamic variables that can be updated in real time with market conditions and policies. This allows the optimization model to promptly capture changes in cost structure and compliance requirements, resulting in a calculated solution that more closely reflects the current situation and effectively avoiding the risk of cost underestimation or compliance violations due to information lag. Second, by inputting the quality vector and dynamic cost matrix into the multi-objective optimization engine, while satisfying preset refurbishment rate thresholds and total cost constraints, the optimization can achieve optimal results. Under the conditions of regional compliance, this method generates a Pareto optimal solution set for remanufacturing center location selection and dynamic material flow scheduling. It provides not a single solution, but a series of optimal trade-offs (Pareto solutions), enabling decision-makers to choose based on strategic preferences. This allows them to find a globally optimal solution in complex multi-objective decision-making, rather than a locally optimal one. Thirdly, by responding to changes in the external environment, the method dynamically adjusts the logistics paths and remanufacturing resource allocation strategies corresponding to the Pareto optimal solution set, achieving closed-loop optimization. This transforms the supply chain system from a static, one-off planning system into a dynamic, continuously learning, adaptive system. When external disturbances occur, the system can proactively trigger re-optimization and adjust resource allocation, significantly improving the resilience and robustness of the supply chain and ensuring near-optimal operation in dynamic and uncertain environments. In summary, the technical solution of this application, by integrating multi-source data and constructing a dynamic model, can achieve real-time optimization and dynamic adjustment under multiple objectives, improving decision-making accuracy and supply chain resilience.

[0081] Please see Figure 3 As shown, in one embodiment, a circular supply chain material flow optimization device is provided. This device may include an acquisition module 301, a first generation module 302, a construction module 303, a second generation module 304, and an adjustment module 305, as detailed below:

[0082] The acquisition module 301 is used to collect multi-source heterogeneous data, which includes quality inspection reports of recyclables from different regions around the world, real-time cost information of cross-border logistics, and texts of environmental protection and tariff regulations of various countries.

[0083] The first generation module 302 is used to perform semantic parsing of the quality inspection report of recycled products through the large language model LLM, and generate a quality vector characterizing the remanufacturing potential of recycled products by combining the physical imaging analysis results.

[0084] Module 303 is used to integrate real-time cost information of cross-border logistics with environmental protection and tariff regulations by utilizing the dynamic modeling module to construct a dynamic cost matrix that includes tariff weights, transportation cost fluctuation factors and carbon emission compliance costs.

[0085] The second generation module 304 is used to input the quality vector and dynamic cost matrix into the multi-objective optimization engine to generate a Pareto optimal solution set for remanufacturing center location selection and material flow dynamic scheduling under the conditions of satisfying the preset refurbishment rate threshold, total cost constraint and regional compliance.

[0086] The adjustment module 305 is used to respond to external environmental change events and dynamically adjust the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set to achieve closed-loop optimization.

[0087] From the above appendix Figure 3 As illustrated by the example of the circular supply chain material flow optimization device, firstly, by semantically parsing the quality inspection report of recycled products using a large language model and combining it with physical imaging analysis results to generate a quality vector characterizing the remanufacturing potential of recycled products, the optimization decision is no longer limited to a few structured data points but is based on richer and more accurate information. This provides more reliable input for subsequent optimization, transforming the originally static cost parameters into dynamic variables that can be updated in real time with market conditions and policies and regulations. This allows the optimization model to promptly capture changes in cost structure and compliance requirements, resulting in a calculated solution that is closer to the current reality and effectively avoiding the risk of cost underestimation or compliance violations due to information lag. Secondly, by inputting the quality vector and dynamic cost matrix into the multi-objective optimization engine, while meeting preset refurbishment rate thresholds and total cost constraints, the optimization can achieve optimal results. Under the conditions of regional compliance, this method generates a Pareto optimal solution set for remanufacturing center location selection and dynamic material flow scheduling. It provides not a single solution, but a series of optimal trade-offs (Pareto solutions), enabling decision-makers to choose based on strategic preferences. This allows them to find a globally optimal solution in complex multi-objective decision-making, rather than a locally optimal one. Thirdly, by responding to changes in the external environment, the method dynamically adjusts the logistics paths and remanufacturing resource allocation strategies corresponding to the Pareto optimal solution set, achieving closed-loop optimization. This transforms the supply chain system from a static, one-off planning system into a dynamic, continuously learning, adaptive system. When external disturbances occur, the system can proactively trigger re-optimization and adjust resource allocation, significantly improving the resilience and robustness of the supply chain and ensuring near-optimal operation in dynamic and uncertain environments. In summary, the technical solution of this application, by integrating multi-source data and constructing a dynamic model, can achieve real-time optimization and dynamic adjustment under multiple objectives, improving decision-making accuracy and supply chain resilience.

[0088] Optionally, the above Figure 3The first generation module 302 in the example may include a recognition unit, an extraction unit, an alignment unit, and a quality vector generation unit, wherein:

[0089] The recognition unit is used to identify and standardize non-English quality rating terms using a multilingual pre-trained BERT architecture, mapping non-English quality rating terms to a unified quality level coding system.

[0090] The extraction unit is used to simultaneously extract structured fields from the quality inspection report of recycled products regarding the equipment's service life, maintenance history, and functional test results, as one of the input dimensions of the quality vector;

[0091] The alignment unit is used to perform cross-modal alignment between the parsed quality level code and the electronic component integrity image output by the physical imaging system to obtain aligned text semantic features and image spatial features.

[0092] The quality vector generation unit is used to fuse aligned text semantic features and image spatial features based on an attention mechanism to generate a fused quality vector. In the fused quality vector, samples with high renovation probability have a higher clustering density in the vector space.

[0093] Optionally, the above Figure 3 Example building block 303 may include a computation unit and an estimation unit, wherein:

[0094] The calculation unit is used to capture real-time updates of customs commodity code HTS tax rates and combine them with the fuel price index and the dry bulk cargo index of a specific region to calculate the comprehensive cost function of cross-border transportation per unit of cargo volume.

[0095] The estimation unit is used to introduce the carbon footprint tracking module to estimate the carbon dioxide equivalent emissions per ton-kilometer based on the mode of transportation and distance, and convert it into an equivalent carbon tax cost item and incorporate it into the dynamic cost matrix.

[0096] Optionally, the above Figure 3 The example's second generation module 304 may include an optimization unit and an adjustment unit, wherein:

[0097] The optimization unit is used to simultaneously optimize the objective function in the three-dimensional decision space using the improved NSGA-III algorithm: maximizing the average renovation rate, minimizing the weighted total expenditure, and maximizing the regulatory compliance score;

[0098] The adjustment unit is used to dynamically adjust the direction of population evolution through a reference point adaptive mechanism, so that the solution set converges to a high-quality, low-cost, and high-compliance cooperative region.

[0099] Optionally, Figure 3 The example device may also include a driver module, a guide generation module, an embedding module, and a comparison module, wherein:

[0100] The driving module is used to start the compliance avoidance strategy generation subroutine driven by the Large Language Model (LLM) when the compliance risk index of a candidate site selection scheme exceeds the warning threshold.

[0101] The guide generation module is used by the compliance avoidance strategy generation subroutine to generate multilingual compliance execution guidelines to guide operators to adjust the dismantling process sequence or change transport agents to meet the regulatory requirements of specific countries.

[0102] An embedded module is used to embed multilingual compliance implementation guidelines into smart contract templates and deploy them on a blockchain platform to achieve automated verification of the material traceability chain;

[0103] The comparison module is used to automatically compare the actual operation log with the contract terms whenever a material handover occurs, and freeze subsequent transfer nodes when a discrepancy is found.

[0104] Optionally, the above Figure 3 The example adjustment module 305 includes a listening unit, a calling unit, and a selection unit, wherein:

[0105] The monitoring unit is used to continuously monitor policy announcements, natural disaster warnings, or port congestion signals to identify emergencies that may affect existing scheduling strategies.

[0106] The invocation unit is used to invoke a reinforcement learning model to evaluate the expected effectiveness of various emergency response strategies if a sudden event is detected.

[0107] The selection unit is used to select the strategy with the highest cumulative reward recommended by the reinforcement learning model as the response action. The response action includes switching the transportation mode, temporarily transferring the production load to the backup center, and releasing the safety stock buffer.

[0108] Optionally, Figure 3 The example device may also include a periodic acquisition module and a redistribution module, wherein:

[0109] The periodic data collection module is used to establish a load balancing feedback loop across regional remanufacturing centers and periodically collect the actual processing capacity utilization of each center.

[0110] The redistribution module is used to dynamically redistribute the upstream recyclable flow direction by combining the Pareto optimization results of the Pareto optimal solution set when a center is overloaded for multiple consecutive cycles, so as to avoid the formation of local bottlenecks.

[0111] In one embodiment, a computer device is provided, the computer device may be... Figure 1 The internal structure diagram of the circular supply chain material flow optimization system, as shown in the application scenario example, can be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When executed by the processor, the computer program implements the functions or steps of a circular supply chain material flow optimization method.

[0112] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the following steps:

[0113] Collect multi-source heterogeneous data, including quality inspection reports of recyclables from different regions around the world, real-time cost information of cross-border logistics, and texts of environmental protection and tariff regulations of various countries;

[0114] The quality inspection report of recycled products is semantically parsed using a large language model (LLM), and a quality vector representing the remanufacturing potential of recycled products is generated by combining the results of physical imaging analysis.

[0115] By integrating real-time cost information of cross-border logistics with environmental and tariff regulations using a dynamic modeling module, a dynamic cost matrix is ​​constructed that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs.

[0116] By inputting the quality vector and dynamic cost matrix into the multi-objective optimization engine, Pareto optimal solution sets for remanufacturing center location selection and dynamic scheduling of material flow are generated under the conditions of satisfying the preset refurbishment rate threshold, total cost constraints and regional compliance.

[0117] In response to changes in the external environment, the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set are dynamically adjusted to achieve closed-loop optimization.

[0118] The aforementioned computer program, by fusing multi-source data and constructing dynamic models, can achieve real-time optimization and dynamic adjustment under multiple objectives, thereby improving decision-making accuracy and supply chain resilience.

[0119] In one embodiment, a computer-readable storage medium is provided that stores a computer program, which, when executed by a processor, performs the following steps:

[0120] Collect multi-source heterogeneous data, including quality inspection reports of recyclables from different regions around the world, real-time cost information of cross-border logistics, and texts of environmental protection and tariff regulations of various countries;

[0121] The quality inspection report of recycled products is semantically parsed using a large language model (LLM), and a quality vector representing the remanufacturing potential of recycled products is generated by combining the results of physical imaging analysis.

[0122] By integrating real-time cost information of cross-border logistics with environmental and tariff regulations using a dynamic modeling module, a dynamic cost matrix is ​​constructed that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs.

[0123] By inputting the quality vector and dynamic cost matrix into the multi-objective optimization engine, Pareto optimal solution sets for remanufacturing center location selection and dynamic scheduling of material flow are generated under the conditions of satisfying the preset refurbishment rate threshold, total cost constraints and regional compliance.

[0124] In response to changes in the external environment, the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set are dynamically adjusted to achieve closed-loop optimization.

[0125] When the aforementioned computer program is executed by the processor, it integrates multi-source data and constructs a dynamic model, enabling real-time optimization and dynamic adjustment under multiple objectives, thereby improving decision-making accuracy and supply chain resilience.

[0126] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0127] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0128] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0129] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for optimizing material flow in a circular supply chain, characterized in that, Includes the following steps: Collect multi-source heterogeneous data, including quality inspection reports of recyclables from different regions around the world, real-time cost information of cross-border logistics, and texts of environmental protection and tariff regulations of various countries; The quality inspection report of the recycled products is semantically parsed using a large language model (LLM), and a quality vector characterizing the remanufacturing potential of the recycled products is generated by combining the results of physical imaging analysis. By using the dynamic modeling module to integrate the real-time cost information of cross-border logistics with the text of environmental protection and tariff regulations, a dynamic cost matrix is ​​constructed that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs. The quality vector and the dynamic cost matrix are input into the multi-objective optimization engine to generate a Pareto optimal solution set for remanufacturing center location selection and dynamic scheduling of material flow, under the conditions of satisfying the preset refurbishment rate threshold, total cost constraint and regional compliance. In response to changes in the external environment, the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set are dynamically adjusted to achieve closed-loop optimization. The process of semantically parsing the quality inspection report of recycled products using a large language model and generating a quality vector representing the remanufacturing potential of recycled products by combining the results of physical imaging analysis includes: using a multilingual pre-trained bidirectional encoder representation model BERT to identify and standardize non-English quality rating terms, mapping these terms to a unified quality level coding system; simultaneously extracting structured fields from the quality inspection report regarding equipment service life, maintenance history, and functional test results as one of the input dimensions of the quality vector; performing cross-modal alignment between the parsed quality level coding and the electronic component integrity image output by the physical imaging system to obtain aligned text semantic features and image spatial features; and fusing the aligned text semantic features and image spatial features based on an attention mechanism to generate a fused quality vector, wherein high refurbishment probability samples in the fused quality vector have a higher clustering density in the vector space. The method of using a dynamic modeling module to integrate real-time cost information of cross-border logistics with environmental protection and tariff regulations to construct a dynamic cost matrix that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs includes: real-time capture of updated customs commodity code (HTS) tariff rate data, combined with fuel price index and dry bulk index of specific regions, to calculate the comprehensive cost function of cross-border transportation per unit volume of cargo; and the introduction of a carbon footprint tracking module to estimate the carbon dioxide equivalent emissions per ton-kilometer based on transportation mode and distance, and convert it into an equivalent carbon tax cost item and incorporate it into the dynamic cost matrix.

2. The method for optimizing material flow in a circular supply chain according to claim 1, characterized in that, The process of generating a Pareto optimal solution set for remanufacturing center location selection and dynamic material flow scheduling, under the conditions of satisfying a preset refurbishment rate threshold, total cost constraints, and regional compliance, includes: An improved non-dominated sorting genetic algorithm, NSGA-III, is used to simultaneously optimize the objective function in a three-dimensional decision space: maximizing the average renovation rate, minimizing the weighted total expenditure, and maximizing the regulatory compliance score. By dynamically adjusting the direction of population evolution through a reference point adaptive mechanism, the solution set converges towards a collaborative region characterized by high quality, low cost, and high compliance.

3. The method for optimizing material flow in a circular supply chain according to claim 2, characterized in that, The method further includes: When the compliance risk index of a candidate site selection scheme exceeds the warning threshold, the compliance avoidance strategy generation subroutine driven by the Large Language Model (LLM) is activated. The compliance avoidance strategy generation subroutine generates multilingual compliance execution guidelines to guide operators in adjusting the dismantling process sequence or changing transport agents to meet the regulatory requirements of specific countries. The multilingual compliance implementation guidelines are embedded into a smart contract template and deployed on a blockchain platform to achieve automated verification of the material traceability chain; Whenever a material handover occurs, the system automatically compares the actual operation log with the contract terms and freezes subsequent transfer nodes if any discrepancies are found.

4. The method for optimizing material flow in a circular supply chain according to claim 1, characterized in that, The responses to external environmental changes include: Continuously monitor policy announcements, natural disaster warnings, or port congestion signals to identify unforeseen events that may affect existing scheduling strategies; If a sudden event is detected, the reinforcement learning model is invoked to evaluate the expected effectiveness of various emergency response strategies. The strategy with the highest cumulative reward recommended by the reinforcement learning model is selected as the response action, which includes switching the transportation mode, temporarily transferring the production load to the backup center, and releasing the safety stock buffer.

5. The method for optimizing material flow in a circular supply chain according to claim 1, characterized in that, The method further includes: Establish a cross-regional remanufacturing center load balancing feedback loop and regularly collect the actual processing capacity utilization of each center; When a center operates under overload for multiple consecutive cycles, the upstream recyclables flow is dynamically redistributed based on the Pareto optimization results of the Pareto optimal solution set to avoid the formation of local bottlenecks.

6. A circular supply chain material flow optimization device, characterized in that, The device includes: The acquisition module is used to collect multi-source heterogeneous data, which includes quality inspection reports of recyclables from different regions around the world, real-time cost information of cross-border logistics, and texts of environmental protection and tariff regulations of various countries. The first generation module is used to perform semantic parsing of the recycled product quality inspection report using a large language model (LLM) and generate a quality vector representing the remanufacturing potential of the recycled product by combining the physical imaging analysis results. The process of semantic parsing of the recycled product quality inspection report using a large language model and generating the quality vector representing the remanufacturing potential of the recycled product by combining the physical imaging analysis results includes: using a multilingual pre-trained bidirectional encoder representation model (BERT) to identify and standardize non-English quality rating terms, mapping the non-English quality rating terms to a unified quality level coding system; simultaneously extracting structured fields from the recycled product quality inspection report regarding equipment service life, maintenance history, and functional test results as one of the input dimensions of the quality vector; performing cross-modal alignment between the parsed quality level coding and the electronic component integrity image output by the physical imaging system to obtain aligned text semantic features and image spatial features; and fusing the aligned text semantic features and image spatial features based on an attention mechanism to generate a fused quality vector, wherein samples with high refurbishment probability have a higher clustering density in the vector space. The construction module is used to integrate the real-time cost information of cross-border logistics and the text of environmental protection and tariff regulations using the dynamic modeling module to construct a dynamic cost matrix that includes tariff weights, transportation cost fluctuation factors, and carbon emission compliance costs. This process involves: real-time capture of updated customs commodity code (HTS) tariff rate data, combined with fuel price indices and dry bulk cargo indices for specific regions, to calculate a comprehensive cost function for cross-border transportation per unit volume; and the introduction of a carbon footprint tracking module to estimate the carbon dioxide equivalent emissions per ton-kilometer based on transportation mode and distance, and converting this into an equivalent carbon tax cost item for inclusion in the dynamic cost matrix. The second generation module is used to input the quality vector and the dynamic cost matrix into the multi-objective optimization engine, and generate a Pareto optimal solution set for remanufacturing center location selection and material flow dynamic scheduling under the conditions of satisfying the preset refurbishment rate threshold, total cost constraint and regional compliance. The adjustment module is used to respond to external environmental changes and dynamically adjust the logistics path and remanufacturing resource allocation strategy corresponding to the Pareto optimal solution set to achieve closed-loop optimization.

7. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the circular supply chain material flow optimization method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the circular supply chain material flow optimization method as described in any one of claims 1 to 5.

Images