Logistics operation management and decision-making method, device and equipment based on large model

Abstract

The invention discloses a logistics operation management and decision-making method, device and equipment based on a large model, and the method comprises the steps: collecting structured data and unstructured data in logistics operation in real time, converting the unstructured data into a text vector through a preset embedded model, building an index association with a key field of the structured data, and carrying out the index association with the key field of the unstructured data; constructing a supply chain knowledge base; when the external event information is monitored, extracting an event entity of the external event information by using a large language model, and performing association analysis in the supply chain knowledge base based on the event entity to generate a risk event object; receiving a query statement, decomposing the query statement based on the risk event object by using a large language model to obtain a plurality of sub-tasks, and processing retrieval data obtained by performing retrieval based on different sub-tasks to generate abstract type answers and a plurality of alternative solutions; and taking a scheme selected by the user from the plurality of alternative solutions as a decision scheme, and generating and executing a cross-department workflow task.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and in particular to a method, apparatus and equipment for logistics operation management and decision-making based on a large model. Background Technology

[0002] In modern supply chain management, enterprises face increasingly complex global operating environments and uncertainties. To achieve effective monitoring of the entire supply chain, including procurement, production, warehousing, logistics, and delivery, supply chain control towers have emerged as centralized information hubs. These towers aim to provide end-to-end visualization support by integrating structured data from business systems such as order management, warehouse management, and transportation management. Traditional control tower systems primarily rely on pre-set dashboards, fixed reports, and threshold-based rule engines for status display and anomaly alerts, possessing some capabilities in structured data processing.

[0003] However, with the dynamic changes in the internal and external operating environment, significant limitations remain when facing the multi-source, heterogeneous, and highly dynamic characteristics of modern supply chains. Existing technological solutions primarily rely on structured data processing at the data level, struggling to effectively integrate and semantically understand massive amounts of unstructured information (such as supplier notifications, industry news, and weather reports). This information often serves as an early indicator of supply chain risks, and current technologies cannot semantically correlate it with internal business entities, resulting in "information silos" and a "semantic gap" in risk perception. At the human-computer interaction level, users must navigate complex interfaces for data filtering and multi-dimensional analysis, requiring complex multi-system operations and professional training. They cannot engage in interactive exploration and decision support using natural language, leading to high costs and low efficiency in information acquisition and understanding. Furthermore, cross-departmental collaborative execution still heavily relies on human experience and offline communication, lacking information synchronization mechanisms, resulting in low efficiency and failing to meet the real-time, resilience, and intelligent decision-making requirements of modern supply chains. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a logistics operation management and decision-making method, device, equipment and computer program product based on a large model, which aims to solve the technical problems of insufficient intelligence in data fusion, risk perception and collaborative response in existing supply chain management solutions.

[0005] To achieve the above objectives, this invention provides a logistics operations management and decision-making method based on a large model, the method comprising: The system collects both structured and unstructured data from logistics operations in real time. It uses a pre-defined embedding model to convert unstructured data into text vectors and establishes an index association with key fields of structured data to build a unified supply chain knowledge base. When external event information is detected, the event entities of the external event information are extracted using a large language model, and the event entities are used for correlation analysis in the supply chain knowledge base to generate structured risk event objects. The system receives a query statement input by the user, decomposes the query statement based on the risk event object using a large language model to obtain multiple sub-tasks, processes the retrieval data obtained based on different sub-tasks, and generates a summary-style answer and multiple alternative solutions. The solution selected by the user from multiple alternative solutions is taken as the decision solution, and the large language model is invoked to generate and execute cross-departmental workflow tasks based on the decision solution.

[0006] Preferably, the real-time data collection for logistics operations includes both structured and unstructured data, including: Data is collected by connecting to business systems including ERP, WMS, TMS, and CRM systems through application programming interfaces (APIs) to obtain structured data. Unstructured data, including supplier emails, news reports, weather services, social media, and customs announcements, is collected from external information sources through web crawlers or third-party data interfaces.

[0007] Preferably, the step of converting unstructured data into text vectors using a preset embedding model and establishing an index association with key fields of structured data includes: The text corresponding to unstructured data is vectorized using a pre-defined embedding model to obtain text vectors; Key entities are extracted from unstructured data as metadata, and the metadata is bound and associated with text vectors and stored in the supply chain knowledge base. Key entities include location information, event type, organization name and / or material identifier. By matching key entities in the metadata with corresponding key fields in the structured data, an index association is established between unstructured data and structured data.

[0008] Preferably, when external event information is detected, the event entities of the external event information are extracted using a large language model, and correlation analysis is performed on the event entities in the supply chain knowledge base to generate structured risk event objects, including: The external event information is semantically understood and entity recognized using a large language model, and event entities including event type, location of occurrence, time window of impact, and scope of impact are extracted. Based on the event entity, the supply chain entity is searched and matched in the supply chain knowledge base to identify the supply chain entity with influence, and the business system interface is called to obtain the current business data corresponding to the supply chain entity. The current business data includes order planning time, transportation route and inventory level. Using a large language model, risk impact analysis is performed based on the event entities and the current business data to generate risk impact analysis results. Based on the risk impact analysis results, the risk event object is generated, including event identifier, risk level, list of entities affecting the supply chain, specific risk causes, and response recommendations.

[0009] Preferably, the step of using a large language model to perform risk impact analysis based on the event entity and the current business data, and generating risk impact analysis results, includes: The overlap between the impact time window of external events and the order planning time is analyzed to obtain the time dimension analysis results; By analyzing the spatial correlation between the impact area of ​​external events and the geographical location and transportation routes of supply chain entities, spatial dimension analysis results are obtained. Based on a preset risk matrix, the risk impact analysis results are evaluated by combining the time dimension analysis results and the spatial dimension analysis results, resulting in the risk impact analysis results including the risk level.

[0010] Preferably, the step of using a large language model to decompose the query statement based on the risk event object to obtain multiple sub-tasks includes: The semantic intent of the query statement is identified using a large language model, and the query statement is decomposed in combination with the business context provided by the risk event object to generate multiple sub-tasks, including counting affected orders, querying alternative resources, and assessing cost impact. Based on the semantic intent and business context, the ReAct framework is used to plan the execution sequence for each subtask, and to assign a calling tool and generate an execution instruction containing query conditions, filtering parameters and return fields for each subtask in the execution sequence; wherein, the type of calling tool includes SQL query engine, business system API interface or vector retrieval interface.

[0011] Preferably, the process of processing the retrieval data obtained from different subtasks to generate a summary-style answer and multiple alternative solutions includes: The data is retrieved from the business system and supply chain knowledge base according to the execution instruction; The retrieved data is analyzed using a large language model to generate a summary-style answer that includes a description of the current situation and a quantitative analysis of the business impact. By combining a pre-set business rule base with historical records of similar events retrieved from the supply chain knowledge base, multiple alternative solutions are generated, including resource allocation methods, delivery plan adjustments, and customer communication strategies.

[0012] Preferably, the step of invoking the large language model to generate and execute cross-departmental workflow tasks based on the decision scheme includes: The decision scheme is analyzed using a large language model, and task elements including task type, execution action and participating departments are identified based on a predefined workflow template. Based on the task elements, generate structured instructions that include task allocation schemes and execution priorities. Then, based on the structured instructions, call the corresponding business system's API interface to create and distribute workflow tasks to the corresponding participating departments.

[0013] To achieve the above objectives, the present invention also provides a logistics operation management and decision-making device based on a large model, the device comprising: The data processing unit is used to collect both structured and unstructured data in logistics operations in real time. It uses a preset embedding model to convert unstructured data into text vectors and establishes an index association with key fields of structured data to build a unified supply chain knowledge base. The monitoring unit is used to extract event entities from external event information using a large language model when external event information is detected, and to perform correlation analysis on the event entities in the supply chain knowledge base to generate structured risk event objects. The solution generation unit is used to receive the query statement input by the user, decompose the query statement based on the risk event object using a large language model to obtain multiple sub-tasks, process the retrieval data obtained based on different sub-tasks, and generate a summary answer and multiple alternative solutions. The decision-making unit is used to select the solution chosen by the user from multiple alternative solutions as the decision solution, and call the large language model to generate and execute cross-departmental workflow tasks based on the decision solution.

[0014] To achieve the above objectives, the present invention also proposes a logistics operation management and decision-making device based on a large model, including a processor, a memory, and a computer program stored in the memory. The computer program is executed by the processor to implement the steps of a logistics operation management and decision-making method based on a large model as described in the above embodiments.

[0015] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a computer program that is executed by a processor to implement the steps of a large-model-based logistics operation management and decision-making method as described in the above embodiments.

[0016] To achieve the above objectives, the present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of a large-model-based logistics operation management and decision-making method as described in the above embodiments.

[0017] Beneficial effects: The above solutions, by constructing a unified supply chain knowledge base, achieve deep integration of massive heterogeneous data and core business data, breaking down the "data silos" and "semantic gaps" present in traditional systems. Utilizing large language models for automatic risk event identification and attribution fundamentally transforms supply chain management from passive monitoring to proactive intelligent decision-making, significantly shortening the risk discovery and awareness cycle. By decomposing natural language queries into executable data retrieval tasks and generating decision suggestions with multi-scenario simulations, the barriers to data analysis are greatly lowered, enabling users to quickly gain in-depth insights through conversational interaction, improving the efficiency and scientific rigor of decision-making. Furthermore, by automatically transforming decision-making solutions into executable cross-departmental workflows, an end-to-end automated closed loop from decision-making to action is achieved, significantly improving the speed of supply chain anomaly response and the efficiency of cross-departmental collaboration, thereby enhancing the adaptability and intelligence of supply chain management.

[0018] By clearly defining the scope of data collection from multiple internal and external sources, comprehensive coverage of all elements of the supply chain information was achieved, laying a solid data foundation for all-round risk perception and effectively breaking down data silos in traditional systems. By converting text into vectors and extracting key entities as metadata, machines can understand the semantics of unstructured data. By matching metadata with structured data fields, the correlation between external events and internal business data was established efficiently and accurately, laying a solid data foundation for subsequent accurate risk identification and decision analysis.

[0019] By extracting event entities through LLM and combining them with real-time business data for in-depth analysis, risk analysis is no longer limited to the event itself but can be integrated with specific business contexts. This achieves a leap from information to cognition, avoiding false positives or false negatives caused by the lack of contextual association in traditional systems. By introducing quantitative analysis along both temporal and spatial dimensions and combining it with a pre-defined risk matrix for assessment, risk level determination becomes a dynamic and accurate assessment process based on business context. This significantly improves the accuracy and reliability of risk warnings and the resilience of the supply chain, providing high-quality, structured input for subsequent decision support.

[0020] The ReAct framework automatically breaks down complex natural language queries into executable task sequences, endowing the system with human-expert-like task planning and tool invocation capabilities. This allows users to perform in-depth analyses that previously required specialized knowledge and complex operations through dialogue. By integrating real-time data, business rules, and historical experience, it generates data-driven and feasible multi-solution recommendations, thereby improving the efficiency and scientific rigor of strategic decision-making. By parsing natural language decision-making schemes into structured task instructions and automatically calling business system APIs to create and distribute tasks, it solves the problem of relying on manual communication for cross-departmental collaboration in traditional models. It achieves intelligent workflow orchestration and triggering, ensuring that optimal decisions can be efficiently translated into collaborative actions across departments, shortening the response cycle from risk identification to actual intervention, reducing communication costs and human error, and realizing end-to-end automation of supply chain management. 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.

[0022] Figure 1 This is a flowchart illustrating a logistics operation management and decision-making method based on a large model, as provided in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of a logistics operation management and decision-making device based on a large model, provided as an embodiment of the present invention.

[0024] The realization of the invention's objective, its functional characteristics, and advantages will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. 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. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the 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.

[0026] The present invention will be described in detail below with reference to the embodiments.

[0027] Reference Figure 1 The diagram shown is a flowchart of a logistics operation management and decision-making method based on a large model, provided by an embodiment of the present invention.

[0028] In this embodiment, the method is implemented based on a supply chain control platform, which includes a data source perception module, a data fusion and processing module, an AI engine module, an application service module, and a user interaction module. The data source perception module is used to connect to and collect multi-source heterogeneous data from business systems and external information systems. The data fusion and processing module is used to clean, standardize, and semantically integrate the collected multi-source heterogeneous data, and construct a supply chain knowledge base. The AI ​​engine module integrates a large language model and an AI agent framework to perform semantic parsing and entity extraction on incoming external events, and to conduct risk correlation analysis and in-depth assessment based on the supply chain knowledge base. The application service module encapsulates the capabilities of the AI ​​engine into reusable microservice interfaces, including risk perception services, attribution analysis services, intelligent question answering services, solution simulation services, and collaborative workflow services. The user interaction module provides a conversational user interface for interaction with the platform.

[0029] The method includes: S11 collects both structured and unstructured data from logistics operations in real time. It uses a pre-defined embedding model to convert unstructured data into text vectors and establishes an index association with key fields of structured data to build a unified supply chain knowledge base.

[0030] Furthermore, in step S11, the real-time collection of logistics operations includes both structured and unstructured data, including: S11-1 connects to business systems including ERP, WMS, TMS, and CRM systems via application programming interfaces to collect data and obtain structured data; S11-2, collecting unstructured data from external information sources, including supplier emails, news reports, weather services, social media, and customs announcements, through web crawlers or third-party data interfaces.

[0031] Furthermore, in step S11, the process of converting unstructured data into text vectors using a preset embedding model and establishing an index association with key fields of structured data includes: S11-3, The text corresponding to the unstructured data is vectorized using a preset embedding model to obtain text vectors; S11-4 Extract key entities from unstructured data as metadata, and bind and associate the metadata with text vectors before storing them in the supply chain knowledge base. Key entities include location information, event type, organization name and / or material identifier. S11-5, by matching the key entities in the metadata with the corresponding key fields in the structured data, an index association is established between the unstructured data and the structured data.

[0032] In this embodiment, business systems such as ERP (orders), WMS (inventory), TMS (transit management system), and CRM (customer management system) are accessed via API interfaces or direct database connections. Structured data is collected in real time from these internal business systems. Specifically, by interfacing with the ERP system, order data including purchase order numbers, supplier information, material codes, and delivery dates is obtained; real-time inventory records, including warehouse location, material number, and current inventory quantity, are extracted from the WMS system; the transportation status, current location, and estimated arrival time of goods in transit are received from the TMS system; and basic customer information and order-related data are obtained from the CRM system. This structured data, after data cleaning and format standardization, serves as the foundation for structured business data. Unstructured data, such as supplier emails, news portals, weather services, social media, and customs announcements, are obtained through web crawling services or third-party data APIs. For example, a distributed web crawler cluster can be used to periodically crawl supply chain-related information published on news portals, including port operation status, geopolitical events, and industry trends; by accessing the API interface of meteorological data service providers, weather forecasts and disaster warnings can be obtained from around the world; an email parsing engine can be used to automatically process supply anomaly notifications sent by suppliers; and social media APIs can be used to monitor public opinion information related to the supply chain.

[0033] Embedding models pre-trained with text-embedding-3-large or doubao-embedding-text are used to vectorize the collected unstructured text data. For example, when receiving news text such as "The Central Meteorological Observatory issued a red typhoon warning, predicting that it will make landfall in XX city within the next 24 hours, and the city's port will suspend operations," the model converts it into a 768-dimensional floating-point vector (text vector), which accurately represents the deep meaning of the text in the semantic space. Using natural language processing techniques or by calling a large language model through prompting engineering, key entities are extracted from the text as metadata, such as the location entity: XX city, and the event type entity: typhoon warning. This metadata is then bound to the text vector of the news article and stored in the vector database of the supply chain knowledge base, with an HNSW index established to support efficient similarity retrieval.

[0034] When establishing index associations, the location entity in the metadata (e.g., City XX) is matched with the supplier address and warehouse location fields in the structured data. If it's found that supplier A's registered address and main production base are located in "City XX," and supplier A's purchase order PO-123456 is scheduled to ship from that city's port within the next 48 hours, an automatic association index is created between the unstructured typhoon warning and the structured order data. Through this matching of "metadata" and "key business data fields," the previously isolated typhoon warning news is linked to specific business entities (supplier A) and business activities (order PO-123), establishing an automatically identifiable index association that allows for rapid location of all affected business entities. All processed vector data, metadata, and their relationships together constitute a continuously updated, semantically searchable unified supply chain knowledge base, providing a data foundation for subsequent risk perception.

[0035] The entire supply chain knowledge base adopts a strategy of "initial full construction + real-time incremental updates." During the platform initialization phase, historical data is processed in batches to establish the base knowledge base. During operation, real-time data streams are continuously processed to dynamically maintain the timeliness and completeness of the knowledge base. A unified supply chain knowledge base capable of understanding business semantics and supporting complex relational queries has been built, providing reliable data support for subsequent risk perception and intelligent decision-making.

[0036] S12, when external event information is detected, the event entities of the external event information are extracted using a large language model, and the event entities are used for correlation analysis in the supply chain knowledge base to generate structured risk event objects.

[0037] Furthermore, in step S12, when external event information is detected, the event entities of the external event information are extracted using a large language model, and correlation analysis is performed on the event entities in the supply chain knowledge base to generate structured risk event objects, including: S12-1, Using a large language model, perform semantic understanding and entity recognition on the external event information, and extract event entities including event type, location of occurrence, time window of impact, and scope of impact area; S12-2, based on the event entity, search and match supply chain entities in the supply chain knowledge base, identify the supply chain entities that have an impact, and call the business system interface to obtain the current business data corresponding to the supply chain entity, wherein the current business data includes order planning time, transportation route and inventory level; S12-3, Using a large language model, perform risk impact analysis based on the event entity and the current business data, and generate risk impact analysis results; S12-4, Based on the risk impact analysis results, generate the risk event object, including event identifier, risk level, list of entities affecting the supply chain, specific risk causes, and response recommendations.

[0038] Furthermore, in step S12-3, the step of using a large language model to perform risk impact analysis based on the event entity and the current business data, and generating risk impact analysis results, includes: S12-3-1, Analyze the overlap between the impact time window of external events and the order planning time to obtain the time dimension analysis results; S12-3-2, analyze the spatial correlation between the impact area of ​​external events and the geographical location and transportation routes of supply chain entities to obtain spatial dimension analysis results; S12-3-3, based on the preset risk matrix, the results of time dimension analysis and spatial dimension analysis are combined to evaluate and obtain the risk impact analysis results including the risk level.

[0039] In this embodiment, when the perception module detects a new incoming external event, such as "The XX City Meteorological Observatory issued a red typhoon warning at 10:00 AM today, predicting that a super typhoon will make landfall in XX City within the next 48 hours, with central winds reaching level 16 or higher. All ports and wharves in the city have suspended operations since 6:00 PM today, expected to last for 48 hours," the platform immediately activates the large language model in the AI ​​engine module to process the information. Based on preset prompt word templates, the large language model performs semantic understanding and entity recognition on this news, accurately extracting structured event entities. For example, the extraction result might be: Event type: Typhoon warning; Location: XX City; Impact time window: from 6:00 PM on the current date, lasting 48 hours; Affected area: XX City, especially coastal areas.

[0040] Based on the extracted event entities, a correlation analysis is performed in the supply chain knowledge base to initiate a multi-dimensional search. First, semantic search is performed using the key entity "Location of Occurrence: City XX". Then, the relationships in the supply chain knowledge graph are traversed to quickly locate all supply chain entities (such as suppliers, materials, orders, or inventory) located in City XX. For example, the factory address of the core supplier "Supplier A" is identified as being located in City XX. Then, by calling the API of the ERP system, all confirmed but not yet delivered purchase orders (such as order PO-123) supplied by "Supplier A" are obtained, and their planned delivery times (ETDs) are extracted. At the same time, the API of the TMS system is called to query all in-transit transportation orders passing through the port of City XX and their estimated arrival times (ETAs) and detailed routes.

[0041] After obtaining this real-time business data, the large language model begins in-depth risk impact analysis. This includes calculating, in the time dimension, whether the typhoon impact window completely overlaps with the planned order delivery time (e.g., determining whether the planned delivery time of order PO-123 falls within the 48-hour window of typhoon impact); and in the spatial dimension, identifying land transport routes that directly pass through the typhoon's expected landfall area (e.g., analyzing whether the order's planned transport route relies on the port of XX city). Assuming order PO-123 is scheduled to be shipped tomorrow morning and does indeed require transshipment through the port of XX city, both the time and spatial dimensions match successfully. The large language model compares the analysis results (high temporal overlap, high spatial correlation) with a pre-set risk matrix. This risk matrix comprehensively considers the degree of event impact (port closure leading to complete transportation disruption, high impact) and the probability of occurrence (official warning, high probability), thus outputting a high-risk level risk impact analysis result. The risk matrix is ​​a pre-constructed multi-dimensional quantitative assessment model used to determine the values ​​of the risk level field in the risk event object. Finally, the platform automatically generates a structured risk event object, which includes a unique event identifier (e.g., EVT-2024-TYPHOON-001), a risk level (“High”), a list of affected supply chain entities (e.g., suppliers [“Supplier A”], orders [“PO-123”]), and clearly records the specific risk cause (“A red typhoon warning was issued in XX city where Supplier A is located, leading to port closure and directly affecting the shipment and transportation of order PO-123”). This risk event object provides accurate and rich contextual information for subsequent interactive decision support.

[0042] S13: Receive the query statement input by the user, decompose the query statement based on the risk event object using a large language model to obtain multiple sub-tasks, process the retrieval data obtained based on different sub-tasks, and generate a summary answer and multiple alternative solutions.

[0043] Furthermore, in step S13, the query statement is decomposed based on the risk event object using a large language model to obtain multiple sub-tasks, including: S13-1, The semantic intent of the query statement is identified using a large language model, and the query statement is decomposed in combination with the business context provided by the risk event object to generate multiple sub-tasks including counting affected orders, querying alternative resources, and assessing cost impact. S13-2, Based on the semantic intent and business context, the ReAct framework is used to plan the execution sequence for each subtask, and a calling tool is assigned to each subtask in the execution sequence, as well as an execution instruction containing query conditions, filtering parameters and return fields are generated; wherein, the type of calling tool includes SQL query engine, business system API interface or vector retrieval interface.

[0044] Furthermore, in step S13, the retrieval data obtained from searching based on different subtasks is processed to generate a summary-style answer and multiple alternative solutions, including: S13-3, retrieve the data from the business system and supply chain knowledge base according to the execution instruction to obtain the retrieved data; S13-4, Analyze the retrieved data using a large language model to generate the summary-style answer, which includes a description of the current situation and a quantitative analysis of the business impact; S13-4, combining the preset business rule base and historical similar event handling records retrieved from the supply chain knowledge base, generates multiple alternative solutions including resource allocation methods, delivery plan adjustments, and customer communication strategies.

[0045] In this embodiment, based on the previously generated risk event object (EVT-2024-TYPHOON-001) regarding the impact of the typhoon on XX city, the supply chain manager inputs a natural language query through the platform's dialogue interface, such as "Assess the impact of this typhoon on our third-quarter revenue and recommend countermeasures." Upon receiving this query, the AI ​​engine module is immediately activated for processing. The LLM first identifies the semantic intent of the query as "impact assessment" and "solution recommendation," and combines this with the business context provided by the generated risk event object (e.g., affected suppliers are "Supplier A," and high-risk orders are "PO-123"). The LLM then uses the ReAct framework for reasoning and planning, decomposing this complex query into an ordered sequence of subtasks. For example, the execution sequence is: Subtask 1 (Statistics on Affected Orders): Accurately compile statistics on all purchase orders affected by this typhoon, their total amount, and associated customer information. The tool assigned to this task is to generate SQL query commands to retrieve order details from the ERP database's order table for orders with a status of "Confirmed" or "In Transit" and supplier "Supplier A".

[0046] Subtask 2 (Query Alternative Resources): Check for alternative inventory or suppliers. The tool assigned to this task is to call the WMS system's API to query the inventory levels of the same material (e.g., SKU-X) in other regional warehouses; simultaneously, it calls the supplier management system's API to query a list of alternative suppliers with the same material supply qualifications.

[0047] Subtask 3 (Assess Cost Impact): Evaluate the additional costs of different response options (such as emergency allocation, air freight). The tool assigned to this task is to call the cost calculation engine's API and perform simulation calculations based on the logistics quotation database.

[0048] LLM generates executable instructions with specific parameters for each subtask. For example, for subtask one, the generated SQL instruction might include conditions such as WHERE supplier_id = 'Supplier A' AND status IN ('confirmed', 'in_transit'). The platform executes this plan, running the above instructions sequentially, retrieving the required data from business systems such as ERP and WMS, as well as the knowledge base. Examples include: the total amount of affected orders is 5 million yuan (of which 2 million yuan involves orders from important customer C); warehouse B has 1000 units of available inventory; the delivery cycle for alternative supplier D is 7 days; and air freight costs will increase by 3 times the original logistics costs. LLM intelligently integrates and analyzes this multi-source heterogeneous data, generating a summary response: "This typhoon is expected to directly affect 10 orders from supplier A, totaling 5 million yuan, potentially leading to a 5% decrease in Q3 revenue. Among them, the 2 million yuan order from important customer C faces the highest risk of delay." LLM combines a pre-set business rule base (such as "orders from high-value customers should be prioritized for allocation from the backup warehouse") with historical knowledge base records of similar risk events (such as "in the past, offering a 5% discount to customer C could obtain leniency") to generate multiple alternative solutions and simulate their potential effects: Option A (Emergency Transfer): Immediately transfer inventory from warehouse B to prioritize fulfilling customer C's order. The estimated additional logistics costs are 50,000 yuan, but this will guarantee 2 million yuan in revenue, with no impact on customer satisfaction.

[0049] Option B (Negotiate Delay and Compensation): Communicate with Customer C to delay delivery and offer an 8% price discount. There are no additional logistics costs, but revenue will decrease by 160,000 yuan, and there is a risk of customer churn.

[0050] Option C (Activate Alternative Supplier): Immediately place an order with alternative supplier D. Material costs will increase by 2%, and delivery will be delayed by 7 days. The overall impact is pending evaluation.

[0051] These summaries and proposals, which include quantitative analysis, are ultimately presented in a clear and structured format to provide direct and in-depth data support for decision-making.

[0052] S14, the solution selected by the user from multiple alternative solutions is taken as the decision solution, and the large language model is invoked to generate and execute cross-departmental workflow tasks based on the decision solution.

[0053] Furthermore, in step S14, the invocation of the large language model to generate and execute cross-departmental workflow tasks based on the decision scheme includes: S14-1, The decision scheme is analyzed using a large language model, and task elements including task type, execution action and participating departments are identified based on a predefined workflow template; S14-2: Generate structured instructions containing task allocation schemes and execution priorities based on task elements, and call the corresponding business system API interface to create and distribute workflow tasks to the corresponding participating departments based on the structured instructions.

[0054] In this embodiment, when the supply chain manager selects "Solution A: Immediately transfer inventory from warehouse B to prioritize fulfilling customer C's order" as the final decision solution from among several alternative solutions recommended by the platform, the platform immediately activates the collaborative workflow automation function. The large language model first parses the selected text decision solution, identifying the key action instructions: "transfer inventory from warehouse B" and "fulfill customer C's order"; this decision solution is then matched against the system's predefined "solution-workflow template" mapping library, successfully matching the "emergency inventory transfer" template. This template predefines the task elements required to execute this solution: the task types are "inventory transfer execution" and "customer communication," the execution actions are "create transfer order" and "draft customer notification," the participating departments are clearly defined as "logistics department" and "sales department," and the task priority is set to "high."

[0055] Based on the identified task elements, the LLM leverages its function calling capabilities to generate a set of structured, executable instructions. These instructions are essentially API call requests encapsulated with specific parameters. For example, the LLM might generate a JSON-formatted instruction requesting a call to the Jira system's API to create a task titled "Execute Emergency Transfer for Order PO-123," specifying the responsible party as the logistics department, prioritizing it as high, and automatically filling in key information extracted from the risk event object and decision-making scheme in the task description (e.g., the source warehouse is warehouse B, and the target customer is customer C). Upon receiving this series of structured instructions generated by the LLM, the execution engine calls the corresponding business system API. For instance, by calling the WMS (Warehouse Management System) API, it automatically generates a transfer order from warehouse B to the required location of customer C; simultaneously, by calling the APIs of WeChat Work or DingTalk, it instantly pushes the task notification to the workbenches of the logistics and sales department managers.

[0056] Furthermore, as part of the workflow, LLM automatically drafts necessary communication text based on task content, preset email templates, and real-time information returned by the API (such as the expected new shipping date). For example, it generates a draft reassurance email for the sales department to send to customer C, explaining the shipping delay caused by force majeure and the emergency remedial measures taken, along with an updated estimated delivery time. All created tasks and drafted texts are executed only after final confirmation by relevant personnel, forming a complete closed loop from intelligent decision-making to automated collaborative execution, greatly improving the efficiency and accuracy of emergency response.

[0057] In another embodiment, the method further includes: The complete processing of each risk event is recorded as historical processing data, including input context data, generated alternative solutions, user-selected decision-making options, and final business results. Based on the historical processing data, the large language model is updated and optimized using parameter-efficient fine-tuning techniques. The update and optimization process includes using LoRA technology to build an adapter module for supply chain decision-making scenarios, training only a small number of parameters while maintaining the model's generality. A reward mechanism is built based on successful cases of historical decisions, and reinforcement learning is used to optimize decision-making strategies from human feedback.

[0058] In another embodiment, the method further includes: Real-time analysis of surveillance video streams in supply chain scenarios, including warehouses and ports, using computer vision models to identify anomalies in cargo quantity or packaging damage, determining abnormal states, generating early warning events based on these states, and triggering risk management processes; and, Optical character recognition (OCR) technology is used to analyze logistics document images and extract key business fields, including order number, material code, and quantity information. These key business fields are then automatically compared with business records in the business system, and a data correction process is triggered when discrepancies are found.

[0059] In another embodiment, the method further includes: Analyze the supplier's historical transaction data to obtain basic performance indicators, including on-time delivery rate, quality pass rate, and price stability; Based on the supplier communication records, assess the corresponding response time and cooperation level to generate service capability assessment results; Based on the aforementioned basic performance indicators and service capability assessment results, a multi-dimensional evaluation indicator system for suppliers is constructed. Based on the multi-dimensional evaluation indicator system, a large language model is used for analysis and trend prediction to generate a supplier profile that includes an overall risk score and development potential assessment.

[0060] Reference Figure 2The diagram shown is a structural schematic of a logistics operation management and decision-making device based on a large model, provided in an embodiment of the present invention.

[0061] In this embodiment, the device 20 includes: The data processing unit 21 is used to collect structured and unstructured data in logistics operations in real time, convert unstructured data into text vectors using a preset embedding model, and establish index associations with key fields of structured data to build a unified supply chain knowledge base. The monitoring unit 22 is used to extract event entities from external event information using a large language model when external event information is detected, and to perform correlation analysis on the event entities in the supply chain knowledge base to generate structured risk event objects. The solution generation unit 23 is used to receive the query statement input by the user, decompose the query statement based on the risk event object using a large language model to obtain multiple sub-tasks, process the retrieval data obtained based on different sub-tasks, and generate a summary answer and multiple alternative solutions. Decision unit 24 is used to take the solution selected by the user from multiple alternative solutions as the decision solution, call the large language model to generate and execute cross-departmental workflow tasks according to the decision solution.

[0062] Each unit module of the device 20 can execute the corresponding steps in the above method embodiment, so the details of each unit module will not be elaborated here. Please refer to the description of the corresponding steps above for details.

[0063] This invention also provides a logistics operation management and decision-making device based on a large model. This device includes the logistics operation management and decision-making apparatus based on a large model as described above. The large model-based logistics operation management and decision-making apparatus can employ... Figure 2 The structure of the embodiment, correspondingly, can be executed Figure 1 The technical solutions of the method embodiments shown are similar in implementation principle and technical effect. For details, please refer to the relevant records in the above embodiments, which will not be repeated here.

[0064] The device includes: a mobile phone, digital camera, or tablet computer, or other device with a camera function; or a device with an image processing function; or a device with an image display function. The device may include components such as a memory, processor, input unit, display unit, and power supply.

[0065] The memory can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc.; the data storage area can store data created according to the use of the device, etc. In addition, the memory can include high-speed random access memory, and can also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory can also include a memory controller to provide access to the memory for the processor and input units.

[0066] The input unit can be used to receive input numerical, character, or image information, and to generate keyboard, mouse, joystick, optical, or trackball signal inputs related to user settings and function control. Specifically, in addition to a camera, the input unit of this embodiment may also include a touch-sensitive surface (e.g., a touch screen) and other input devices.

[0067] The display unit can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces of the device. These graphical user interfaces can be composed of graphics, text, icons, video, and any combination thereof. The display unit may include a display panel, optionally configured as an LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), or other similar display panel. Furthermore, a touch-sensitive surface may cover the display panel. When the touch-sensitive surface detects a touch operation on or near it, it transmits the information to the processor to determine the type of touch event. Subsequently, the processor provides corresponding visual output on the display panel based on the type of touch event.

[0068] This invention also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the memory described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement... Figure 1 The method for logistics operations management and decision-making based on a large model is shown. The computer-readable storage medium may be a read-only memory, a disk, or an optical disk, etc.

[0069] This invention also provides a computer program product, including a computer program / instructions, which are loaded and executed by a processor to implement... Figure 1 This paper presents a logistics operations management and decision-making method based on a large model.

[0070] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the device embodiments, equipment embodiments, and storage medium embodiments, since they are basically similar to the method embodiments, the descriptions are relatively simple, and relevant parts can be referred to the descriptions in the method embodiments.

[0071] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0072] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A large model-based logistics operation management and decision method, characterized in that, The method comprises: Real-time collection of structured data and unstructured data in logistics operation, conversion of unstructured data into text vectors by using a preset embedding model, and index association with key fields of structured data to build a unified supply chain knowledge base; When external event information is monitored, event entities of the external event information are extracted by using a large language model, and associated analysis is performed in the supply chain knowledge base based on the event entities to generate a structured risk event object; A query sentence input by a user is received, the query sentence is decomposed based on the risk event object by using the large language model to obtain a plurality of subtasks, and the search data obtained based on different subtasks is processed to generate an abstract answer and a plurality of alternative solutions; A solution selected by the user from the plurality of alternative solutions is taken as a decision solution, and a cross-department workflow task is generated and executed according to the decision solution by calling the large language model.

2. The logistics operation management and decision method based on a large model according to claim 1, characterized in that, The real-time collection of structured data and unstructured data in logistics operation comprises: Data collection is performed by connecting business systems including ERP systems, WMS systems, TMS systems, and CRM systems through application program interfaces to obtain structured data; Unstructured data including supplier emails, news reports, weather services, social media, and customs announcements is collected from external information sources through web crawlers or third-party data interfaces.

3. The logistics operation management and decision method based on a large model according to claim 1, characterized in that, The conversion of unstructured data into text vectors by using a preset embedding model and the establishment of index association with key fields of structured data comprise: Text corresponding to unstructured data is vectorized by using a preset embedding model to obtain text vectors; Key entities in unstructured data are extracted as metadata, and the metadata are stored in the supply chain knowledge base after being bound and associated with the text vectors, wherein the key entities include location information, event types, organization names, and / or material identifiers; Index association between unstructured data and structured data is established by matching key entities in the metadata with corresponding key fields in the structured data.

4. The logistics operation management and decision method based on a large model according to claim 1, characterized in that, When external event information is monitored, event entities of the external event information are extracted by using a large language model, and associated analysis is performed in the supply chain knowledge base based on the event entities to generate a structured risk event object, which comprises: Semantic understanding and entity recognition of the external event information are performed by using a large language model to extract event entities including event types, occurrence locations, influence time windows, and influence area ranges; Supply chain entity retrieval matching is performed in the supply chain knowledge base based on the event entities to identify affected supply chain entities, and business system interfaces are called to obtain current business data corresponding to the supply chain entities, wherein the current business data includes order planning time, transportation routes, and inventory levels; Risk impact analysis is performed based on the event entities and the current business data by using a large language model to generate risk impact analysis results; The risk event object is generated according to the risk impact analysis result, and includes event identification, risk level, a list of supply chain entities affected, specific risk causes and countermeasures.

5. The logistics operation management and decision method based on a large model according to claim 4, characterized in that, The risk impact analysis result is generated by using a large language model based on the event entity and the current business data, including: Analyze the overlap between the influence time window of the external event and the order planning time to obtain the time dimension analysis result; Analyze the spatial correlation between the influence area range of the external event and the geographical position of the supply chain entity and the transportation route to obtain the spatial dimension analysis result; Based on the preset risk matrix, the time dimension analysis result and the spatial dimension analysis result are evaluated to obtain the risk impact analysis result including the risk level.

6. The logistics operation management and decision-making method based on a large model according to claim 1, characterized in that, The query statement is decomposed based on the risk event object by using a large language model to obtain a plurality of subtasks, including: The semantic intent of the query statement is identified by using a large language model, and the query statement is decomposed based on the business context provided by the risk event object to generate a plurality of subtasks including statistical affected orders, query alternative resources and cost impact evaluation; Based on the semantic intent and business context, the ReAct framework is used to plan the execution sequence for each subtask, and each subtask in the execution sequence is assigned a calling tool and an execution instruction containing query conditions, filtering parameters and return fields; wherein the type of calling tool includes SQL query engine, business system API interface or vector retrieval interface.

7. The logistics operation management and decision-making method based on a large model according to claim 6, characterized in that, The retrieval data obtained based on different subtasks is processed to generate a summary answer and a plurality of alternative solutions, including: Data retrieval is performed from the business system and the supply chain knowledge base according to the execution instruction to obtain the retrieval data; The retrieval data is analyzed by using a large language model to generate the summary answer including the current situation description and quantitative business impact analysis; Combined with the preset business rule library and the historical similar event processing records retrieved from the supply chain knowledge base, a plurality of alternative solutions including resource allocation method, delivery scheme adjustment and customer communication strategy are generated.

8. The logistics operation management and decision-making method based on a large model according to claim 1, characterized in that, The large language model is called to generate and execute cross-department workflow tasks according to the decision scheme, including: The decision scheme is parsed by using a large language model, and task elements including task type, execution action and participating department are identified based on a predefined workflow template; According to the task elements, structured instructions containing task allocation scheme and execution priority are generated, and the API interface of the corresponding business system is called to create and distribute workflow tasks to the corresponding participating departments according to the structured instructions.

9. A large model-based logistics operation management and decision device, characterized by, The device includes: The data processing unit is used to collect structured data and unstructured data in real time, convert unstructured data into text vectors by using a preset embedding model, and index and associate with the key fields of structured data to build a unified supply chain knowledge base; The monitoring unit is configured to extract event entities of external event information by using a large language model when the external event information is monitored, and perform correlation analysis in the supply chain knowledge base based on the event entities to generate a structured risk event object; The scheme generation unit is configured to receive a query sentence input by a user, decompose the query sentence based on the risk event object by using a large language model to obtain a plurality of subtasks, and process retrieval data obtained based on different subtasks to generate an abstract answer and a plurality of alternative solutions; The decision unit is configured to take a scheme selected by the user from the plurality of alternative solutions as a decision scheme, and call a large language model to generate and execute cross-department workflow tasks according to the decision scheme.

10. A logistics operation management and decision device based on a large model, comprising a processor, a memory, and a computer program stored in the memory, wherein the computer program is executed by the processor to implement the steps of the logistics operation management and decision method based on a large model according to any one of claims 1 to 8.

Images