A simulation-based supply chain analysis system using large language models

Abstract

The present invention relates to a simulation-based supply chain analysis system using a large language model, which simulates the state of a supply chain using a simulator and stores the results, searches for simulation results and related knowledge in response to a user's supply chain query, and responds using an LLM model together with the searched data. The system comprises: a simulator that simulates a supply chain model and outputs simulation results; an LLM module equipped with a large language model (LLM); a supply chain database that stores data; a simulation processing unit that performs a simulation through the simulator according to a given supply chain model, collects simulation results, and stores them in the supply chain database; a knowledge storage unit that stores modeling information of the supply chain model, knowledge about the supply chain model, analysis knowledge for analyzing the simulation results, and the collected simulation results in the supply chain database; a query processing unit that receives a first query message regarding the supply chain; and a knowledge search unit that extracts search information from the first query message and searches the supply chain database using the search information. A configuration is provided that includes a response processing unit that generates a second query message containing the first query message along with the retrieved data, queries the LLM module with the second query message to obtain a response message, and outputs the obtained response message. According to the system described above, by additionally searching for and providing relevant knowledge and simulation results for supply chain analysis, hallucinations can be minimized, thereby enabling the provision of highly reliable responses to user queries.

Description

Technology Field

[0001] The present invention relates to a simulation-based supply chain analysis system using a large language model, which simulates the state of a supply chain using a simulator and stores the results, searches for simulation results and related knowledge in response to a user's supply chain query, and responds using an LLM model together with the searched data. Background Technology

[0002] Generally, production management refers to the activity of properly allocating given ordered products to production facilities or factories to ensure they are produced without disruption within a given timeframe.

[0003] In particular, many companies possess multiple production facilities or plants, and production capacity varies among them. That is, factors such as the number of equipment units, the performance of the equipment itself, the skill level and employment standards of the workforce, and the availability of raw materials may differ for each plant or production process. Consequently, even when producing the same product, differences in production capacity arise.

[0004] Furthermore, when multiple types of products must be produced in specific demand or order quantities over a given period, those quantities must be allocated and distributed among a number of production plants. Allocating production resources to match such given demand is a critical issue for improving production efficiency. Traditionally, resource allocation problems were mostly addressed using a rule-based approach based on specific conditions; however, the lack of clear evaluation metrics made it difficult to assess the performance of the resulting allocation outcomes.

[0005] To solve these problems, the applicant presents a technology for scheduling processes using simulation [Patent Documents 1, 2, 3] or a technology for handling the allocation problem of production plants using simulation [Patent Document 4]. Therefore, through simulation, an optimal scenario can be predicted for a given demand or order quantity.

[0006] However, the task of interpreting and analyzing the results of simulation models relies primarily on individuals with expertise in this field. Simulation experts or data analysts evaluate the status of a factory or process based on the simulation results and propose necessary adjustments. This process is not only time-consuming but also has limitations, as the accuracy and reliability of the results vary depending on the analyst's experience and capabilities.

[0007] Therefore, even if a general manager does not possess specialized knowledge of simulation analysis, there is a need for technology that supports decision-making regarding the production plant by interpreting or analyzing simulation results. Prior art literature

[0009] Korean Registered Patent Publication No. 10-2338304 (Published Dec. 13, 2021) Korean Published Patent Publication No. 10-2023-0165026 (Published Dec. 05, 2023) Korean Published Patent Publication No. 10-2024-0000926 (Published Jan. 03, 2024) Korean Patent Application No. 10-2024-0133037 (Filed Sep. 30, 2024) The problem to be solved

[0010] The objective of the present invention is to solve the problems described above by providing a simulation-based supply chain analysis system using a large language model, which simulates the supply chain state using a simulator and stores the results, searches for simulation results and related knowledge in response to a user's supply chain query, and responds using an LLM model along with the retrieved data.

[0011] In particular, the objective of the present invention is to provide a simulation-based supply chain analysis system using a large language model that searches for a simulation result most similar to a currently given supply chain state and provides the searched simulation result together with an LLM model to generate a response from the result.

[0012] In addition, the objective of the present invention is to provide a simulation-based supply chain analysis system using a large language model that additionally searches for background knowledge for supply chain analysis and related knowledge regarding methods for interpreting and analyzing simulation results, and provides them together with the LLM model. means of solving the problem

[0013] To achieve the above objective, the present invention relates to a simulation-based supply chain analysis system using a large language model, comprising: a simulator that simulates a supply chain model and outputs a simulation result; an LLM module equipped with a large language model (LLM); a supply chain database that stores data; a simulation processing unit that performs a simulation through the simulator according to a given supply chain model, collects the simulation result, and stores it in the supply chain database; a knowledge storage unit that stores modeling information of the supply chain model, knowledge about the supply chain model, analysis knowledge for analyzing the simulation result, and the collected simulation result in the supply chain database; a query processing unit that receives a first query message regarding the supply chain; a knowledge search unit that extracts search information from the first query message and searches the supply chain database using the search information; and a response processing unit that generates a second query message including the first query message along with the searched data, queries the LLM module with the second query message to obtain a response message, and outputs the obtained response message.

[0014] In addition, the present invention is characterized in that, in a simulation-based supply chain analysis system using a large language model, the modeling information of the supply chain model includes processes, facilities, working hours, products, product connection structures, demand, and delivery deadlines.

[0015] Furthermore, the present invention relates to a simulation-based supply chain analysis system using a large language model, wherein the simulation result is composed of an original result generated by the simulation and summary information summarizing the original result, the original result includes items regarding delivery compliance prediction, product flow record, production plan optimization, equipment operation record, and inventory level management, and the summary information is composed of performance indicators (KPIs), including delivery compliance rate and equipment utilization rate, and the simulation result is characterized by being indexed using modeling identification information identifying the supply chain model and execution identification information identifying the runtime of the simulation executed with the modeling of the supply chain model.

[0016] In addition, the present invention relates to a simulation-based supply chain analysis system using a large language model, wherein the analysis knowledge is knowledge for analyzing the simulation results and is characterized by including guidelines and rules for performing the analysis method.

[0017] In addition, the present invention is characterized in that, in a simulation-based supply chain analysis system using a large language model, the supply chain database provides a function to search for the simulation results using an SQL query, and provides a function to search for modeling information of the supply chain model, knowledge about the supply chain model, and analysis knowledge for analyzing the simulation results using an embedding vector.

[0018] In addition, the present invention relates to a simulation-based supply chain analysis system using a large language model, wherein the query processing unit is characterized by reconstructing a query message by including information from a previous conversation.

[0019] Furthermore, the present invention relates to a simulation-based supply chain analysis system using a large language model, wherein the knowledge search unit extracts search information including a search term from the query message, extracts a data table name or data item of a simulation result mapped to the search term, generates an SQL query statement using the extracted table name or data item, searches for the simulation result using the generated SQL query statement, generates an embedding vector of the search information, and searches for modeling information of the supply chain model, knowledge about the supply chain model, and analysis knowledge for analyzing the simulation result using the generated embedding vector.

[0020] In addition, the present invention relates to a simulation-based supply chain analysis system using a large language model, wherein the second query message includes a retrieved simulation result, retrieved knowledge, and the first query message, and is characterized by being written to request a response to the first query message by referring to the retrieved simulation result and the retrieved knowledge. Effects of the invention

[0021] As described above, according to the simulation-based supply chain analysis system using a large language model according to the present invention, by performing the interpretation and analysis of simulation results regarding the supply chain through an LLM model, it is possible to support optimal decision-making regarding the supply chain even without specialized knowledge of simulation analysis, thereby achieving the effect of managing the supply chain more efficiently and quickly.

[0022] In addition, according to the simulation-based supply chain analysis system using a large language model according to the present invention, by analyzing through an LLM model by referring to the optimal scenario results obtained by the simulator, the production efficiency of the production plant can be increased, potential problems can be identified and resolved in advance, and the overall plant can be optimized.

[0023] In addition, the present invention minimizes hallucination by additionally searching for relevant knowledge regarding supply chain analysis and providing it together with the LLM model, thereby enabling the provision of highly reliable responses to user queries.

[0024] Furthermore, by additionally searching for and providing relevant knowledge regarding background knowledge for supply chain analysis, interpretation of simulation results, and analysis methods, the present invention enables the LLM model to generate complex responses, such as bottleneck analysis, based on the searched data, and provides insights for prediction and analysis, thereby enabling data processing and analysis at a level that was previously impossible.

[0025] In addition, the present invention derives a summary of simulation results and key performance indicators (KPIs) through a simulator, and generates complex responses such as bottleneck analysis based on search data through an LLM model, thereby enabling users to understand the supply chain status in real time and make optimal decisions. Brief explanation of the drawing

[0026] FIG. 1 is a configuration diagram of the entire system for implementing the present invention. FIG. 2 is an exemplary diagram of a supply chain model according to an embodiment of the present invention. FIG. 3 is an example diagram of parameters and values ​​of a supply chain model according to an embodiment of the present invention. FIG. 4 is a table showing simulation result items and descriptions thereof according to an embodiment of the present invention. FIG. 5 is a table illustrating simulation result items and their values ​​according to an embodiment of the present invention. FIG. 6 is a block diagram of the configuration of a simulation-based supply chain analysis system using a large language model according to an embodiment of the present invention. FIG. 7 is an exemplary diagram of a supply chain query and response screen according to an embodiment of the present invention. Specific details for implementing the invention

[0027] Hereinafter, specific details for implementing the present invention will be explained with reference to the drawings.

[0028] In addition, in describing the present invention, identical parts are denoted by the same reference numerals, and their repeated description is omitted.

[0029] First, the configuration of the entire system for implementing the present invention will be explained with reference to FIG. 1.

[0030] As shown in FIG. 1(a), the simulation-based supply chain analysis system using a large language model according to the present invention (hereinafter the analysis system) can be implemented as a program system on a computer terminal (10) that analyzes or predicts the supply chain.

[0031] That is, the analysis system (30) can be implemented as a program system on a computer terminal (10), such as a PC, smartphone, or tablet PC. In particular, the analysis system can be configured as a program system and installed and executed on a computer terminal (10). The analysis system (30) provides a service that analyzes the supply chain or provides a forecast by utilizing the hardware or software resources of the computer terminal (10).

[0032] At this time, the computer terminal (10) is equipped with a processor, memory, etc. The processor performs simulation, analysis, prediction, etc., and stores intermediate result data, final result data, etc. in memory or retrieves them from memory. At this time, the processor may be composed of a central processing unit (CPU), a graphics processing unit (GPU), an AI accelerator, a neural processing unit (Neural Processing Unit), an AI chip, etc.

[0033] In addition, as another embodiment, as shown in FIG. 1(b), the analysis system (30) may be configured and executed as a server-client system consisting of an analysis client (30a) and an analysis server (30b) on a computer terminal (10).

[0034] Meanwhile, the analysis client (30a) and the analysis server (30b) can be implemented according to the configuration method of a conventional client and server. That is, the functions of the entire system can be distributed according to the performance of the client or the amount of communication with the server. Although described below as an analysis system, it can be implemented in various forms of distribution depending on the configuration method of the server and client.

[0036] Next, a supply chain model according to an embodiment of the present invention will be described with reference to FIGS. 2 and 3.

[0037] A supply chain model consists of demand, which represents the products ordered and requested and the quantity required, and facilities and processes that produce the products (hereinafter referred to as the production network).

[0038] As shown in Figure 2, the production network of the supply chain model consists of multiple processes (OPER), and each process (OPER) is equipped with a machine (MAC) and processes (processes, produces, etc.) products (or raw materials, RM) using the machine (MAC). That is, each process (OPER) receives products (or raw materials, RM) and produces processed (processes, produces, etc.) products (RM).

[0039] The production network of the supply chain model is composed of multiple processes to ultimately produce a final product. That is, the production network is a process of producing a final product in a single factory, representing the process of inputting raw materials (RM) to ultimately produce a final product (FINAL_PROD). In the example of Figure 2, starting from the input of raw materials, the final product is completed through process steps such as OPER_1, OPER_2, OPER_3, and OPER_4.

[0040] In addition, products (or raw materials) produced in each process may be fed back into other processes. Here, the product ultimately produced by the final process is referred to as the final product, and the product produced in an intermediate process and fed into the next process is referred to as the intermediate product.

[0041] In addition, products produced in one process can be stored in a warehouse before being fed into another process. That is, products produced in one process are stored in a warehouse, and the next process retrieves the products from the warehouse for processing. A warehouse can be added between processes.

[0042] Furthermore, preferably, the production network of the supply chain model can be defined by each process, facility, product, working time, product linkage structure, etc. The definitions of processes, facilities, and products are as previously described. Additionally, working time represents the time required to perform the corresponding process at each stage. Moreover, the product linkage structure represents information regarding the transformation from a specific product to another. That is, each product is produced into another product (processed product) by a specific process, and the product linkage structure represents the structure of each process at this stage. The product linkage structure reflects the flow of the process.

[0043] Next, demand in a supply chain model is defined by the products requested (ordered) for production by the relevant production network, their demanded quantity (order quantity), and delivery deadlines. The product demanded quantity is the quantity for which the production of the final product is requested, and the delivery deadline represents the deadline for which the delivery of the product is requested. In this context, the product refers to the final product.

[0044] As described above, the supply chain model (modeling information) consists of processes, facilities, working hours, products, product connection structures, demand, delivery time information, etc., which represent the state of the supply chain exactly.

[0045] Meanwhile, when each modeling variable in a supply chain model is fixed to a specific numerical value, the model represents a single supply chain state. In other words, a supply chain model representing a single supply chain state is one in which each modeling variable (or modeling data) is specifically configured.

[0046] Preferably, as shown in the example of FIG. 3, each modeling data (variables, parameters, values, etc.) of the supply chain model required for simulation can be represented as matrix data, etc.

[0047] In FIG. 3, (a) is operation information, consisting of information on the time required for each process, etc. Also, (b) is machine information, consisting of production time, production unit information, etc. Also, (c) is product information, representing product information managed (produced) in the factory.

[0048] Additionally, (d) represents a product structure that illustrates the process in which an input product is converted into an output product through a specific process. In this structure, when an input product passes through a specific process, it ultimately becomes an output product and moves to the next stage. In particular, the product structure information consists of input quantity, production quantity information, etc. These products include raw materials, intermediate products, and final products.

[0049] Additionally, (e) is demand information consisting of the delivery date (delivery deadline), delivery quantity, and customer. Processes necessary to meet the specified delivery deadline must be performed, through which the simulator can derive the optimal scheduling result for meeting the delivery deadline.

[0050] Next, the results of a supply chain simulation according to an embodiment of the present invention will be described with reference to FIGS. 4 and 5.

[0051] The results of the supply chain simulation are data generated by simulating a specific supply chain model by the simulator (31). In particular, the result data generated by the simulator (31) consists of original result data generated by the simulation, such as equipment operation records, product flow, and inventory levels, and summary data (or performance indicator data) that summarizes the original results to provide performance indicators (KPIs), etc.

[0052] Figure 4 is a description of the simulation result (original result) items, and Figure 5 shows an example of the simulation result items.

[0053] As seen in Figure 4 or Figure 5, the simulation results consist of items such as delivery time compliance prediction, product flow records, production plan optimization, equipment operation records, and inventory level management.

[0054] First, the delivery compliance forecast item predicts whether products can be shipped on time according to the customer's desired date. This item consists of product delivery compliance and quantity. In other words, the delivery compliance forecast result verifies whether products were shipped in a timely manner to meet demand and represents the factory's performance as a ratio.

[0055] In the example of Fig. 5, it can be seen that 48 units of product "FINAL_PROD" are to be produced for order "DEMAND_1", and exactly 48 units are produced and delivered by the target time (8:00 AM on May 18, 2024). It can be seen that production finished earlier than planned (10:48 PM on May 12, 2024), and all quantities met the deadline without any issues.

[0056] Furthermore, the product flow record items consist of records documenting the process of a product passing through each stage. These items comprise records of the product passing through each process and equipment operation information. In other words, the product flow record results show how products move depending on the specific processes and equipment used for production, allowing for the prediction of bottlenecks. Additionally, by utilizing these results to understand the production flow, the entire process can be managed to ensure smooth operation.

[0057] Through the example in Fig. 5, it can be seen that 12 products of a bundle named "FINAL_01" ("FINAL_PROD") begin to be processed by operator "OPER_4" and equipment "MAC_4". This event is recorded at midnight on May 8, 2024, with the "TRACK_IN" status (start of processing). These products are products associated with an order named "DEMAND_1".

[0058] Furthermore, the production planning optimization feature supports supply chain optimization through the results of production target calculations. This feature consists of production plans and progress status for each process stage. Here, resources refer to the personnel, equipment, and time required for the process, and unnecessary waste can be reduced through resource allocation planning. Additionally, the accuracy of the forecast can be verified by comparing the scheduling results with the actual factory process and equipment operation plans. This enables review and adjustment before implementation.

[0059] Additionally, in the example of Fig. 5, the demand for "DEMAND_1" requires the completion of 156 units of the "FINAL_PROD" product by 8:00 AM on May 18, 2024. The current working quantity is 156 units, and the minimum input start time is the same as the deadline. There are no quantities scheduled for transfer yet, and the product is being processed at the "PROD_1 WH" warehouse. The work is in progress in the "OPER_1" process.

[0060] In addition, the equipment operation record items are items that record how well the factory equipment is operating. These items consist of the equipment's operating status, idle time, and work changeover records. They help reduce downtime and manage the equipment to ensure it operates well. Also, in the example of Fig. 5, it can be seen that equipment "MAC_01" started the "CUTTING" process through the "TRACK_IN" event at midnight on May 8, 2024, and the processed product is "PROD_1".

[0061] In addition, the items for inventory level management are items that manage whether sufficient products required for the process are prepared. These items consist of inventory quantities and flow by process. By maintaining smooth process flow and product supply, productivity can be increased and costs reduced. Also, in the example of Fig. 5, as of midnight on May 9, 2024, it can be seen that the "PROD_1" product is stored in 10 lots in the "PROD_1 WH" warehouse, with a total quantity of 120 units.

[0062] Meanwhile, the simulation results are outcomes simulated by a single scenario (or an optimized scenario) under specific supply chain conditions. In other words, simulation is a process performed to derive an optimized scenario by reflecting the factory conditions at a specific point in time. Therefore, through simulation, it is possible to predict and prepare for potential problems that may occur at each process stage in advance.

[0063] In addition, summary data (or performance indicator data) is data that summarizes the original results generated as a result of the simulation. Summary data consists of key performance indicators (KPIs), such as on-time delivery rates and equipment utilization rates. Performance indicators (KPIs) are utilized as evaluation metrics in the process of deriving optimized results from the simulation. KPI analysis derives key performance indicators to quantitatively evaluate the corresponding simulation (or scenario).

[0064] In particular, performance indicators (KPIs) consist of equipment utilization rates, on-time delivery rates, etc.

[0065] Here, the equipment utilization rate is a ratio indicating how well the machines in a factory are running. For example, if a machine can be operated for 24 hours a day in a factory, but actually spends 20 hours making a product, the utilization rate of this machine is approximately 83% (20 hours ÷ 24 hours = 0.83, i.e., 83%).

[0066] In addition, the on-time delivery rate is the percentage of products that can be supplied on time according to the date ordered by the customer. If the customer wants a quantity of 100 units, and the factory can prepare only 80 units on time, the on-time delivery rate is 80%. If the factory prepares all orders accurately, the on-time delivery rate becomes 100%.

[0067] Meanwhile, the simulation results (or summary data) are stored in the form of tables or records.

[0068] Next, a simulation-based supply chain analysis system (30) using a large language model according to an embodiment of the present invention will be described with reference to FIG. 6.

[0069] As shown in FIG. 6, a simulation-based supply chain analysis system (30) using a large language model according to an embodiment of the present invention comprises a simulator (31) for simulating a supply chain model, an LLM module (32) equipped with a large language model (LLM), a supply chain database (33) for storing the supply chain model and simulation results, a simulation processing unit (34) for performing a simulation and collecting results, a knowledge building unit (35) for building supply chain simulation and analysis knowledge, a query processing unit (35) for receiving a supply chain query and analyzing the query, a knowledge search unit (36) for searching knowledge related to the query, and a response processing unit (37) for generating a response to the supply chain query.

[0070] First, the simulator (31) is a conventional simulator that simulates a supply chain model. For example, the supply chain model uses a model such as the one shown in FIGS. 2 and FIGS. 3 above. For example, the supply chain model of the simulation consists of a given demand and a number of processes that produce products according to that demand. Each supply chain model of the simulation is modeled with the actual demand and production facility status at the site.

[0071] That is, as shown in FIG. 2 or FIG. 3, the modeling data of the supply chain model consists of processes, equipment, working hours, products, product connection structures, demand (demand products and their quantities), delivery deadlines, etc., which represent a specific supply chain state as is. In particular, the supply chain model is modeled by assigning specific numerical values, etc., to each modeling variable.

[0072] In particular, when a supply chain model (or supply chain state) is given, the simulator (31) simulates an optimal scenario in the given supply chain state. Preferably, the simulator (31) is composed of a model in which internal variables are learned and set by a learning model, and is optimized by learning [Patent Documents 1, 2, 3]. Therefore, when the simulator (31) simulates the given supply chain model (or supply chain state), it simulates an optimized supply chain scenario.

[0073] Additionally, the simulator (31) simulates a supply chain model to generate simulation results. In particular, as shown in FIG. 4 or FIG. 5, the simulator (31) generates (outputs) items related to delivery compliance prediction, product flow records, production plan optimization, equipment operation records, inventory level management, etc. The simulation results consist of results simulated by an optimized scenario in a given supply chain model. The result data generated in this way is the original result of the simulation.

[0074] In addition, the simulator (31) generates summary data by summarizing the original simulation results. As previously seen, the summary data consists of key performance indicators (KPIs) such as on-time delivery rates and equipment utilization rates.

[0075] Next, the LLM module (32) is equipped with a large language model (LLM). The LLM model may be an artificial intelligence model designed to enable natural language processing. The LLM model can generate an optimal response (or answer) to the question by identifying the context of the question associated with the input prompt and analyzing the relationships between words. The LLM model can be pre-trained based on large-scale data and can be further trained and fine-tuned to suit a specific task or domain.

[0076] In particular, the LLM module (32) receives a prompt, transmits the prompt to an LLM model, and receives a response from the LLM model and replies. That is, the LLM module (32) generates a response by predicting the next word or sentence based on learned knowledge regarding the prompt entered by the user. This allows for a natural and consistent conversation to continue.

[0077] Preferably, the LLM module (32) is composed of a conventional LLM model. In particular, the LLM module (32) utilizes a pre-trained large language model such as GPT-4. Preferably, the LLM model is based on a conventional Transformer-style neural network architecture. The Transformer-style neural network architecture understands context by considering the interactions between all words in an input sentence through a self-attention mechanism.

[0078] In particular, LLM models learn from vast amounts of text data in various forms, such as web pages, books, and news articles. In doing so, LLM models primarily use unsupervised learning methods. That is, LLM models learn patterns from data without separate labels, allowing the model to understand the structure and meaning of language on its own.

[0079] Furthermore, LLM models divide text into small units (tokens) and convert them into numeric vectors (embeddings), enabling computer processors to process the text. Additionally, LLM models learn by predicting the next word or sentence of a given text, thereby making natural text generation possible.

[0080] In addition, LLM models enable rapid transfer learning (fine-tuning) for new tasks with a small amount of additional data through large-scale pre-training.

[0081] Next, the supply chain database (33) is a database that stores the modeling content, simulation results, and related knowledge of the supply chain model.

[0082] Specifically, the supply chain database (33) stores modeling information of the supply chain model and related knowledge (or background knowledge) about the modeling.

[0083] The modeling information of the supply chain model consists of the modeling data of the supply chain model described earlier.

[0084] Furthermore, knowledge related to the supply chain model consists of background knowledge designed to aid in understanding the model and modeling condition knowledge that defines the dataset and its conditions. The background knowledge comprises explanations of the dataset and its schema structure, simulation-related terminology, and key performance indicators (KPIs). Additionally, the modeling condition knowledge defines the dataset format and its constraints, which are essential for running the supply chain simulator. This ensures data accuracy and consistency, thereby preventing simulation errors.

[0085] The relevant knowledge (background knowledge) of the supply chain model is organized into knowledge items in a hierarchical structure, designed to enable a clear understanding of the structure and relationships of the supply chain simulation through each item. Additionally, items such as conditional knowledge are provided in a table format, allowing for an intuitive understanding of the specifications and meanings of each table and column.

[0086] Preferably, background knowledge of the supply chain model is written in a document format including text, tables, etc.

[0087] In addition, the supply chain database (33) stores simulation results, and result search knowledge, analysis knowledge, etc. related to the simulation results.

[0088] As previously explained, the simulation results consist of the original results generated by the simulation and summary information that summarizes the original results. The original results of the simulation consist of items such as delivery on-time prediction, product flow records, production plan optimization, equipment operation records, and inventory level management.

[0089] The simulation summary information is a summary of the original simulation results and consists of key performance indicators (KPIs). In other words, the KPI summary data is a quantitative result that briefly summarizes key performance indicators, such as on-time delivery rates and equipment utilization rates.

[0090] In particular, the data from simulation results (or summary data) is stored as structured data, such as in the form of tables or records. For example, it is stored in the form of an RDBMS (Relational Database Management System).

[0091] In particular, simulation results are indexed by the modeling ID (modeling identification information) and the simulation execution number (execution identification information). That is, the modeling ID identifies a specific supply chain model. In other words, when each modeling variable in a supply chain model is determined to a specific numerical value, that model represents a single supply chain state, and a simulation is performed for that specific supply chain state. Multiple simulations can be performed on the same model. In this case, the simulation execution number identifies the execution of each simulation.

[0092] Result retrieval knowledge consists of knowledge regarding rules for retrieving simulation results. It comprises rules or knowledge regarding the structure of simulation result data (the structure of a relational database) and methods for writing SQL queries. In other words, result data structure knowledge consists of rules or knowledge regarding the table structure of the simulation results, the structure of each record, and each field item / format. By utilizing result retrieval knowledge, one can extract desired simulation results by constructing SQL queries.

[0093] Analytical knowledge refers to knowledge used to analyze simulation results, specifically documents (documented knowledge) that explain analysis methods. In other words, analytical knowledge includes methods for analyzing bottlenecks in supply chain models, strategies for optimizing equipment operation, and methods to improve on-time delivery rates. Analytical knowledge consists of specific guidelines and rules necessary for in-depth analysis, such as procedures for analyzing bottlenecks and methods (procedures) for optimizing equipment or processes.

[0094] The analytical knowledge is utilized by the LLM module (32) (or LLM model) to generate an analysis report or to write a response to a query. In particular, the LLM module (32) refers to the analytical knowledge to generate an in-depth analysis and optimization proposal.

[0095] Meanwhile, the supply chain database (33) is provided with a search method suitable for each type of data.

[0096] Simulation results (original results or summary information) are retrieved using a standard field indexing method that searches by referencing data fields (or key fields), etc. Specifically, queries such as SQL are used to extract the values ​​of the requested fields of records that match the values ​​of specific fields (key fields, etc.).

[0097] Furthermore, related knowledge is stored and retrieved based on search terms or vectors. In other words, to efficiently store vast amounts of knowledge data and to quickly and accurately search for related knowledge according to user queries, a vector database is introduced to enable semantic-based search.

[0098] Specifically, knowledge data is converted into numerical vectors so that computers can process them. This allows the meaning and characteristics of knowledge to be represented in a vector space. The content of knowledge is converted into vectors using a semantic embedding method.

[0099] At this point, the conversation content is converted into a vector using the LLM model (or the embedding module of the LLM model). Embedding is the process of representing the meaning and context of text as a high-dimensional vector. In other words, the LLM model provides an embedding function or interface (API, Application Programming Interface) that converts into a vector, and the conversation content is converted into a vector using this embedding function or interface.

[0100] Alternatively, as another embodiment, conventional vector databases such as FAISS, Pinecone, etc. are used.

[0101] Next, the simulation processing unit (34) simulates the supply chain model through the simulator (31) and collects the results.

[0102] That is, the simulation processing unit (34) initializes the supply chain model or simulator (31) by specifying each modeling data (modeling variables, etc.) of the supply chain model described above. That is, the supply chain model or simulator (31) is specified by assigning specific numerical values, etc. to each modeling variable. The specified supply chain model represents a specific supply chain state.

[0103] Additionally, the simulation processing unit (34) causes the simulator (31) to simulate the supply chain model and collects the simulation results of the simulator (31). The simulation results generated by the simulator (31) consist of original result data generated by the simulation, such as equipment operation records, product flow, and inventory levels, and summary data (or performance indicator data) that summarizes the original results to provide performance indicators (KPIs), etc.

[0104] The original results of the simulation consist of items related to the previously described delivery compliance prediction, product flow records, production plan optimization, equipment operation records, inventory level management, etc. Preferably, the simulation results consist of results simulated by a scenario optimized in the corresponding supply chain model (or supply chain state).

[0105] Summary information from the simulation is generated by deriving performance indicator (KPI) results from the simulation results. Preferably, the performance indicators (KPIs) consist of equipment utilization rate and on-time delivery rate. The equipment utilization rate is a ratio indicating how well the machines in the factory are operating, and the on-time delivery rate is the ratio of the ability to supply products on time to the date ordered by the customer.

[0106] Next, the query processing unit (36) receives a query message regarding the supply chain. In particular, the query processing unit (36) reconstructs the query message by including information from the previous conversation for the context of the conversation.

[0107] Specifically, the query processing unit (36) receives a query message through the LLM module (32) or the LLM model. That is, it receives a query message regarding the user's supply chain through the prompt of the LLM model, etc.

[0108] In addition, the query processing unit (35) analyzes the input supply chain query message (or query prompt) and extracts search information such as search terms.

[0109] For example, a supply chain query can be entered as follows.

[0110] [Example of Query Message #1]

[0111] What should be done to optimize the current facility utilization rate?

[0112] [Example of Query Message #2]

[0113] What measures are needed to raise the on-time delivery rate to 95%?

[0114] [Example of Query Message #3]

[0115] The current on-time compliance rate for Demand A is 85%. Would additional equipment investment or work schedule adjustments be necessary to raise the compliance rate to 95%?

[0116] [Example of Query Message #4]

[0117] "How to resolve the current bottleneck between the OPER_3 and OPER_4 processes?"

[0118] Additionally, the query processing unit (36) stores information from previous conversations for the context of the conversation and reconstructs the current query by including the information from previous conversations. That is, in order to understand the context of the conversation, information from previous conversations is stored (or data provided in each conversation is stored) so that the LLM model can maintain the flow of the conversation.

[0119] For example, this illustrates a case where the previous conversation and the current conversation are given as follows.

[0120] Previous Conversation: "How to Improve the Bottleneck in the OPER_3 Process?"

[0121] Current Conversation: "How should I adjust OPER_4?"

[0122] At this point, related data is searched, including the information of OPER_3 mentioned in the previous conversation. To do this, an LLM prompt is generated using the information saved from the previous conversation.

[0123] □ Context-based prompt example □

[0124] [Previous Conversation Info] OPER_3 Process: Utilization Rate 70%, Bottleneck Occurred

[0125] [Current Inquiry] How do I adjust the OPER_4 process?

[0126] Additionally, the query processing unit (36) determines whether it is a simple data request or a (deepened) analysis request through the LLM module (32).

[0127] The LLM model determines whether a query is a simple data request or a request for deep analysis by considering factors such as whether relevant keywords are included in the query, the contextual pattern of the query, and the previous conversational context. For example, the LLM model determines a request for deep analysis if keywords such as "optimization," "improvement," "solution," "prediction," or "analysis" are included. Alternatively, the LLM model determines a request for deep analysis if the query involves comparison, a request for a solution, or a prediction. For instance, a question like "How do you resolve the bottleneck in the OPER_3 process?" is determined to require a bottleneck analysis method, while a question like "How do you optimize the equipment utilization rate?" is determined to require an equipment optimization plan.

[0128] In addition, the LLM model considers the context of the previous conversation, and determines it as an in-depth analysis if the conversation proceeds as follows.

[0129] Query: "What is the current facility utilization rate?" → Simple data request

[0130] Question: "What should be done to improve the facility utilization rate?" → Request for in-depth analysis

[0131] Meanwhile, preferably, the query processing unit (36) can generate a prompt asking what kind of query it is and transmit it to the LLM module (32) to determine the type of query through the answer.

[0132] □ Examples of prompts for query types □

[0133] [Prompt] Please let me know if the following query is a simple data request or a request for analysis. <Query> "What should I do to improve the equipment utilization rate?"

[0134] Next, the knowledge search unit (37) extracts search information, such as search terms and keywords, from the previous query message, searches the supply chain database (33) using the search information, and retrieves the searched related information or related knowledge.

[0135] Additionally, the knowledge search unit (37) extracts search information, such as search terms (or keywords) and search phrases, from the supply chain query message to search for related information in the supply chain database (33). Preferably, the query message is a reconstructed query message that includes context (previous conversation information). That is, the LLM module (32) understands the context of the query based on the previous conversation information and performs the role of supplementing necessary search terms or recommending suitable data items.

[0136] In particular, the knowledge search unit (37) analyzes supply chain query messages through the LLM module (32) to extract key keywords, phrases, and entities.

[0137] In the previous example #1, search terms (or keywords) such as "equipment utilization rate" and "optimization" are extracted. Additionally, in example #2, search terms such as "on-time delivery rate", "95%", and "solution" are extracted.

[0138] In addition, search terms and phrases such as "on-time compliance rate," "on-time standards," "work schedule optimization," and "benefits of adding equipment" are extracted from the query message of #3.

[0139] In particular, the knowledge search unit (37) generates a prompt to inquire about search information for a query message, queries the LLM module (32), and obtains the response to obtain search information.

[0140] □ Example of Search Information Extraction □

[0141] [Inquiry Message] "What measures are needed to raise the on-time delivery rate to 95%?"

[0142] [Prompt] "Extract the key keywords or key phrases from the following query message: {Query Message}."

[0143] [LLM Response] "On-time delivery rate", "95%", "Measure"

[0144] Additionally, the knowledge search unit (37) extracts data table names or data items of simulation results that are mapped to the extracted keywords. That is, it performs mapping between the extracted keywords and data tables (records), fields, and search rules defined in related knowledge (Knowledge, background knowledge, modeling knowledge, etc.).

[0145] At this time, the knowledge search unit (37) refers to a predefined mapping rule to link the extracted keyword with a data structure (table and field) stored in the supply chain database (33). The mapping rule describes the relationship between each table and field stored in the database and the keyword, and as an example, "on-time compliance rate" is mapped to the DUE_IN_QTY and PLAN_QUANTITY fields of the HT_DMD_RTF table.

[0146] That is, if the keyword is "on-time compliance rate", the data mapped to it is extracted as the DUE_IN_QTY and PLAN_QUANTITY fields of the HT_DMD_RTF table. Alternatively, the HT_RES table, TO_STATUS, and UTILIZATION are extracted as tables or items related (mapped) to "bottleneck".

[0147] Additionally, the knowledge search unit (37) extracts mapping results through the LLM module (32) when the keyword and data mapping rules are incomplete or when the query message deviates somewhat from the existing rules. That is, the knowledge search unit (37) recommends mapping by supplementing the data based on the learned content of similar query cases through the LLM module (32). This supports compensating for the limitations of the rule-based approach and enabling the derivation of appropriate results for various types of queries entered by the user.

[0148] That is, the knowledge search unit (37) maintains (stores) a query case database (or log system) that records previous queries and mapping results. When a new query comes in, the knowledge search unit (37) searches for past cases similar to the query and transmits them to the LLM module (32) along with the cases to derive more accurate mapping results.

[0149] Additionally, the knowledge search unit (37) automatically generates SQL queries based on data structures defined in knowledge, such as search rules. This process is performed based on database structures defined by data mapping rules, and the generated SQL is designed to efficiently search for simulation results. In the search knowledge (or search rules), conditions and filtering logic required for writing SQL are defined based on database structures and logical relationships.

[0150] For example, you can automatically generate SQL queries as follows.

[0151] □ SQL Generation Example □

[0152] "SELECT DUE_IN_QTY, PLAN_QUANTITY FROM HT_DMD_RTF WHERE DUE_IN_QTY / PLAN_QUANTITY < 0.95; "

[0153] Next, the knowledge search unit (37) searches the supply chain database (33) using the previously extracted search information and retrieves the searched relevant information or relevant knowledge.

[0154] In the previous example, the supply chain database (33) is searched using search terms or search phrases such as "equipment utilization rate," "optimization," "delivery compliance rate," "delivery criteria," "work schedule optimization," and "benefits of adding equipment." Through this, simulation results or related knowledge related to the query are retrieved.

[0155] At this time, the knowledge search unit (37) retrieves and obtains the simulation results through the previously automatically generated SQL, etc. That is, since the simulation results are stored as structured data such as RDBMS, they are retrieved through structured query statements such as SQL.

[0156] Additionally, the knowledge search unit (37) searches for and obtains background knowledge, analysis knowledge (knowledge about analysis methods), etc. by means of an embedding vector (or embedding-based search). At this time, an embedding vector for searching is constructed using a search term or search phrase, etc., and related knowledge (background knowledge, analysis knowledge, etc.) is searched using the constructed embedding vector.

[0157] In particular, knowledge of analysis methods is retrieved when a user's query goes beyond simply requesting data and requires in-depth analysis or optimization strategies.

[0158] For example, in the example #1 mentioned earlier, data related to the optimization of equipment utilization rates is searched. At this time, the embedding vector is composed of search terms and phrases such as "equipment utilization rate" and "optimization". Additionally, in the query message of #3, data related to on-time delivery rates (delivery criteria, work schedule optimization methods, benefits of adding equipment, etc.) is searched from the supply chain database (33). Additionally, in the query message of #4, the embedding vector is composed of search information such as "bottleneck analysis" and "process optimization".

[0159] That is, the knowledge search unit (37) converts the text of unstructured data, such as a report, into a vector and performs a similarity search through the operation of the embedding vector.

[0160] In addition, the knowledge search unit (37) searches differently depending on whether the supply chain query is a simple data request or an in-depth analysis request.

[0161] That is, in the case of a simple data request, an SQL query is automatically generated to retrieve the necessary data from the supply chain database (33).

[0162] Example: "SELECT UTILIZATION FROM HT_RES WHERE UTILIZATION < 0.8;"

[0163] Furthermore, in the case of a request for in-depth analysis, embedding vector-based search is performed alongside SQL search to additionally retrieve background knowledge and analysis methods. For example, in addition to the SQL search mentioned above, the embedding vector searches for semantically similar data by performing operations on analysis-related keywords such as "equipment optimization," "bottlenecks," and "work schedule adjustment" within the vector space.

[0164] Next, the response processing unit (38) provides the retrieved information or knowledge along with the supply chain query to the LLM module (32) to generate a response to the supply chain query.

[0165] That is, the response processing unit (38) injects the data searched by the knowledge search unit (37) along with the supply chain query message earlier into the LLM module (32) as additional input. The LLM module (32) generates a response by combining the previously learned knowledge and the searched information.

[0166] In particular, the response processing unit (38) generates a prompt requesting a response to a supply chain query by referring to the retrieved data. Preferably, a template is set up, and a response request template is generated using the corresponding prompt template.

[0167] □ Response Request Prompt Template #1] □

[0168] Related Knowledge: {Search information on related knowledge}

[0169] Simulation Results: {Search information for simulation results}

[0170] Please answer the following question based on the information above: {Question Message}

[0172] Here, the {…} symbol is used to indicate the parts to be replaced. The part of {search data for related knowledge} is replaced with data retrieved from the previously retrieved knowledge (background knowledge, analysis knowledge, etc.). Also, the part of {search data for simulation results} is replaced with data retrieved from the simulation. Additionally, {query message} is the query message received from the previously retrieved query processing unit (36).

[0173] Additionally, preferably, in order to provide a clearer and easier-to-understand response to the user's query, the retrieved quantitative data (simulation results) is converted into a sentence. For example, if the original data (in table form) is extracted with equipment utilization rate = 75% and additional introduction effect = 85%, it is converted into a sentence such as "The equipment utilization rate is 75% and can be increased to 85% upon additional introduction of equipment." At this time, the sentence conversion is transmitted to the LLM module (32), and the LLM can optimize it into a natural sentence by supplementing the context. Alternatively, the quantitative data can be processed into a sentence using a predefined template or conversion rule. Alternatively, the sentence based on the template or conversion rule can be supplemented through the LLM module (32).

[0174] Meanwhile, preferably, the query message to be included in the prompt may be a query message that includes context. That is, the query message is reconstructed by including previous conversation information to maintain the context of the preceding conversation. The {Query Message} section may be entered (replaced) with the query message reconstructed in this way.

[0175] The example of the response request prompt for the template above is as follows.

[0176] □ Example of Response Request Prompt #1] □

[0177] Related Knowledge: {Information retrieved from related knowledge (knowledge information retrieved using terms such as 'equipment utilization rate' and 'optimization')}

[0178] Simulation Results: "The current facility utilization rate is 75%, and it is expected to rise to 85% with the introduction of additional facilities."

[0179] Please answer the following questions based on the information above:

[0180] What should be done to optimize the current facility utilization rate?

[0181] Meanwhile, the response processing unit (38) provides the LLM module (32) with analysis knowledge related to the query message, so that the LLM module (32) can perform advanced analysis (bottleneck analysis, cause and solution of bottleneck process equipment, optimization plan, equipment layout and work schedule optimization plan, etc.) based on this. That is, the previously searched analysis knowledge (or knowledge about analysis methods) consists of specific contents such as "bottleneck analysis procedures" and "equipment optimization plans," and in particular, methodologies or procedures for performing advanced analysis such as bottleneck analysis and equipment optimization. The LLM module (32) can perform analysis by referring to these analysis knowledge methods.

[0182] Additionally, the response processing unit (38) transmits the response request prompt to the LLM module (32) and receives and outputs a response message that is responded to by the prompt.

[0183] That is, when the response processing unit (38) generates a response request prompt and transmits it to the LLM module (32), the LLM module (32) replies with a response message in response to the prompt. At this time, the replied response message is output.

[0184] When the above response request prompt is provided to the LLM module (32), the following response is received.

[0185] [Example of response to response request prompt #1]

[0186] To increase the current on-time compliance rate for Demand A from 85% to 95%, the following adjustments may be necessary: ​​1. Additional Equipment Investment: Production speed can be increased by introducing additional equipment to key bottleneck processes to improve on-time compliance. 2. Work Schedule Adjustment: Current process operating schedules can be adjusted to reduce downtime in key processes and maximize resource utilization between processes. 3. Reallocation of Existing Resources: The target on-time compliance rate can be achieved by efficiently allocating inventory and resources without additional equipment investment.

[0187] Also, print the received response.

[0188] Figure 7 shows a screen illustrating an example of a response to a supply chain query.

[0189] Next, a method for responding to supply chain queries using simulation results and related knowledge according to an embodiment of the present invention is described as an example.

[0190] First, we will provide examples of supply chain questions and answers related to on-time delivery rates.

[0191] Regarding the delivery fulfillment rate, a supply chain query can be entered as follows.

[0192] [Question Prompt #11]

[0193] The current on-time delivery rate for the FINAL_PROD product is 85%. What measures are needed to raise the on-time delivery rate to 95%?

[0194] As follows, the related simulation results include items such as delivery on-time prediction, product flow records, and equipment operation records.

[0195] [Related Simulation Results Found #11]

[0196] 1. On-time delivery prediction

[0197] DEMAND_ID: ID identifying a specific demand

[0198] PLAN_QUANTITY: Total quantity requested based on demand

[0199] DUE_IN_QTY: Quantity of products received within the delivery date of the corresponding demand

[0200] 2. Product Flow Record

[0201] PROD_CODE: Product code currently being processed in the process

[0202] TO_OPER_ID: Current process ID where the product is in transit

[0203] TO_RES_ID: ID of the facility where the product is being processed

[0204] EVENT_TIME: The point in time when the task start / end time is recorded

[0205] 3. Equipment Operation Records

[0206] RES_ID: Facility ID

[0207] EVENT_TIME: Equipment operation start and end times

[0208] TO_STATUS: Current status of the facility (e.g., WAIT, PROCESS)

[0209] In addition, the related background knowledge and analytical knowledge are as follows.

[0210] [Related Knowledge Found #11]

[0211] 1. When running the simulation, a score expressed as a percentage called the delivery compliance rate is generated, which serves as a metric to evaluate the results. The delivery compliance rate refers to the percentage of products that meet the demand's due date. For example, if all products meet the delivery dates for all demand, the result is 100%. In other words, the extent to which delivery dates are met in the simulation results is the most important evaluation factor.

[0212] 2. If the delivery date was not met, it is necessary to check the delivery compliance rate for each product.

[0213] 3. If there are products with particularly poor on-time delivery rates or if the overall RTF is low, it is necessary to identify which process is exhibiting a bottleneck.

[0214] 4. To improve on-time delivery rates, you can reduce the number of SETUP CHANGE operations by adding equipment to perform processes where necks occur, or by changing the DSP_RULE_ID of the corresponding process in the PP_OPER table to ORP_RES_REQ.

[0215] 5. There may be multiple demands for the same product. This occurs when the orderers or due dates differ. Therefore, the delivery satisfaction rate by demand may be more meaningful than the delivery compliance rate by product.

[0216] 6. To calculate the on-time delivery rate, check DUE_IN_QTY and PLAN_QUANTITY at the last Warehouse (=Demand Warehouse). The formula is as follows. From this calculation result, the RTF per demand can be obtained, and calculating all demands at once gives the total RTF. RTF = sum(Quantity completed before the due date) / sum(Total planned quantity)

[0217] Next, the response obtained by referring to the relevant simulation results and knowledge above is as follows.

[0218] [Response Example #11]

[0219] It has been confirmed that the current on-time delivery rate for the FINAL_PROD product is 85%. To improve this to 95%, we can consider increasing the utilization rate of the OPER_3 process, where a bottleneck is occurring, by 10%, and deploying one additional piece of equipment to supplement the processing capacity of the OPER_4 process. Additionally, by reducing the number of equipment job changes from the existing 4 to 2, the on-time delivery rate can be improved to the target level.

[0220] Next, examples of supply chain queries and answers regarding inventory or process optimization are provided.

[0221] [Question Prompt #12]

[0222] "Currently, inventory is accumulating excessively between the OPER_3 and OPER_4 processes. What is the optimal equipment operation schedule to resolve this bottleneck?"

[0223] As follows, the related simulation results include items such as delivery compliance prediction, product flow records, equipment operation records, and inventory level management.

[0224] [Related Simulation Results Found #11]

[0225] 1. On-time delivery prediction

[0226] DEMAND_ID: ID identifying a specific demand

[0227] PLAN_QUANTITY: Total quantity requested based on demand

[0228] DUE_IN_QTY: Quantity of products received within the delivery date of the corresponding demand

[0229] 2. Product Flow Record

[0230] PROD_CODE: Product code being processed in each process

[0231] TO_OPER_ID: Current process ID where the product is in transit

[0232] TO_RES_ID: ID of the facility processing the product

[0233] EVENT_TIME: The time when an event (task start / end) occurs

[0234] 3. Equipment Operation Records

[0235] RES_ID: ID of the facility in operation

[0236] EVENT_TIME: Equipment operation start and end times

[0237] TO_STATUS: Equipment status (WAIT, PROCESS, MOVE, etc.)

[0238] 4. Inventory Level Management

[0239] LOT_QTY: Total quantity of products in a specific process or warehouse

[0240] In addition, the related background knowledge and analytical knowledge are as follows.

[0241] [Related Knowledge Found #12]

[0242] Analysis method for finding the neck point

[0243] Wait Time Analysis: Calculate wait times for each process to identify bottleneck processes.

[0244] Utilization Analysis: Analyze the difference in processing speed between the bottleneck process and the next process by checking the equipment utilization rate.

[0245] Analysis of Neck Phenomenon Caused by Insufficient Aux Units: Determine whether additional equipment or resources are required.

[0246] Next, the response obtained by referring to the relevant simulation results and knowledge above is as follows.

[0247] [Response Example #12]

[0248] To resolve the current issue of excessive inventory accumulation between the OPER_3 and OPER_4 processes, we propose the following measures: 1. Adjustment of OPER_3 Process Operating Rate: OPER_3 Process: Reduce operation time by 10% and increase the operation interval by 15 minutes. 2. Verification of OPER_4 Process Process Processing Capacity: OPER_4 Process: Deploy one additional piece of equipment or extend the operation time of the existing equipment by 20%.

[0250] Next, we provide examples of supply chain queries and responses regarding facility utilization rates.

[0251] [Question Prompt #13]

[0252] "Currently, the MAC_1 facility operates for 16 hours a day. Please suggest measures to maximize the facility's utilization rate under these circumstances."

[0253] As follows, the related simulation results include items such as delivery on-time prediction, product flow records, production plan optimization, and equipment operation records.

[0254] [Related Simulation Results Found #13]

[0255] Regarding facility utilization rates

[0256] 1. On-time delivery prediction

[0257] DEMAND_ID: Demand ID

[0258] PLAN_QUANTITY: Target Production Quantity

[0259] DUE_IN_QTY: Quantity of products received within the delivery date of the corresponding demand

[0260] LPST_DATE: Minimum input time

[0261] 2. Product Flow Record

[0262] PROD_CODE: Product Code

[0263] TO_OPER_ID: Process ID

[0264] TO_RES_ID: Related Facility ID

[0265] EVENT_TIME: Work time information

[0266] 3. Production Planning Optimization

[0267] CURR_IN_QTY: Input quantity required in the current process

[0268] NEXT_OUT_QTY: Quantity to send to the next process

[0269] 4. Equipment Operation Records

[0270] RES_ID: Facility ID

[0271] EVENT_TIME: Equipment operation start and end times

[0272] In addition, the related background knowledge and analytical knowledge are as follows.

[0273] [Related Knowledge Found #13]

[0274] 1. On-time Delivery and Capacity Utilization Analysis

[0275] Check the overall RTF (On-Time Failure Rate) through DEMAND_ID, PLAN_QUANTITY, and DUE_IN_QTY.

[0276] Check the task completion status by referring to TOT_OUT_QTY and TO_STATUS.

[0277] Analyze the EVENT_TIME and TO_STATUS of equipment MAC_1 to extract the Idle Time (waiting time).

[0278] 2. Reduced number of setup changes

[0279] Calculate the operation switching frequency (Event Count) of the bottleneck process (OPER_3) and equipment MAC_1 based on TO_OPER_ID and TO_RES_ID.

[0280] Based on this, adjust the DSP_RULE_ID to reduce the number of Setup attempts from the existing 4 to 2.

[0281] 3. Distributing Bottleneck Processes

[0282] Check the LOT_CNT and LOT_QTY of the bottleneck process (OPER_3).

[0283] Resolve the bottleneck by deploying the MAC_2 facility to TO_RES_ID.

[0284] Next, the response obtained by referring to the relevant simulation results and knowledge above is as follows.

[0285] [Response Example #13]

[0286] An analysis of the current MAC_1 facility's operating rate based on a 16-hour cycle revealed an average waiting time of 30 minutes between tasks, resulting in approximately 2 hours of unnecessary idle time. To address this, one solution involves readjusting the work schedule and applying a rule to reduce the number of task changes from the current 4 to 2, thereby securing an additional 1 hour of operating time. Furthermore, by deploying the MAC_2 facility as a backup to the OPER_3 process—where a bottleneck is occurring—to distribute the load, the overall operating rate can be improved from the current 80% to 95%.

[0287] Although the invention made by the inventor has been specifically described above according to the embodiments, the invention is not limited to the above embodiments and can be modified in various ways without departing from the gist thereof. Explanation of the symbols

[0288] 10 : Computer terminal 30: Analysis System 30a: Analysis Client 30b : Analysis Server 31 : Simulator 32 : LLM Module 33: Supply Chain Database 34: Mock Processing Unit 35: Knowledge Storage Unit 36: Query Processing Unit 37 : Knowledge Search Unit 38 : Response Processing Unit 80 : Network

Claims

Claim 1 A simulation-based supply chain analysis system using a large language model comprises: a simulator that simulates a supply chain model and outputs simulation results; an LLM module equipped with a large language model (LLM); a supply chain database that stores data; a simulation processing unit that performs a simulation through the simulator according to a given supply chain model, collects simulation results, and stores them in the supply chain database; a knowledge storage unit that stores modeling information of the supply chain model, knowledge about the supply chain model, analysis knowledge for analyzing the simulation results, and the collected simulation results in the supply chain database; a query processing unit that receives a first query message regarding the supply chain; and a knowledge search unit that extracts search information from the first query message and searches the supply chain database using the search information. and, the system includes a response processing unit that generates a second query message containing the first query message along with the retrieved data, queries the LLM module with the second query message to obtain a response message, and outputs the obtained response message; the simulation result is composed of an original result generated by the simulation and summary information that summarizes the original result; the original result includes items regarding delivery compliance prediction, product flow record, production plan optimization, equipment operation record, and inventory level management; the summary information is composed of performance indicators (KPIs), including delivery compliance rate and equipment operation rate; the simulation result is indexed using modeling identification information that identifies the supply chain model and execution identification information that identifies the runtime of the simulation executed with the modeling of the supply chain model; the supply chain model comprises process information composed of time required per process; equipment information composed of production time and production unit information; product linkage information that indicates the process of an input product being converted into a production product through a specific process and is composed of input quantity and production quantity;A simulation-based supply chain analysis system using a large language model, characterized by including delivery information including a delivery date, delivery quantity, and customer; wherein the items for delivery compliance prediction are items that predict whether a product can be produced in accordance with the customer's delivery date, and consist of whether the product is on the delivery date and quantity; wherein the items for product flow records are items that record the process of a product passing through each process, and consist of records of the product passing through each process and equipment operation information; wherein the items for production plan optimization are composed of production plans and progress status by process; wherein the items for equipment operation records include the operating status of the equipment, idle time, and work changeover records; wherein the items for inventory level management include inventory quantities by process; and wherein the simulation result is a result simulated by at least one optimal scenario scheduled in a specific supply chain state, and the supply chain state is a state in which each modeling variable in the corresponding supply chain model is set to a specific value. Claim 2 A simulation-based supply chain analysis system using a large language model according to claim 1, characterized in that the modeling information of the supply chain model includes processes, facilities, working hours, products, product connection structures, demand, and delivery deadlines. Claim 3 delete Claim 4 A simulation-based supply chain analysis system using a large language model according to claim 1, wherein the analysis knowledge is knowledge for analyzing the simulation results and includes guidelines and rules for performing the analysis method. Claim 5 A simulation-based supply chain analysis system using a large language model according to claim 1, characterized in that the supply chain database provides a function to search for the simulation results using SQL queries, and provides a function to search for the modeling information of the supply chain model, knowledge about the supply chain model, and analysis knowledge for analyzing the simulation results using embedding vectors. Claim 6 A simulation-based supply chain analysis system using a large language model, wherein, in claim 1, the query processing unit reconstructs a query message including information from a previous conversation. Claim 7 A simulation-based supply chain analysis system using a large language model according to claim 1, wherein the knowledge search unit extracts search information including a search term from the query message, extracts a data table name or data item of a simulation result mapped to the search term, generates an SQL query statement using the extracted table name or data item, searches the simulation result using the generated SQL query statement, generates an embedding vector of the search information, and searches for modeling information of the supply chain model, knowledge about the supply chain model, and analysis knowledge for analyzing the simulation result using the generated embedding vector. Claim 8 A simulation-based supply chain analysis system using a large language model according to claim 7, wherein the second query message includes a retrieved simulation result, retrieved knowledge, and the first query message, and is written to request a response to the first query message by referring to the retrieved simulation result and the retrieved knowledge.

Images