Apparatus, Driving Method, and System for Military Analysis Report Generation Using AI Technology

Abstract

The present invention relates to a military analysis report generation device utilizing artificial intelligence technology, a method for operating the device, and an analysis report generation system. A military analysis report generation device according to an embodiment of the present invention may include a communication interface unit for receiving simulation result data generated during or as a result of a simulation execution process from a defense modeling and simulation (M&S) device, and a control unit that generates a quantitative analysis result by performing functional rule-based analysis of maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the simulation result data received through the communication interface unit, searches for a reference context from a storage unit using a search augmentation generation (RAG) method using a vector index in which an external reference knowledge document, including military doctrine, rules of engagement (ROE), sensor model, or simulation logic analysis document corresponding to the quantitative analysis result, is stored as an embedding vector, and inputs the quantitative analysis result and the reference context into a local large language model operating in a closed network environment to automatically generate a military analysis report on the defense M&S simulation result.

Description

Technology Field

[0001] The present invention relates to a military analysis report generation device utilizing artificial intelligence technology, a method for operating the device, and an analysis report generation system. More specifically, it relates to a military analysis report generation device, a method for operating the device, and an analysis report generation system that automatically preprocesses and analyzes simulation log data by utilizing information retrieval and inference techniques based on a Large Language Model (LLM) and Retrieval-Augmented Generation (RAG) operating in a local environment, in order to automatically analyze simulation result data generated in the field of defense modeling and simulation (M&S) and generate reports by applying, for example, artificial intelligence (AI), natural language processing (NLP), data processing, information processing devices, and software technologies. Background Technology

[0002] In the defense sector, Modeling and Simulation (M&S) is essential for evaluating weapon system performance, verifying engagement procedures, analyzing tactical effects, and conducting operational studies. In this defense M&S environment, a large volume of diverse simulation result data is generated, including maneuver information, sensor detection and tracking data, engagement and hit results, and changes in weapon system status. As scenarios become more complex, the scale of the data and the difficulty of analysis increase further. Existing analysis methods require analysts to manually review simulation log files, extract and organize key events, and write reports. This results in significant time consumption and inconsistencies in the quality of results depending on the analyst's experience and capabilities.

[0003] Meanwhile, with the recent advancement of Large Language Model (LLM) technology, which performs the functions of analyzing, interpreting, and summarizing text-based information, Retrieval-Augmented Generation (RAG) technology, which combines external knowledge into the model's inference process, is gaining attention. However, general LLM-based analysis technologies often rely on cloud environments requiring internet connectivity or depend on vast amounts of external training data, making it difficult to meet the security and closed-system requirements demanded by the defense sector.

[0004] Defense M&S data may contain sensitive information, such as operational concepts (CONOPS), weapon system performance parameters, engagement algorithms, and tactical operational procedures; therefore, processing it in environments where there is a risk of data leakage to external networks is restricted. In particular, the server environments of defense research institutions and the military are equipped with strict information security systems, including internet isolation, network segmentation, and restrictions on external access, making the use of cloud-based LLM services virtually impossible. For this reason, to analyze defense M&S data, it is essential to apply a local LLM model that can operate in a closed network or single-equipment environment without an internet connection.

[0005] In addition, since defense M&S analysis requires the simultaneous analysis of quantitative figures (e.g., time to detection, time to engagement, probability of hit, TTD / TTK, etc.) and tactical and doctrine-based simulation logic, it is difficult to ensure adequate analysis quality using only simple natural language generation models. Therefore, there is a need for information retrieval technology (RAG), preprocessing technology, and technology that combines quantitative analysis and text analysis to efficiently analyze large amounts of log data while satisfying the characteristics and security requirements of military simulation. Prior art literature

[0006] Korean Registered Patent Publication No. 10-2239377 (April 6, 2021) Korean Published Patent Publication No. 10-2016-0116144 (October 7, 2016) Korean Published Patent Publication No. 10-2022-0085125 (June 22, 2022) Korean Published Patent Publication No. 10-2022-0118822 (August 26, 2022) Korean Published Patent Publication No. 10-2023-0139974 (October 11, 2023) International Published Patent Publication No. 2023 / 230417 (October 26, 2023) US Published Patent Publication No. 2023 / 0259836 (August 17, 2023) US Published Patent Publication No. 2024 / 0047628 (2024.02.08)

[0007] A Study on LLM-based Log Anomaly Detection for Integrated Ship Cyber ​​Defense Systems (2025 Korea Society for Military Science and Technology Conference: Cyber ​​Sector (Pilot), Ok Ji-won et al., pp. 1908–1909) Analysis of the Effectiveness of an LLM-based xAPI Converter for Military Education and Training Data Standardization (2025 Korea Society for Military Science and Technology Conference: Systems Engineering Sector, Im Do-eon et al., pp. 1644–1645) A Study on Improvement Measures for Korean Military Terminology Processing in sLLM-based Anomaly Analysis Systems (2025 Korea Society for Military Science and Technology Conference: Information and Communication Sector, Ji Seung-bae et al., pp. 1024–1025) The problem to be solved

[0008] The technical problem to be solved by the present invention is to provide a military analysis report generation device, a method for operating the device, and an analysis report generation system that automatically preprocess and analyze simulation log data by utilizing information retrieval and inference techniques based on a Large Language Model (LLM) and Retrieval-Augmented Generation (RAG) operating in a local environment, for the purpose of automatically analyzing simulation result data generated in the field of defense modeling and simulation (M&S) and generating reports by applying, for example, artificial intelligence (AI), natural language processing (NLP), data processing, information processing device, and software technology.

[0009] Furthermore, the technical problem that the present invention aims to solve is to provide a military analysis report generation device, a method for operating the device, and an analysis report generation system that apply a new technical methodology capable of efficiently analyzing and organizing complex simulation results by utilizing a local LLM capable of operating without an internet connection and an information retrieval technique dedicated to military simulation, in a situation where technical means are required to process large-scale simulation result log data generated in a defense M&S environment in a secure closed network environment, reduce reliance on manual work by analysts, and consistently perform quantitative and qualitative analysis.

[0010] The technical objective of this invention is to overcome the limitations of the existing method, which requires analysts to manually review and summarize large-scale simulation data generated in defense modeling and simulation (M&S) environments, to ensure the reliability and consistency of analysis results, and to enable automated analysis using artificial intelligence even in a closed-network-based security environment. Defense M&S data consists of various forms of quantitative and qualitative information, such as maneuver routes, detection and tracking information, engagement and hit results, and damage assessments. As the scale of data increases rapidly with the increase in scenario complexity, it is difficult to simultaneously ensure accuracy and efficiency through manual analysis by analysts alone. Furthermore, since the defense sector has strict security requirements that prevent the use of external network-based cloud LLM services, an automated analysis system utilizing LLM that operates in a local environment without an internet connection is essential.

[0011] Meanwhile, although Large Language Models (LLM) and Retrieval-Augmented Generation (RAG) technologies have recently been utilized across various industries, these technologies focus on general text analysis or web-based knowledge retrieval and fail to provide the capabilities to interpret the structural characteristics of defense M&S logs (such as time-based continuous data, engagement events, and quantitative analysis parameters). Furthermore, the lack of a structure capable of automatically performing analyses combined with core defense knowledge—such as specific doctrines, rules of engagement, and sensor detection models—has led to a problem where variations in analysts' experience affect the analysis results.

[0012] Accordingly, the present invention aims to solve the following technical problems. First, to reduce the reliance on manual work by analysts by automatically classifying and quantifying simulation results such as maneuver, detection, engagement, and damage assessment to provide an automated analysis structure for defense M&S log data. Second, to ensure security and operational continuity by utilizing a local LLM capable of operating even in a closed network environment without an internet connection to establish a local LLM-based analysis system that meets defense security requirements. Third, to improve the accuracy and substantiality of LLM analysis by searching and reflecting simulation logic analysis documents, such as rules of engagement, sensor models, and tactical procedures, in RAG-based military knowledge and simulation logic search functions. Fourth, to improve the consistency and quality of analysis results by automatically generating reports that reflect military expression methods, analysis formats, and tactical judgment criteria to automatically generate and standardize defense analysis reports. Fifth, to provide a multifaceted evaluation structure equivalent to that of a professional analyst by reflecting both the numerical characteristics of simulation data and text-based explanations for the integrated processing of quantitative data and text-based analysis.

[0013] In short, the purpose of the present invention is to achieve automation, accuracy, and standardization of defense M&S simulation result analysis, resolve security constraints in closed network operating environments, and simultaneously provide new technical means to intelligently generate military analysis reports by utilizing RAG and local LLM.

[0014] The objectives of the present invention are not limited to those mentioned above, and other unmentioned objectives will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0015] A military analysis report generation device according to an embodiment of the present invention for achieving the above objectives comprises: a communication interface unit for receiving simulation result data generated during or as a result of a simulation execution process from a defense modeling and simulation (M&S) device; and a control unit that generates a quantitative analysis result by performing functional rule-based analysis of maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the simulation result data received through the communication interface unit, searches for a reference context from a storage unit using a search augmentation generation (RAG) method using a vector index in which an external reference knowledge document, including military doctrine, rules of engagement (ROE), sensor model, or simulation logic analysis document corresponding to the quantitative analysis result, is stored as an embedding vector, and inputs the quantitative analysis result and the reference context into a local large language model operating in a closed network environment to automatically generate a military analysis report on the defense M&S simulation result.

[0016] The control unit comprises: a data collection and preprocessing unit that, when simulation results generated from a defense modeling and simulation (M&S) device are provided in the form of log data, receives said log data, performs time alignment, missing value processing, and event classification, and calculates quantitative parameters as analysis reference values ​​that numerically express the state or performance of maneuver, detection, engagement, and damage assessment; a rule-based analysis unit that generates quantitative analysis results by performing functional rule-based analysis on maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using said preprocessed data; a RAG-based information retrieval and context reinforcement unit that constructs a reference context for LLM input by searching for paragraphs related to said quantitative analysis results based on similarity, based on a vector index in which external reference knowledge documents including military doctrine, rules of engagement (ROE), sensor models, and simulation logic analysis documents are stored as embedding vectors; a local LLM-based comprehensive analysis unit that integrates said quantitative analysis results and said reference context using a local Large Language Model (LLM) operating in a closed network environment and generates an analysis statement with a military modified sentence structure; and an analysis statement generated by said comprehensive analysis unit It may include a report generation unit that automatically configures and generates a final report according to the standard defense analysis report template structure of simulation overview, maneuver analysis, detection effect, engagement result, damage assessment, and comprehensive judgment.

[0017] The above data collection and preprocessing unit can perform timestamp correction and common timeline alignment of simulation logs, interpolation of missing values ​​for coordinates, speed, and course, rule-based classification of maneuver, detection, engagement, and damage assessment events, calculation of quantitative parameters such as time to first detect (TTD), time to kill (TTK), detection range, and relative approach speed, and assignment of metadata including tags based on platform, scenario, event type, and doctrine.

[0018] The above rule-based analysis unit may include determining maneuver patterns using course change amount, speed change rate, distance and bearing change amount, evaluating detection performance based on sensor signal-to-noise ratio (SNR), detection probability (Pd), and false alarm probability (Pfa), determining interception success based on engagement initiation conditions and kill radius, and calculating damage levels (Level 0~3) and remaining combat power.

[0019] The above RAG-based information retrieval and context reinforcement unit subdivides military doctrine, rules of engagement, and sensor model documents into paragraph units and converts them into embedding vectors, stores the vectors as an approximate nearest neighbor (ANN)-based vector index, generates a search query vector based on quantitative analysis results to search for related paragraphs using a cosine similarity-based Top-k method, and can configure the searched paragraphs and source information into context blocks for LLM input.

[0020] The above-mentioned local LLM-based comprehensive analysis unit assigns control instructions to the local LLM, including military writing style, analysis stage structure, rules for direct use of quantitative values, and rules for prioritizing RAG evidence, through a system prompt generator; performs local LLM inference by configuring quantitative analysis result blocks, RAG search context blocks, and scenario metadata blocks as integrated inputs; and provides the output analysis statement to the report generation unit after performing duplicate removal, quantitative value matching verification, and confidence score calculation in the post-processor.

[0021] The above report generation unit can automatically assign LLM output statements to six sections corresponding to the simulation overview, maneuver analysis, detection analysis, engagement analysis, damage assessment, and comprehensive judgment of the standard defense analysis report structure, organize RAG citations, source information, and reliability index to include them as additional information, and automatically insert graphs and tables based on quantitative analysis values ​​to output a final report in PDF, DOCX, or TXT format.

[0022] The above control unit can control the execution order of each module configured internally, perform data structure conversion and routing, re-search or apply alternative literature when RAG search is insufficient, request regeneration when LLM output is missing or quantitative values ​​are distorted, and generate an audit record including model version, input data hash, similarity score, and confidence score.

[0023] In addition, a method for operating a military analysis report generation device according to an embodiment of the present invention for achieving the above objectives comprises: a step in which a communication interface unit receives simulation result data generated during or as a result of a simulation execution process from a defense modeling and simulation (M&S) device; and a step in which a control unit generates a quantitative analysis result by performing a functional rule-based analysis of maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the simulation result data received through the communication interface unit, and searches for a reference context in a search augmentation generation (RAG) manner from a storage unit storing said vector index using a vector index in which an external reference knowledge document, including military doctrine, rules of engagement (ROE), sensor model, or simulation logic analysis document corresponding to said quantitative analysis result, is stored as an embedding vector, and inputs said quantitative analysis result and said reference context into a local large language model operating in a closed network environment to automatically generate a military analysis report on the defense M&S simulation result.

[0024] Furthermore, a military analysis report generation system according to an embodiment of the present invention for achieving the above objectives includes a defense modeling and simulation (M&S) device that provides simulation result data generated during or as a result of a simulation execution process, and a military analysis report generation device that generates quantitative analysis results by performing functional rule-based analysis of maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the simulation result data provided from the defense modeling and simulation (M&S) device, searches for a reference context from a storage unit using a search augmentation generation (RAG) method using a vector index in which an external reference knowledge document, including military doctrine, rules of engagement (ROE), sensor model, or simulation logic analysis document corresponding to the quantitative analysis results, is stored as an embedding vector, and inputs the quantitative analysis results and the reference context into a local large language model (LLM) operating in a closed network environment to automatically generate a military analysis report on the defense M&S simulation results.

[0025] Meanwhile, to solve the aforementioned problem, the present invention relates to a system and method for automatically analyzing and generating reports on defense M&S simulation results based on local LLM and RAG. The problem of the present invention is solved by a system and method for automatically analyzing and generating reports on defense M&S simulation results, comprising the following configuration and procedure.

[0026] First, the present invention includes a data collection and preprocessing unit that receives simulation result data, such as maneuver, detection, engagement, and damage assessment generated from defense M&S simulations, and organizes it into a standardized format. The preprocessing unit parses simulation log files to perform time-based data sorting, missing value processing, event classification, and quantitative parameter calculation.

[0027] Second, the present invention includes a rule-based analysis module that analyzes preprocessed M&S data by function. The analysis module automatically performs analysis of maneuver trajectories, detection performance analysis (SNR, Pd, Pfa), engagement event analysis (Kill / Miss), damage level analysis, etc., based on rules or formulas, and generates analysis results in a structured form.

[0028] Third, the present invention includes a RAG-based information retrieval and context reinforcement unit for searching external reference knowledge, such as military doctrines, rules of engagement, and sensor models. The above configuration utilizes a vector index constructed from documents related to defense simulation to search for documents and paragraphs highly relevant to simulation results, and is configured to include the retrieved context in the input of a large language model.

[0029] Fourth, the present invention includes a local LLM-based comprehensive analysis unit that integrates and processes analysis results and RAG search context through a local LLM (Local Large Language Model) capable of operating in a closed network environment without an internet connection. The comprehensive analysis unit includes procedures for generating a system prompt, configuring analysis input text, performing LLM inference, and post-processing output text.

[0030] Fifth, the present invention includes a report generation unit for automatically structuring and outputting the response results of a local LLM according to a report format required by a defense analyst, wherein the report generation unit is configured to allow a user to set a result analysis report template according to the purpose of the scenario or the nature of the analysis task.

[0031] Specifically, the report generation unit receives simulation result data and analysis responses from the local LLM as input, and generates a final report by arranging and summarizing the results according to a report template selected or defined by the user, in accordance with the military bulleted sentence structure, item-by-item analysis system, and standard report table of contents.

[0032] At this time, the above report template can be configured to include at least one item, such as a simulation overview, maneuver characteristics, detection effects, engagement results, damage assessment, and operational implications, and the composition, arrangement order, or inclusion of items can be variably defined depending on the purpose of the scenario, the level of analysis, or the intended use of the report.

[0033] Accordingly, the present invention is configured to automatically generate reports for different purposes, such as operational analysis, power evaluation, or decision support, based on user-configured templates, even for the same simulation result.

[0034] Sixth, the present invention includes a control unit for automating the above-mentioned series of processing flows and integrates and manages the execution of each analysis step, input and output data flow, RAG search call, and local LLM call procedures.

[0035] Through the system and method according to the present invention, defense M&S results can be performed through an automated procedure ranging from collection, refinement, analysis, knowledge retrieval, LLM inference, to report output. In particular, a technical configuration is provided that automatically generates military analysis reports by combining RAG and local LLM, even in a closed network environment.

[0036] In short, the present invention provides a new technical means to automatically generate reliable and consistent defense analysis reports even in a closed network environment by integrating quantitative and qualitative analysis of defense M&S logs, military knowledge search, and local LLM-based natural language generation functions. Effects of the invention

[0037] According to an embodiment of the present invention, large-scale simulation result data generated in a defense modeling and simulation (M&S) environment can be automatically refined, analyzed, and summarized, thereby significantly improving analysis efficiency compared to the existing method where analysts manually reviewed logs. Through a preprocessing unit and a rule-based analysis module, M&S results such as maneuver, detection, engagement, and damage assessment are automatically generated in a structured form, ensuring standardization and consistency in the analysis process.

[0038] Furthermore, according to an embodiment of the present invention, since it includes a RAG-based information retrieval structure for searching external reference knowledge such as military doctrine, rules of engagement, and sensor models, it is possible to automatically reinforce the contextual basis required for analysis results. Accordingly, the reliability of explanations or evaluations generated by the LLM can be increased, and deviations due to the analyst's experience or subjectivity can be minimized.

[0039] Furthermore, according to an embodiment of the present invention, since a local LLM capable of operating in a closed network environment without an internet connection is used, AI-based analysis can be performed without transmitting information requiring security, such as defense M&S data, externally. This enables the stable utilization of the system of the present invention even in limited network environments, such as defense agencies and military research institutes.

[0040] In addition, according to an embodiment of the present invention, the inference results of a local LLM, rule-based analysis results, and RAG search information are combined to automatically generate a report format for military analysts, thereby reducing the time required to write analysis reports and maintaining the homogeneity of report quality. Standardized report structures such as simulation overview, maneuver characteristics, detection effects, engagement results, damage assessment, and operational implications can be automatically generated, making it easy to use as data for the analyst's judgment.

[0041] In short, embodiments of the present invention provide functional and technical effects that can simultaneously achieve the automation of defense M&S result analysis, standardization and explainability of analysis results, maintenance of security in a closed network environment, and efficiency in generating military reports. Brief explanation of the drawing

[0042] FIG. 1 is a block diagram illustrating the overall configuration of a local LLM and RAG-based defense M&S simulation result automatic analysis and report generation system according to the present invention. Figure 2 is a flowchart illustrating the procedure for inputting and refining simulation log data in the data collection and preprocessing unit of the present invention. FIG. 3 is a flowchart illustrating a procedure in which a rule-based analysis module according to the present invention automatically performs quantitative analysis by function according to a scenario purpose or analysis task. Figure 4 is a configuration diagram showing the process of constructing a vector index and similarity search for the RAG-based information retrieval and context reinforcement unit utilized in the present invention. FIG. 5 is a flowchart illustrating the procedure in which a local LLM-based comprehensive analysis unit according to the present invention integrates analysis results and search context to generate text output. FIG. 6 is a configuration diagram showing the process of the report generation unit of the present invention automatically generating a final report according to a defense analysis report template. Figure 7 is a system control diagram showing the method by which the control unit of the present invention manages the overall analysis flow and the data flow. FIG. 8 is a screen example illustrating an example of a user interface (UI) of a defense M&S simulation result automatic analysis and report generation system according to the present invention. FIG. 9 is a configuration diagram showing a military analysis report generation system according to another embodiment of the present invention. FIG. 10 is a block diagram illustrating the detailed structure of the military analysis report generating device of FIG. 9. Figure 11 is a flowchart showing the operation process of the military analysis report generating device of Figure 9. Specific details for implementing the invention

[0043] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims.

[0044] Specific details for implementing the present invention will be described in detail with reference to the drawings attached below. Regardless of the drawings, identical reference numerals refer to identical components, and "and / or" includes each of the mentioned items and all combinations of one or more.

[0045] The terms used herein are for describing the embodiments and are not intended to limit the invention. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used herein, "comprises" and / or "comprising" do not exclude the presence or addition of one or more other components in addition to the components mentioned.

[0046] Unless otherwise defined, all terms used in this specification (including technical and scientific terms) may be used in a meaning that is commonly understood by those skilled in the art to which the present invention pertains. Additionally, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0047] The present invention relates to a system and method for automatically analyzing and generating reports on defense M&S simulation results based on local LLM and RAG. Preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the embodiments of the present invention are for illustrative purposes only, and the technical concept of the present invention is not limited to the description below, and various modifications are possible within the technical concept described in the claims. Additionally, the same reference numerals in the drawings refer to the same or corresponding components.

[0048] FIG. 1 is a block diagram illustrating the overall configuration of a local LLM and RAG-based defense M&S simulation result automatic analysis and report generation system (100) according to the present invention.

[0049] Each component of the above system (100) may be implemented as a computer program (or software module) executed on an information processing device including one or more processors and memory, and some or all of them may be arranged as a distributed system or a server-client structure.

[0050] The system (100) according to the present invention includes, for example, the following components to perform the entire process from receiving simulation result data to generating an analysis and report. First, the data collection and preprocessing unit (200) receives a simulation log file and performs preprocessing functions such as time alignment, missing value processing, event classification, and quantitative parameter calculation. Second, the rule-based analysis module (300) performs quantitative analysis by function based on the preprocessed data, including a maneuver analysis unit, a detection analysis unit, an engagement analysis unit, and a damage assessment analysis unit. Third, the RAG-based information retrieval and context reinforcement unit (400) searches for external reference knowledge, such as military doctrine, rules of engagement, and sensor models, through a vector index to provide contextual information necessary for LLM input. Fourth, the local LLM-based comprehensive analysis unit (500) generates a draft of a military analysis report by integrating the analysis results and the search context through a large language model executed in a closed network environment. Fifth, the report generation unit (600) structures the output text of the comprehensive analysis unit according to a report template and automatically generates a final analysis report. Sixth, the control unit (700) manages the execution order of each component and controls the input / output data flow and the overall processing procedure. The above components may be implemented in the form of software modules on a computing device including one or more processors and memory, and some or all of them may be distributed as independent servers or nodes. Depending on the security characteristics of the defense sector, the system (100) is based on the premise of being installed and operated in a closed network environment separated from the external network or on standalone equipment.

[0051] FIG. 2 is a flowchart showing the overall procedure in which a data collection and preprocessing unit (200) according to the present invention receives log data generated from a defense M&S simulation and refines it into an analyzable standard format.

[0052] The preprocessing unit (200) is composed of a data input unit (210), a preprocessing engine (220), and a meta-tagging unit (230), each of which can be implemented as a software module in a computing device that includes one or more processors.

[0053] First, in the simulation log input step (S210), the data input unit (210) receives the original log file generated from the simulation engine. The input log may include, for example, the following data fields. First, motion information (Log_Motion): time (timestamp), platform ID, coordinates (x, y, z), speed, course, etc.; second, detection / tracking information (Log_Detection): sensor ID, detection status (flag), detection distance, SNR (Signal-to-Noise Ratio), Pd, Pfa, etc.; third, engagement event (Log_Engagement): weapon firing time, target ID, interception status (Kill / Miss), hit location, flight time, etc. Fourth, damage assessment (Log_Damage): Since the log structure differs in various M&S simulation engines, such as damage level, whether major equipment is damaged, and judgment of destruction (sinking), the input unit (210) supports multiple formats such as JSON, CSV, and Binary Log, and may include an automatic format recognition function or a conversion rule table. The input log is stored in an internal buffer and is sorted by time in a subsequent step.

[0054] Next, in the time alignment and missing value handling step (S220), the preprocessing engine (220) aligns all events based on a common reference time (Time 0) or the simulation start time to ensure temporal consistency of the logs. The preprocessing engine performs the following functions: (1) Time alignment converts log timestamps into a common timeline by applying a Δt (offset) correction value when different references are used (e.g., time delay per sensor). The aligned result is managed in the form of a Time-Ordered Event Sequence. (2) Missing Value Handling: The following missing value handling is performed: I) For missing quantitative data such as coordinates / velocities, linear interpolation or forward fill methods are applied; 2) For missing detection flags, a sensor model-based default value (Default = False) can be applied; 3) For missing firing event information, reconstruction is performed using other sensor / engagement logs near the weapon firing time.

[0055] The preprocessing engine can separately store logs regarding the missing value processing results to be used for explainability and future verification. Next, in the event classification step (S230), the preprocessing engine (220) refers to an Event Rule Set to classify simulation logs by function. Events can be classified into events such as motion, detection, engagement, and damage assessment, and the rule set can be configured as follows. Motion Events occur when the position or speed of the platform changes by more than Δ standard; Detection Events occur when the sensor flag switches to True or a tracking ID is assigned; Engagement Events occur when a firing command is issued, an intercept trajectory is generated, or a hit is determined; and Damage Events occur when the damage level increases or the status of major equipment changes to “Degraded” or “Failed”.

[0056] Rule-based classification can be performed using conditions defined in the tactical model (Doctrine) or the rule of engagement (ROE), and this is utilized for context reinforcement in the subsequent RAG search unit (400). Next, in the quantitative parameter calculation step (S240), the preprocessing engine (220) automatically calculates key quantitative parameters required for M&S analysis based on the classified events. These are structured so that they can be immediately utilized in the maneuver analysis unit, detection analysis unit, and engagement analysis unit. In the embodiment, the following parameters are calculated. (1) The maneuver parameters may include distance traveled = Σ|pos(t) - pos(t-1)|, average speed = total distance traveled / total elapsed time, whether evasive maneuvering is performed = course change amount > θ_threshold, etc. (2) The detection parameters may include time-to-detection (TTD), detection range = |platform position - target position|, detection performance = function (SNR, Pd, Pfa), etc. (3) Engagement parameters may include the time to start engagement (TEI), whether the interception was successful (Kill / Miss), Time-To-Kill (TTK) = KillTime - FireTime, etc. (4) Damage assessment parameters may include the amount of change in damage level, whether critical damage occurred, and the rate of loss of function (e.g., sensor / weapon / propulsion series). In the embodiment, parameter values ​​may be entered as TTD=12.4s, TTK=31.7s, and detection range=18.1km. The calculated quantitative parameters are stored in a JSON or Key-Value structure and are transmitted as input to the subsequent rule-based analysis module (300) and LLM comprehensive analysis unit (500).

[0057] Next, in the meta-tagging step (S250), the meta-tagging unit (230) assigns identification tags and metadata for analysis to the preprocessed data. This is an essential step for ensuring contextual connectivity when configuring RAG search and LLM input. Examples of the assigned metadata are as follows: (1) Platform metadata may include Blue / Red distinction, platform type (DDG, FFG, UAV, etc.), sensor / weapon configuration, etc. (2) Scenario metadata may include terrain characteristics, weather conditions, target characteristics, etc. (3) Event metadata may include structured information such as {“event_type”: “detection”, “time”: 125.4, “platform_id”: A1, “target_id”: B3}. (4) Doctrine tags may include detection failure conditions, engagement initiation conditions, ROE violations, etc. The aforementioned meta-tagged data is used as part of the “query” during subsequent RAG searches, enabling explainability of evidence-based analysis results during the LLM inference process.

[0058] Next, in the step of generating a standard structure for analysis (S260), the preprocessing unit finally generates a standardized analysis data structure that includes all maneuver / detection / engagement / damage assessment data. The structure of the example is as follows.

[0059] {

[0060] "scenario_info": {...},

[0061] "motion_data": [...],

[0062] "detection_events": [...],

[0063] "engagement_events": [...],

[0064] "damage_assessment": [...],

[0065] "quantitative_parameters": {...},

[0066] "meta_tags": {...}

[0067] }

[0068] The structure of the above embodiment serves as an interface commonly used in the rule-based analysis module (300), the RAG-based information retrieval unit (400), and the LLM comprehensive analysis unit (500).

[0069] Next, in the output and error log recording step (S270), the preprocessing engine (220) records errors or data outliers that occurred during the processing to support explainability and reproducibility verification. Examples of errors include time discrepancies, abnormal sudden changes in coordinate values, missing sensor events, and input file format errors. The error information is also transmitted to the audit log of the control unit (700) and used for verification by analysts or managers.

[0070] FIG. 3 is a flowchart illustrating a procedure in which a rule-based analysis module (300) according to the present invention receives preprocessed M&S simulation result log data and automatically performs a quantitative analysis by function on one or more simulation result analysis items, such as a maneuver analysis unit (310), a detection analysis unit (320), an engagement analysis unit (330), and a damage assessment analysis unit (340), which are set according to the purpose of the scenario or the analysis task.

[0071] The above rule-based analysis module (300) is a key component that quantitatively and structurally interprets simulation results and provides them for subsequent analysis in a stage prior to the local LLM-based comprehensive analysis being performed within the system (100) of the present invention.

[0072] The rule-based analysis module (300) includes an analysis rule set, a formula-based parameter calculator, an event inference engine, etc., and all of these components can be implemented as software modules that run on one or more processors. The rule-based analysis module (300) may be configured to perform one or more simulated result analyses, such as maneuver analysis, detection performance analysis, engagement analysis, and damage assessment analysis, depending on the characteristics of the input log data and the purpose of the analysis, but is not limited thereto.

[0073] In one embodiment, the rule-based analysis module (300) can perform quantitative analysis by function, including a maneuver analysis unit (310), a detection analysis unit (320), an engagement analysis unit (330), and a damage assessment analysis unit (340).

[0074] First, in the motion analysis step (S310), the motion analysis unit (310) receives the preprocessed motion log (motion_data) and quantitatively calculates the trajectory, speed, and motion pattern of the platform and weapon. (1) To analyze the motion trajectory, the total distance traveled is calculated by integrating the position coordinates (x, y, z) of the platform and weapon in chronological order. This can be expressed as <Equation 1>.

[0075]

[0076] (2) To analyze speed and course, the average speed, maximum speed, and rate of change of speed (dv / dt) are calculated. This determines whether the change in course meets the criteria defined in the rules of engagement (ROE or manual). For example, the attack maneuver is determined by satisfying the condition (course converges toward the target direction). The definition of course convergence toward the target direction is that it satisfies the criteria set in CPA and TCPA. (3) To analyze relative maneuver, the approach speed and relative distance (time series) between enemy and friendly platforms and weapons are calculated, and the maneuver pattern immediately before detection and engagement is automatically tagged. As an example, maneuver patterns such as "Direct Approach" and "Avoidance Turn" are defined.

[0077] In the relative maneuver analysis phase, the maneuver characteristics of friendly platforms and weapons can be automatically classified as "Direct Approach" or "Avoidance Turn" by utilizing changes in relative position, bearing, and speed with respect to the target.

[0078] For example, whether a direct approach is established can be determined based on whether the azimuth error of the friendly platform's course relative to the target direction remains below a preset threshold angle for a certain period of time, and whether the relative approach speed toward the target within the same period exceeds a threshold value. In this embodiment, if both of the above conditions are satisfied by a predetermined ratio or higher, the maneuver is determined to be a Direct Approach pattern.

[0079] According to an embodiment, whether an evasive maneuver is performed can be determined based on whether the course of the friendly platform changes by more than a predetermined threshold angle during the analysis time interval, whether the lateral movement component relative to the target increases by more than a preset ratio due to the change in course, and whether, as a result, the relative distance to the target increases or no longer decreases. If all of these conditions are satisfied, the present invention classifies the maneuver as an avoidance turn.

[0080] The simplified judgment rules described above can be adjusted according to various simulation environments and platform combinations, and the relative maneuver analysis unit generates these judgment results as tag information within the rule-based analysis module of the present invention and provides them as input data utilized in subsequent detection analysis, engagement analysis, and LLM-based comprehensive analysis units.

[0081] Next, in the detection analysis step (S320), the detection analysis unit (320) quantitatively calculates the detection performance, initial detection time, and tracking maintenance status for each sensor using the sensor detection log and the activation log. (1) The initial detection time (TTD is Time-to-Detect) can be calculated by measuring the first time point t_detect when the detection flag switches from False to True. The initial detection time can be defined as shown in the following <Equation 2>.

[0082]

[0083] Here, t_appearance refers to the time when the target first appears in the simulation space.

[0084] (2) The distance at detection can be calculated by determining the distance d_detect between the platform and the target at the time of detection. The value of d_detect can be classified into “normal detection,” “delayed detection,” and “detection failure (Over Range)” by determining whether it is within the performance range of the sensor model (or depending on whether it is included within the performance range). (3) The detection performance calculation, i.e., quantitative evaluation, calculates the detection quality according to the sensor performance model S = f(SNR, Pd, Pfa). For example, when the signal-to-noise ratio (SNR) is above a preset threshold (SNR_threshold), the detection success rate can be evaluated as increased, and reliability calculation based on the detection probability (Pd) and tracking stability evaluation based on whether the tracking identifier (ID) is maintained for a set time interval Δt can be performed. If the above tracking identifier (ID) is not maintained for Δt, it may be determined as “tracking instability.” (4) For the determination of false alarms and missed detections, if the sensor indicates “detection” but there is no actual target, it is determined as a false alarm, and if there is a target but detection does not occur, it is determined as a missed detection. The results of the detection analysis above may be used as input for a subsequent engagement analysis step (e.g., determination of conditions for initiating engagement).

[0085] Next, in the engagement analysis stage (S330), the engagement analysis unit (330) automatically determines engagement performance by comprehensively analyzing launch events, interception paths, hit events, and failure events. First, regarding the determination of the start of engagement, the engagement analysis unit (330) determines the conditions for the start of engagement according to tactical rules or rules of engagement (ROE). Specifically, it is determined that engagement has started if at least one of the following is satisfied: when the distance between the platform and the target becomes less than or equal to the set engagement range after the enemy platform is detected, when an engagement approval time occurs according to command and control rules, or when a fire command is output from the fire control system (FCS). Next, the determination of whether the interception was successful (kill / miss) is performed by tracking the flight trajectory of the missile or projectile. The engagement analysis unit (330) calculates the minimum distance to the target during the armed flight and determines that the interception is successful (Kill) if the minimum distance is less than or equal to the lethal radius (Kill_Radius). At this time, the time-to-kill (TTK) is calculated by recording the event timestamp (KillTime) at the time of successful interception and calculating the time difference from the time of fire (FireTime), which can be defined as the following mathematical formula.

[0086]

[0087] If an engagement is determined to be a miss, the engagement analysis unit (330) subdivides and analyzes the type of engagement miss. Specifically, the engagement miss may be classified into at least one of the following: insufficient engagement range due to detection delay; impossible guidance due to tracking instability of the fire control system (FCS); timing error due to delay in firing time; the target performing maneuver evasion; or weapon malfunction or abnormal operation. Additionally, the engagement analysis unit (330) can perform multi-engagement analysis. For example, when multiple weapons are projected sequentially or simultaneously onto a single target, the engagement performance of each weapon is compared and analyzed, and the effectiveness of the engagement in a multi-engagement situation is evaluated by calculating whether a surviving threat exists even after the engagement ends. The results of the engagement analysis are used as key basis information to generate "engagement result summaries" and "operational implications" in the subsequent Local Large Language Model (LLM)-based comprehensive analysis stage.

[0088] Next, in the Damage Assessment Analysis step (S340), the Damage Assessment Analysis Unit (340) quantitatively analyzes the damage caused by a hit or near-miss projectile in conjunction with a platform-specific performance degradation model. First, the damage level assessment is automatically performed based on the change values ​​in the damage event log. The damage level can be defined, for example, as follows. If no damage occurs at all, it is defined as Level 0 (no damage); if limited functional degradation occurs, such as the performance degradation of one or more sensors, it can be defined as Level 1 (minor damage). Additionally, if part of the armament becomes incapacitated or part of the propulsion system's function is lost, it can be defined as Level 2 (moderate damage); and if the platform's combat capability is lost or destroyed, it can be defined as Level 3 (critical damage). The Damage Assessment Analysis Unit (340) analyzes the state change values ​​in the damage event log and automatically calculates the damage level according to the above definitions. Next, the Damage Assessment Analysis Unit (340) performs an impact analysis for each component. For example, if the radar is damaged, the detection range may be assessed as reduced, and if the fire control system (FCS) is damaged, the probability of engagement failure may be assessed as increased. Additionally, if the engine is damaged, the mobility of the platform may be assessed as reduced. These component-specific impacts can be calculated by referring to the function model corresponding to each component. Furthermore, the damage assessment analysis unit (340) performs an analysis of combat persistence. Specifically, it can determine whether the platform can continue to perform missions based on the increase in the calculated damage level, and for example, if the damage level is Level 2 or higher, it may be displayed as being in a “combat persistence” state.Furthermore, the damage assessment analysis unit (340) can calculate an indicator indicating the possibility of sustaining combat or winning or losing combat by comparing and analyzing the remaining combat power of the enemy and friendly forces.

[0089] Next, in the analysis result structuring and output step (S350), the rule-based analysis module (300) organizes the analysis results into a format such as JSON / XML as an example of a standard structure and transmits it to a subsequent module.

[0090] {

[0091] "motion_analysis": {...},

[0092] "detection_analysis": {...},

[0093] "engagement_analysis": {...},

[0094] "damage_assessment": {...},

[0095] "derived_parameters": {

[0096] "TTD": "...",

[0097] "TTK": "...",

[0098] "engagement_success_rate": "...",

[0099] "damage_level": "..."

[0100] }

[0101] }

[0102] In addition, errors and outliers that occur during the analysis process are recorded in the audit log of the control unit (700) and used to verify reproducibility or ensure explainability.

[0103] FIG. 4 is a flowchart illustrating the overall procedure in which a RAG-based information retrieval and context reinforcement unit (400) according to the present invention constructs reference knowledge such as military doctrines, rules of engagement, sensor model documents, and simulation logic analysis documents required for defense M&S analysis in the form of a vector index, searches for paragraphs related to simulation analysis results based on similarity, and provides them as LLM input contexts.

[0104] The RAG-based information retrieval and context augmentation unit (400) is composed of an embedding engine (410), a vector index (420), a query generator (430), a similarity retriever (440), and a context augmentation unit (450), and each component can be implemented in the form of a software module on an information processing device including one or more processors and memory.

[0105] First, in the reference document subdivision and embedding vector generation step (S410), the embedding engine (410) receives external reference documents required for defense simulation analysis, subdivides them into paragraphs or clauses (Chunking), and generates an embedding vector for each text chunk. Examples of reference documents may include (1) military doctrine and tactical manuals (Doctrine, TTP), (2) rules of engagement (ROE) and fire control procedures, (3) sensor detection model and tracking algorithm descriptions, (4) weapon system flight model and hit judgment logic documents, (5) simulation logic analysis reports, experimental plans (OA / OA-LEX), etc.

[0106] The embedding engine applies unsupervised embedding models (e.g., SBERT family, LLM embedding models within the defense network, etc.), and each text chunk can be stored with the following metadata.

[0107] { "doc_id": "...",

[0108] "chunk_id": "...",

[0109] "text": "...",

[0110] "embedding": [ ... ],

[0111] "tags": ["engagement_rule", "sensor_model", ...]

[0112] }

[0113] Next, in the vector index construction and storage step (S420), the vector index (420) is constructed using an approximate nearest neighbor search (ANN) based data structure such as FAISS, HNSW, or DiskANN so that the embedding vectors can be efficiently searched in a high-dimensional space. Due to the characteristics of the defense environment of the present invention, the closed network index is stored independently within a local storage device or security module and is operated without being connected to an external network. The vector index stores paragraph information in chunk units, and each vector is managed together with identification information such as (1) chunk_vector[i], (2) chunk_metadata[i], (3) document source (doc_id), relevant clause number (section_id), and tactical tag (doctrine_tag).

[0114] Next, in the search query generation step (S430), the search query generation unit (430) automatically generates a search vector query based on the quantitative analysis results transmitted from the rule-based analysis module (300) and the metadata generated by the preprocessing unit (200). Here, the quantitative analysis results may include the time to first detection (TTD), time to intercept (TTK), detection range, and the cause of engagement success or failure, and the metadata may include the platform type, sensor type, and target type. In this embodiment, the search query generation unit (430) can generate search queries with multiple configurations. For example, for a search query regarding maneuver analysis, “tactical manual content related to evasive maneuver judgment criteria” may be generated, and for a search query regarding detection analysis, “allowable range of signal-to-noise ratio (SNR) based detection probability (Pd) and false alarm probability (Pfa) and detection performance rules” may be generated. Additionally, search queries regarding engagement analysis may generate “paragraphs related to interception failure types (e.g., tracking instability, firing angle limitations, etc.),” and search queries regarding damage assessment may generate “regulations on combat sustainability degradation based on damage levels.” The above search queries can be constructed by combining natural language sentences, keyword vectors, or JSON-formatted data containing analysis results, and the generated search queries are converted into query embeddings through an embedding engine.

[0115] Next, in the similarity-based paragraph search step (S440), the similarity search unit (440) compares the query vector with the chunk vector stored in the vector index and returns the most relevant paragraphs in a Top-k manner. Examples of similarity evaluation may be performed in the form of (1) cosine similarity, (2) proximity based on L2 distance, (3) dot product-based score calculation, etc.

[0116] In an embodiment of the present invention, to increase the reliability of search results, weights may be assigned according to document types (doctrines / sensors / rules of engagement) or filtering may be performed based on specific scenario metadata (e.g., platform type, threat type), and the searched chunks are returned in the following form.

[0117] { "retrieved_chunks":

[0118] [

[0119] { "text": "...", "similarity": 0.92, "doc_id": "...", "tags": [...]},

[0120] ...

[0121] ]

[0122] }

[0123] Next, in the context augmentation step (S450), the context augmentation unit (450) reconstructs the similarity search results into a prompt format to match the input format of the LLM comprehensive analysis unit (500).

[0124] As an example, an LLM input context can be created with the following structure.

[0125] [Scenario Summary]

[0126] ...

[0128] [Quantitative Analysis Results]

[0129] TTD = 12.4 s

[0130] TTK = 32.7 s

[0131] Detection Range = 18.2 km

[0132] ...

[0134] [Retrieved Military Doctrine / Rules]

[0135] - (Doctrinal Document A, Section 3.2) Sensor detection criteria: ...

[0136] - (Rules of Engagement B, Section 5.4) Criteria for Determining Failure of Engagement When Tracking Instability Occurs: ...

[0137] ...

[0138] Additionally, if necessary, the source (doc_id, section_id) of each searched paragraph can be included in the prompt to configure the LLM to generate an analysis and judgment result with explainability. The contextual information configured in this step is transmitted to the LLM-based comprehensive analysis unit (500) to enable the combination of qualitative judgment and quantitative indicators when generating a military analysis report.

[0139] FIG. 5 is a flowchart illustrating the overall procedure in which a local LLM-based comprehensive analysis unit (500) according to the present invention integrates quantitative analysis results and document search results input from a preprocessing unit (200), a rule-based analysis module (300), and a RAG-based information search and context reinforcement unit (400) to automatically generate a draft of a military analysis report.

[0140] The comprehensive analysis unit (500) is composed of a system prompt generator (510), an LLM input configurator (520), a local LLM inferer (530), and an LLM output post-processor (540), and each component operates as a software module in an information processing device implemented with one or more processors.

[0141] First, in the System Prompt generation step (S510), the System Prompt generator (510) generates control instructions, such as output format, military sentence style, and analysis scope, that the LLM must comply with to suit the analysis purpose of the system of the present invention. As an example, the System Prompt may be configured to include elements such as (1) report structure (mobility analysis → detection analysis → engagement analysis → damage assessment → comprehensive judgment), (2) instructions to apply military modified narrative style, (3) to reflect the RAG search context first, (4) to use numerical and quantitative values ​​as they are without distortion, and (5) to reflect scenario metadata (platform type, operational environment, etc.).

[0142] To ensure consistency and quality, control statements containing multiple system prompts can be generated through a system prompt generator (510). First, as a directive regarding Persona assignment, the local large language model may be designated to perform the role of a defense modeling and simulation (M&S) expert analyst. Specifically, the local large language model may be configured to perform the mission of writing accurate and consistent military analysis reports based on provided defense M&S log analysis results and military doctrine contexts retrieved via the Search Augmentation Generation (RAG) method, acting as a defense analysis expert who performs analysis of tactics, engagement, sensors, and weapon systems. Next, as a directive regarding Task, to clarify the purpose of the output that the local large language model must generate, the local large language model may be designated to write "Analysis and Comprehensive Evaluation of Defense M&S Simulation Results Local LLM-Based Comprehensive Analysis Unit (500), based on the inference results of the local large language model (LLM)," composed of military bulleted text by synthesizing the results of maneuver, detection, engagement, and damage assessment. In addition, as instructions regarding data utilization rules, the method of utilizing quantitative analysis results calculated by the rule-based analysis module (300) and the RAG-based search context can be clearly defined. For example, quantitative values ​​such as Time to Detect (TTD), Time to Kill (TTK), detection range, speed, course, altitude, and damage level calculated during the maneuver, detection, engagement, and damage assessment phases may be designated to be used as is without modification. Additionally, content cited in RAG-based search documents, including military doctrine, rules of engagement, and sensor models, may be designated to be indicated as “(Reference Material)” at the end of the relevant sentence, and all analysis judgments may be restricted to be performed strictly based on the quantitative analysis results and the RAG search context.

[0143] In addition, as instructions regarding object roles and terminology constraints, rules for the use of object names and terms may be established to prevent interpretation errors. For example, enemy weapons and targets may be designated to be labeled as "KN-23," "BM," "enemy aircraft," or "enemy guided missile," and friendly weapons and sensors may be designated to be labeled as "SM-6," "friendly interceptor guided missile," or "friendly air defense radar." Furthermore, platform names may be required to maintain a Blue / Red distinction, arbitrary interpretation or renaming is prohibited, and quantitative values ​​may be restricted from being inferred or altered without basis. Finally, as instructions regarding output formatting, the configuration of the output section may be enforced to match the report template of the report generation unit (600). Specifically, the report may be specified to be organized in the order of "[1] Simulation Overview", "[2] Maneuver Analysis Results", "[3] Detection Effect Analysis", "[4] Engagement Results Analysis", "[5] Damage Assessment", and "[6] Comprehensive Judgment and Operational Implications", and each section may be written in the form of military bulleted sentences and may be restricted to be structured so that quantitative analysis values ​​and RAG-based evidence contexts are naturally connected.

[0144] Next, in the LLM Input Construction step (S520), the LLM input constructor (520) configures the input information to be provided to the local large language model (LLM) into the following three input elements and combines them into a single integrated prompt block.

[0145] First, the Quantitative Analysis Block includes the quantitative analysis results calculated by the rule-based analysis module (300). Specifically, the Quantitative Analysis Block may include detection performance based on the Time to Detect (TTD), detection range, and signal-to-noise ratio (SNR); may include the time of launch, whether the interception was successful or unsuccessful, and the Time-to-Kill (TTK); and may include whether evasive maneuvers were performed, the amount of course change, and the relative approach speed. In addition, information regarding the damage level and the functional impact of each component may be included. An example of the Quantitative Analysis Block in this embodiment may be as follows.

[0146] [Quantitative Results]

[0147] TTD = 12.4 sec

[0148] TTK = 31.7 sec

[0149] Detection Range = 18.1 km

[0150] Relative Closure Rate = +21 m / s (Approach)

[0151] Damage Level = 2 (Medium Damage)

[0152] Next, the RAG Retrieved Context Block contains paragraphs regarding military doctrines, rules of engagement, and sensor models retrieved through the RAG information retrieval unit (400). For example, the RAG Retrieved Context Block may include parts of military doctrine or rule documents such as the following.

[0153] [Retrieved Doctrine / Rules]

[0154] Doctrinal Document A, Section 3.2: “Sensor detection performance is calculated based on SNR and Pd.”

[0155] Rules of Engagement B, Section 5.4: “Increased probability of interception failure when tracking instability occurs”

[0156] In addition, the Scenario Meta Block includes metadata related to the simulation scenario. Specifically, the Scenario Meta Block may include information on the configuration of friendly and enemy platforms, terrain and environmental information, and scenario objectives and limitations. Examples of the Scenario Meta Block in this embodiment may be as follows.

[0157] [Scenario]

[0158] Ally: DDG-II class

[0159] Enemy: MIG-21

[0160] Environment: Standard sea conditions, good visibility

[0161] The LLM input configurator (520) combines the quantitative analysis result block, the RAG search context block, and the scenario / metadata block to form an integrated input statement that enables the local large language model to perform analysis without misjudgment.

[0162] Next, in the Local LLM Inference step (S530), the local LLM inferer (530) calls an LLM model running in a closed network environment to generate analysis text for an integrated input prompt. A key characteristic of this step is that it operates on local equipment (GPU or CPU-based) without an internet connection, does not transmit any data externally to ensure security, outputs an analysis result in the form of "Activation-Detection-Engagement-Damage-Comprehensive Judgment" based on the RAG search context and quantitative analysis values, and the model strictly follows predefined system prompt rules.

[0163] The LLM output has a structure as follows, for example.

[0164] [Operational Overview]

[0165] ...

[0166] [Analysis of Mobility Characteristics]

[0167] ...

[0168] [Analysis of Detection Effectiveness]

[0169] ...

[0170] [Analysis of Engagement Results]

[0171] ...

[0172] [Damage Assessment]

[0173] ...

[0174] [Operational Implications]

[0175] ...

[0176] Next, in the output post-processing step (S540), the LLM output post-processor (540) refines the analysis text generated by the LLM to match the structure required by the final report generation unit (600).

[0177] The LLM output post-processor (540) refines the analysis text generated by the LLM to match the structure required by the final report generation unit (600).

[0178] The main functions of post-processing include removing unnecessary phrases, ambiguous inferences, and duplicate content; checking for potential distortion or misunderstanding of quantitative analysis values; rearranging data to match the section order of the report template; and prioritizing the RAG basis statement in cases where RAG and LLM judgments conflict.

[0179] In addition, during the post-processing process, the reliability of the LLM output (e.g., a self-consistency-based score) can be calculated and transmitted to the control unit (700) and the report generation unit (600) in the following format.

[0180] { "sections": {...},

[0181] "confidence": 0.87,

[0182] "used_references": ["docA_3-2", "ROE_5-4"]

[0183] }

[0184] FIG. 6 is a configuration diagram illustrating the overall procedure in which a report generation unit (600) according to the present invention receives section-by-section analysis text, quantitative analysis data, RAG search basis text, and scenario metadata output from a local LLM-based comprehensive analysis unit (500) and automatically generates a standardized final report in accordance with the format of a defense M&S analysis report.

[0185] The report generation unit (600) consists of a section mapping unit (610), a basis / reliability organizer (620), a strategy analysis / generator (630), and a report assembly and output unit (640).

[0186] First, in the section mapping and structuring step (S610), the section mapper (610) automatically classifies and sorts the analysis results generated by the local LLM-based comprehensive analysis unit (500) according to the standard structure of the military analysis report. The basic structure of the military analysis report used in the present invention may consist of the following six sections: (1) a simulation overview, (2) maneuver analysis results, (3) detection effect analysis, (4) engagement result analysis, (5) damage assessment analysis, and (6) comprehensive judgment and operational implications. The section mapper (610) analyzes the paragraph composition and keyword patterns included in the output results of the local large language model (LLM) and automatically assigns text corresponding to each of the six sections. Additionally, if text corresponding to a specific section is missing, the section mapper (610) transmits an error signal to the control unit (700) to request post-processing of the local large language model or regeneration of the analysis results.

[0187] Next, in the evidence and reliability organization step (S620), the evidence and reliability organizer (620) organizes the RAG search evidence texts (doctrine, ROE, sensor model, etc.) among the LLM comprehensive analysis results in a citation format along with the original text, and enhances explainability by displaying the correlation with the quantitative data used in the internal inference of the LLM.

[0188] The functions performed in this step are (1) organizing RAG source text, (2) calculating a confidence score, and (3) generating a quantitative-qualitative mapping sheet. Examples of each function are as follows. First, regarding the function of organizing RAG source text, sentences marked as “(reference material)” among the analysis text output from the local LLM-based comprehensive analysis unit (500) are automatically extracted, or paragraphs used based on similarity in the RAG context block are automatically identified and extracted. At this time, the document identifier (doc_id) and section identifier (section_id), which are citation source information corresponding to the extracted sentences or paragraphs, are listed together and organized in the form of a footnote or at the bottom of the generated analysis report. Next, regarding the function of calculating a confidence score, a confidence index for each section of the report is calculated based on the confidence value (confidence) transmitted from the local LLM-based comprehensive analysis unit (500). For example, a reliability index of 0.92 may be assigned to the maneuver analysis section, a reliability index of 0.86 may be assigned to the detection analysis section, and a reliability index of 0.88 may be assigned to the engagement analysis section. Additionally, in the quantitative and qualitative mapping sheet generation function, the correspondence between quantitative data, including Time to Detect (TTD), Time to Kill (TTK), detection range, and relative velocity, and analysis sentences generated by the local LLM is organized in a table format. Through this, it is possible to clearly identify which quantitative data each analysis sentence was generated based on.

[0189] Next, in the step of automatically generating corrective measures and improvement plans (S630), the strategy analysis and generator (630) automatically derives military courses of action (CoA), alternative tactics, and equipment or sensor operation strategies based on analysis text generated by a local large language model (LLM) and quantitative analysis values ​​calculated by a rule-based analysis module. In this embodiment, the strategy analysis and generator (630) can generate corrective action items corresponding to various analysis results. For example, if a detection delay occurs, a corrective action item such as "need to adjust sensor elevation or supplement the low-altitude detection model" may be generated, and if an engagement failure occurs, a corrective action item such as "need to adjust Fire Control System (FCS) parameters to ensure tracking stability" may be generated. Additionally, if the damage level is intensified, a corrective action item such as "review measures to improve platform survivability" may be generated, and if the approach speed is excessive, a corrective action item such as "need to apply maneuver evasion procedures" may be generated. In addition, in this step, the reliability assigned to the local large language model output can be utilized to automatically assign a priority or recommendation grade to the generated Course of Action (CoA). Next, in the report assembly and output step (S640), the report assembly and output device (640) combines the section text, supporting documents, quantitative data, and corrective action items configured in the above steps to assemble them into a final defense modeling and simulation (M&S) analysis report.

[0190] In this embodiment, the report assembly and output device (640) can perform the following key functions. First, based on a standard military report template, it can automatically assemble the report according to paragraph composition, title system, and numbering rules, unify sentence connections and expressions between sections, and perform rearrangement of sections to maintain the logical flow of the report leading to maneuver, detection, engagement, damage assessment, and comprehensive judgment. Additionally, the report assembly and output device (640) can automatically generate and insert graphs based on quantitative data. For example, engagement timelines, detection range change graphs, and damage level change graphs can be automatically generated and included in the report, and analysis results generated based on JSON can be automatically converted into a tabular structure and inserted. Furthermore, the report assembly and output device (640) can output the report in a format suitable for a secure environment. Specifically, the output result can be generated in at least one of PDF, DOCX, or TXT formats used in a closed network environment separated from an external network, and transmission to an external network does not occur. In addition, the creation time of the report, input data, and version information of the local large language model can be automatically recorded in conjunction with the audit log steps (S610 to S650). The output analysis report is transmitted to an analyst or manager through the control unit (700), and an internal system interface may be provided to enable re-summarization or the regeneration of specific sections as needed.

[0191] FIG. 7 is a system flowchart illustrating a procedure in which a control unit (700) according to the present invention sequentially manages the flow of the entire process from simulation log input to report output, and integrally controls data transmission and reception and error processing between each component.

[0192] The control unit (700) is composed of a flow controller (710), a data routing manager (720), an error and exception handler (730), and a log and audit record manager (740), and can be implemented as a software module in a device including one or more processors and memory.

[0193] First, in the overall analysis flow initialization and module call step (S710), the flow controller (710) performs system-wide initialization when the analysis starts. Specifically, the flow controller (710) validates the validity of the input data, checks whether the format of the simulation log file is suitable, such as JSON or CSV, and verifies whether the log file is missing or if there is a lack of scenario metadata. Additionally, the flow controller (710) initializes the analysis pipeline to activate the preprocessing unit (200), performs initial setup of the rule-based analysis module (300), prepares the vector index of the RAG-based information retrieval unit (400), loads the model of the local LLM comprehensive analysis unit (500), and initializes the report template of the report generation unit (600). Once the above initialization is completed, the flow controller (710) transmits a control signal to proceed to the preprocessing step (S720).

[0194] Next, in the data flow routing and inter-module input transmission step (S720), the data routing manager (720) performs the role of transmitting the results generated in each processing step to the module of the next step. The transmission procedure in this embodiment is as follows. First, time-aligned log data, event classification results, and normalized parameters are transmitted from the preprocessing unit (200) to the rule-based analysis module (300). Next, analysis parameters including the Time to First Detect (TTD), Time to Intercept (TTK), detection range, maneuver pattern, and engagement results are transmitted from the rule-based analysis module (300) to the RAG-based information retrieval unit (400) to the query generation unit. Additionally, similarity-based top-k search paragraphs, military doctrine and rules of engagement (ROE) context, and sensor model paragraphs are transmitted from the RAG-based information retrieval unit (400) to the local LLM comprehensive analysis unit (500). Next, the LLM output text containing the analysis results by section, the confidence score, and the list of cited paragraphs are transmitted from the local LLM comprehensive analysis unit (500) to the report generation unit (600). The data routing manager (720) converts the data structure into JSON, vector, or text block form suitable for each module and transmits it.

[0195] Next, in the error and exception handling step (S730), the error and exception handler (730) is configured to automatically detect and recover system errors, data structure inconsistencies, and missing analyses that may occur during the entire processing process. Specifically, as a data inconsistency error handling step, if a missing field in a log file is detected, it may generate an automatic correction or warning, and if a calculation error occurs in the rule-based analysis result, it may perform a recalculation. Additionally, as an alternative procedure in case of RAG search failure, if the similarity search result is insufficient (k=0 or the similarity score is below a threshold), it may insert a basic description from an alternative document, such as a basic doctrinal, and the control unit may instruct a re-search request or an expansion of the k value. Furthermore, it may automatically check for missing sections through LLM output quality verification, perform a re-request for the local LLM if quantitative value distortion occurs, and perform a re-analysis request if the confidence score is below a threshold value. In addition, if a report configuration error occurs, the template inconsistency can be automatically corrected, and if a graph or table insertion fails, alternative text can be configured to handle the error, and the result of such error handling is recorded in the log system (740) of the control unit.

[0196] Next, in the log and audit record generation step (S740), the log and audit record manager (740) records all inputs, outputs, model versions, and analysis results in the analysis process in an auditable form. Specifically, as processing step logs, the start and end times of execution for each module, the size and identifier of the input data, the local LLM model version, the embedding model version, and the RAG index version may be recorded. Additionally, as analysis result trace (log trace) information, quantitative analysis values ​​for activation, detection, engagement, and damage assessment, a list of searched paragraphs and document sources, LLM output text, and confidence scores may be recorded. Furthermore, as security and integrity records, authentication information of closed network equipment, hash-based data integrity verification values, and seed and option information to ensure the reproducibility of the same analysis may be recorded. These audit records may be utilized for re-analysis, review by other analysts, or verification of research reports.

[0197] Finally, in the termination and result delivery step (S850), the control unit (700) terminates the entire analysis procedure when the final report output from the report generation unit (600) is successfully completed, and delivers the generated result file to an analyst or manager via a user interface (UI) or an internal file system. Additionally, the control unit (700) may perform subsequent tasks as needed, such as requesting a follow-up analysis, generating additional summaries, calling a re-analysis function for a specific section, or sharing results between nodes within a closed network.

[0198] FIG. 8 is a screen example illustrating an embodiment of a user interface (UI) (800) of a defense M&S simulation result automatic analysis and report generation system (100) according to the present invention.

[0199] The user interface (UI) (800) is designed to visually provide the user with analysis results generated by the preprocessing unit (200), rule-based analysis module (300), RAG-based information retrieval unit (400), local LLM-based comprehensive analysis unit (500), report generation unit (600), and control unit (700). The user interface (UI) (800) is composed of a result display area by clause (810), a basis and reliability display area (820), a corrective action proposal (or strategy analysis and generator) area (830), and a simulation result summary and report output area (840).

[0200] The components of each user interface (UI) (800) are visually and functionally organized to intuitively convey the analysis information and conclusions based on the local large language model (LLM) to the user, and are configured to immediately interact with the automatic analysis flow by minimizing user interaction.

[0201] First, the result display area (810) for each clause is an area that displays quantitative analysis results generated by the preprocessing unit (200) and the rule-based analysis module (300) by dividing them into functional sections for maneuver, detection, engagement, and damage assessment. Examples of the information displayed include whether evasive maneuvers were performed, changes in relative speed, and course rotation amounts in the maneuver analysis section; the time to initial detection (TTD), detection range, and sensor model-based detection probability (Pd) and false alarm probability (Pfa) may be displayed in the detection analysis section. Additionally, the time of launch, whether a hit occurred (Kill / Miss), and time-to-kill (TTK) may be displayed in the engagement analysis section; and the damage assessment section may display the damage level (Level 0~3), component deterioration status, and remaining combat power. The quantitative parameters may be expressed in the form of numerical values, graphs, or timelines, and when a user selects a specific indicator, a detailed log including the event time, location, and sensor value may be provided in a popup form. In addition, sections linked to local LLM analysis statements can be highlighted via an “LLM interpretation reference” icon or color coding.

[0202] Next, the evidence and confidence display area (820) is an area that provides the user with the evidence text retrieved from the RAG-based information retrieval unit (400) and the confidence score of the local LLM output. In this area, a list of RAG document citations including the title, clause number, summary, and similarity score of the retrieved document may be displayed, and the location where the corresponding evidence text is used in the local LLM output sentence may also be displayed. For example, it may be displayed in a form such as “ROE 5.4 - Recommendation to cease engagement when tracking is unstable (Similarity 0.93).” Additionally, the LLM confidence index for each section and for the entire report unit may be provided as a numerical value in the range of 0 to 1 or 0 to 100, and if the confidence is below a threshold value, a warning icon may be displayed on the user interface (UI). Furthermore, a data-evidence mapping table that corresponds 1:1 between the quantitative values ​​included in the analysis results and the RAG evidence statements may be provided in a table format, and this area functions as a core user interface component to ensure explainability of the analysis results.

[0203] Next, the corrective action proposal area (830) is an area that displays operational implications, improvement plans, and corrective actions generated by the local LLM-based comprehensive analysis unit (500) on an item-by-item basis. Examples of items displayed in this area may include "need for sensor elevation and detection model calibration to improve detection delay (see: Sensor Model Section 3.2)," "need for adjustment of fire control system (FCS) parameters to improve tracking stability" to improve engagement failure, "need for application of evasive maneuver procedures and early engagement timing" as a damage reduction measure, and "resetting defense angles in response to increased enemy approach speed" to improve platform operation. Each item may be assigned an automatically generated priority or recommendation level (Level 1 to 3), and the user can select the item to view detailed analysis statements and related quantitative values.

[0204] Next, the simulation result summary and report output area (840) is an area that supports downloading the entire analysis result as a single report document or outputting it directly from a web-based or closed-network-based system. In this area, a summary card containing an overview of the simulation may be provided, and a scenario summary, a summary of key results of maneuver, detection, engagement, and damage assessment in 4 to 6 lines, and an overall assessment may be displayed. Additionally, through the report generation button, at least one of “PDF output,” “DOCX output,” or “TXT export” can be selected to output the template-based assembly result of the report generation unit (600) as is. Furthermore, through the audit log verification button, the analysis date and time, local LLM version, vector index version, and data hash value can be checked, and a replay analysis function based on the same conditions can be called.

[0205] The UI according to the present embodiment provides the following technical effects. It enhances explainability by visually explaining the combined results of LLM, RAG, and quantitative analysis; ensures reliability and ease of verification by transparently presenting the analysis flow step-by-step in the UI; improves operational analysis efficiency by enabling the immediate output of analysis results as standard template-based reports; and maintains security through a lightweight, modular UI structure capable of operating even in a closed-network defense environment.

[0206] The term "part" as used in this embodiment refers to a software or hardware component, such as an FPGA (field-programmable gate array) or ASIC, and the "part" performs certain roles. However, the meaning of "part" is not limited to software or hardware. The "part" may be configured to reside in an addressable storage medium or configured to run one or more processors. Thus, as an example, the "part" includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and "parts" may be combined into a smaller number of components and "parts" or further separated into additional components and "parts." In addition, the components and '~parts' may be implemented to play one or more CPUs within the device or secure multimedia card.

[0207] Although the present invention has been described above with reference to preferred embodiments, those skilled in the art will understand that various modifications and changes can be made to the invention without departing from the spirit and scope of the invention as described in the following claims.

[0208] FIG. 9 is a configuration diagram showing a military analysis report generation system according to another embodiment of the present invention.

[0209] As illustrated in FIG. 9, a military analysis report generation system (890) according to another embodiment of the present invention includes some or all of a defense M&S simulation device (900), a communication network (910), a military analysis report generation device (920), and a third-party device (930).

[0210] Here, "including some or all" means that some components, such as a third-party device (930), may be omitted to configure the military analysis report generation system (890) of FIG. 9, or that some or all components constituting the military analysis report generation device (920) may be integrated into a network device (e.g., wireless switching device, gateway, etc.) constituting the communication network (910), and is described as including all to facilitate a sufficient understanding of the invention.

[0211] A defense modeling and simulation (M&S) device (900) according to another embodiment of the present invention is a device that simulates a defense operational environment in a virtual space to perform a simulation and generates simulation result data from the process or result of the simulation. Modeling can be viewed as a process of defining a defense operational environment, weapon systems, sensors, platforms, rules of engagement, and tactical actions using mathematical formulas, logical rules, state transitions, or sets of parameters.

[0212] The defense M&S simulation device (900) may be configured to perform tactical and operational-level simulations including maneuver, detection, engagement, and damage assessment, and calculates the movement of the platform, the detection operation of the sensor, the firing and interception of weapons, and the process of damage occurring as a result of engagement over time. The defense M&S simulation device (900) may be implemented as a single simulation engine or may be implemented in a form in which multiple simulation modules are linked.

[0213] The defense M&S simulation device (900) can record events and state changes occurring during the simulation in the form of log data. The log data may include time information (timestamp), platform identifier, position, velocity, and course information, sensor detection status, launch and interception events, engagement results, damage status changes, etc., and such log data corresponds to simulation result data generated as a result of the simulation process or execution.

[0214] Additionally, the defense M&S simulation device (900) can perform simulations based on platform-to-platform interaction, sensor models, rules of engagement (ROE), tactical models (Doctrine), and weapon performance models. For example, whether a sensor detects something can be determined based on the signal-to-noise ratio (SNR), detection probability (Pd), and false alarm probability (Pfa), the initiation of engagement can be determined based on set rules of engagement or command and control conditions, and whether an interception is successful can be determined based on kill radius conditions.

[0215] The simulation result data generated by the defense M&S simulation device (900) is transmitted to the military analysis report generation device (920) via the communication network (910). At this time, the simulation result data may be provided in JSON, CSV, or a structured data format equivalent thereto, and is configured so that the preprocessing unit of the military analysis report generation device (920) can perform time alignment, event classification, and quantitative parameter calculation.

[0216] Meanwhile, the defense M&S simulation device (900) may be operated in a closed network environment separated from an external network, and may be operated within the same physical equipment or the same security domain as the military analysis report generating device (920). Additionally, according to an embodiment, some functions of the defense M&S simulation device (900) may be implemented by being logically or physically integrated with the military analysis report generating device (920) or a network device constituting the communication network (910).

[0217] The communication network (910) is a communication infrastructure for transmitting simulation result data, analysis results, and report information between a defense modeling and simulation (M&S) simulation device (900), a military analysis report generation device (920), and a third-party device (930) as needed.

[0218] The communication network (910) is configured to transmit simulation result data regarding the simulation execution process or execution result generated by the defense M&S simulation device (900) to the military analysis report generation device (920). The simulation result data may include event logs related to maneuver, detection, engagement, and damage assessment, state change information, and raw data required for calculating quantitative parameters, and may be transmitted in JSON, CSV, or a structured data format equivalent thereto.

[0219] The communication network (910) may be implemented as a wired communication network, a wireless communication network, or a combination thereof, and may be configured as a local network (LAN), a dedicated network, or a defense-dedicated communication network operating in a closed network environment according to the embodiment. In particular, in an embodiment of the present invention, the communication network (910) may be configured to support a closed network environment that is logically or physically separated from an external network, and the military analysis report generation system may be operated without a direct connection to the external internet.

[0220] Additionally, the communication network (910) can perform data packet ordering, error detection and retransmission, and verification of data corruption during transmission to ensure the reliability and integrity of data transmission. For example, the communication network (910) can transmit a timestamp, data identifier, or hash value together with the data during transmission to enable the military analysis report generating device (920) to verify the integrity of the received data.

[0221] Meanwhile, the communication network (910) may be implemented by logically or physically integrating with some functions of the military analysis report generating device (920). For example, a network device, gateway, or data relay device constituting the communication network (910) may be included inside the military analysis report generating device (920) or implemented by being integrated with the communication interface part of the military analysis report generating device (920).

[0222] Additionally, according to the embodiment, the communication network (910) may be configured to exchange control signals in both directions, such as the status of analysis progress, error information, or retransmission requests, as well as to transmit data from the defense M&S simulation device (900) to the military analysis report generation device (920). Through this, the military analysis report generation system (890) can stably link the entire process of simulation execution, analysis processing, and report generation.

[0223] The military analysis report generation device (920) is a core device that automatically analyzes simulation result data transmitted from the defense modeling and simulation (M&S) simulation device (900) and generates a military analysis report based on the analysis results.

[0224] The military analysis report generating device (920) is configured to perform quantitative and qualitative analyses of maneuver, detection, engagement, and damage assessment in stages, using simulation result data generated as a result of the simulation process or execution received through the communication network (910) as input. The simulation result data may be provided in the form of simulation log data and may include time information, platform status information, sensor detection information, engagement event information, and damage status change information.

[0225] The military analysis report generation device (920) may be configured to include a preprocessing unit, a rule-based analysis module, a RAG-based information retrieval unit, a local LLM-based comprehensive analysis unit, a report generation unit, and a control unit. The preprocessing unit normalizes the received simulation result data into a form suitable for analysis by chronologically aligning the data, classifying events, and processing missing values. The rule-based analysis module performs maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the preprocessed data, and calculates quantitative analysis results such as Time to Detect (TTD), Time to Kill (TTK), detection range, and damage level.

[0226] The RAG-based information retrieval unit searches for external reference knowledge documents, including military doctrines, rules of engagement (ROE), sensor models, and simulation logic analysis documents, using the Search Augmented Generation (RAG) method based on the above quantitative analysis results and metadata, and constructs a reference context to be provided to the local large language model. At this time, the reference knowledge documents are converted into embedding vectors and can be searched through vector indices stored in the storage unit.

[0227] The local LLM-based comprehensive analysis unit utilizes a local Large Language Model operating in a closed network environment to combine the aforementioned quantitative analysis results with reference contexts retrieved via RAG, and generates analysis statements according to military bulleted style and predefined report structures. In this process, the local LLM-based comprehensive analysis unit ensures the consistency and reliability of the analysis results by controlling roles, tasks, data usage rules, and output formats through system prompts.

[0228] The report generation unit assembles the analysis documents generated by the local LLM-based comprehensive analysis unit according to the standard structure of military analysis reports, and generates a final report by automatically inserting graphs and tables based on quantitative analysis values. The generated report can be output in at least one of PDF, DOCX, or TXT formats and is securely provided in a closed network environment.

[0229] The control unit controls the overall operation of the military analysis report generation device (920) and performs control of the execution order of each module, data routing, error and exception handling, re-analysis request processing, and audit record generation. In addition, the control unit can control the delivery of the generated report to an analyst or manager via a third-party device (930) or a user interface (UI) after the analysis is completed.

[0230] Meanwhile, according to an embodiment, the military analysis report generating device (920) may be implemented by being logically or physically integrated with a network device constituting a communication network (910), and may be operated within the same security domain as the defense M&S simulation device (900).

[0231] The third-party device (930) refers to an external device for viewing, utilizing, or using for subsequent tasks the analysis results and military analysis reports generated by the military analysis report generating device (920).

[0232] The third-party device (930) may include a defense personnel terminal, a command and control system terminal, a computer for an analyst or manager, or a defense-related information system, and may be connected to a military analysis report generation device (920) via a communication network (910). The third-party device (930) functions as a consumer or user of the analysis results and performs the role of verifying the generated results and utilizing them for decision-making, without performing the simulation itself or executing the analysis logic.

[0233] A third-party device (930) can display the final analysis report, summary information, quantitative analysis results, or corrective action and countermeasure (CoA) information transmitted from the military analysis report generating device (920) through a user interface (UI). For example, the third-party device (930) can view a military analysis report on the screen that includes maneuver analysis, detection analysis, engagement results, damage assessment, and comprehensive judgment, or download and check a report in PDF, DOCX, or TXT format.

[0234] Additionally, the third-party device (930) can transmit follow-up requests to the military analysis report generating device (920). For example, the user can input a request for a re-summary of specific analysis results, a request for regeneration of specific sections, a request for re-analysis of items that do not meet reliability standards, or a request for re-analysis reflecting additional scenario conditions through the third-party device (930). These requests can be transmitted to the military analysis report generating device (920) via the communication network (910) and processed by the control unit.

[0235] Meanwhile, it is preferable that the third-party device (930) be operated in a closed network environment separated from the external network, and access may be permitted only to users who have been granted access rights within the same security domain as the military analysis report generating device (920). Additionally, depending on the embodiment, the third-party device (930) may be implemented as a separate device physically separated from the military analysis report generating device (920), or it may be implemented in the form of a logically separated user interface module within the same hardware device.

[0236] In addition, the third-party device (930) can provide a function to check the analysis history, report generation history, version information of the local large language model used, and data hash value for auditing and verification purposes, and can be used to verify the reproducibility and reliability of the analysis results.

[0237] In addition to the above, details regarding the defense M&S simulation device (900), communication network (910), military analysis report generation device (920), and third-party device (930) of FIG. 9 have been sufficiently explained through FIG. 1 to FIG. 8, so the details will be replaced by FIG. 8.

[0238] FIG. 10 is a block diagram illustrating the detailed structure of the military analysis report generating device of FIG. 9.

[0239] As illustrated in FIG. 10, the military analysis report generating device (920) of FIG. 9 according to another embodiment of the present invention includes some or all of the communication interface unit (1000), control unit (1010), military analysis report generating unit (1020), and storage unit (1030).

[0240] Here, "including some or all" means that some components, such as the storage unit (1030), may be omitted to form the military analysis report generating device (920) of FIG. 9, or some components, such as the military analysis report generating unit (1020), may be integrated into other components, such as the control unit (1010). To facilitate a sufficient understanding of the invention, it is described as including all components.

[0241] A communication interface unit (1000) of a military analysis report generating device (920) according to another embodiment of the present invention is a component that provides an interface for transmitting and receiving data with a defense modeling and simulation (M&S) simulation device (900) and, if necessary, a third-party device (930).

[0242] The communication interface unit (1000) is configured to receive simulation result data generated during the simulation process or as a result of the simulation from the defense M&S simulation device (900) via the communication network (910). The simulation result data may include event logs related to maneuver, detection, engagement, and damage assessment, state change information, and raw data for analysis, and may be provided in the form of log data.

[0243] The communication interface unit (1000) can verify the format and integrity of the received simulation result data. For example, the communication interface unit (1000) can check whether the received data is in a JSON, CSV, or equivalent structured format, and if there is data omission, format error, or identifier mismatch, it can transmit error information to the control unit (1010).

[0244] Additionally, the communication interface unit (1000) can support the reception of data including timestamps, data identifiers, or sequence information, taking into account delays, losses, or out-of-order events that may occur during the data transmission process. Through this, the control unit (1010) can sort the received simulation result data in chronological order or perform a retransmission request.

[0245] Meanwhile, the communication interface unit (1000) can perform the function of transmitting analysis results, report output files, or processing status information generated by the military analysis report generating device (920) to a third-party device (930) via the communication network (910). For example, the communication interface unit (1000) can transmit an analysis report in PDF, DOCX, or TXT format, or transmit an analysis completion notification, an error notification, or a response to a re-analysis request.

[0246] Additionally, the communication interface unit (1000) may be configured to operate in a closed network environment separated from an external network, and may support authentication, access control, or encrypted data transmission depending on the level of security requirements. According to an embodiment, the communication interface unit (1000) may be implemented as a wired communication interface, a wireless communication interface, or a combination thereof.

[0247] Meanwhile, the communication interface unit (1000) may be implemented by being logically or physically integrated with other components constituting the military analysis report generating device (920). For example, some functions of the communication interface unit (1000) may be implemented by being included in the control unit (1010) or the military analysis report generating unit (1020), and such configuration is also included within the scope of the present invention.

[0248] The control unit (1010) of the military analysis report generation device (920) is a core component that controls the overall operation of the device. The control unit (1010) may be configured as a single chip in the form of an integrated circuit (IC) chip, including a processor such as a central processing unit (CPU) and a memory such as random access memory (RAM) in terms of hardware. In this case, the control unit (1010) can substantially perform the operation of the military analysis report generation unit (1020) by copying and loading a program mounted on the military analysis report generation unit (1020) into memory and then executing it. However, to facilitate a sufficient understanding of the invention, the present specification describes the control unit (1010) as a component that performs the overall control operation of the military analysis report generation device (920).

[0249] The control unit (1010) controls the operation sequence of each component constituting the military analysis report generation device (920) based on the simulation result data of the defense modeling and simulation (M&S) simulation device (900) received through the communication interface unit (1000). For example, the control unit (1010) can control the execution of the preprocessing step, the rule-based analysis step, the RAG-based information retrieval step, the local LLM-based comprehensive analysis step, and the report generation step in sequence.

[0250] Additionally, the control unit (1010) manages the data flow between each stage. Specifically, the control unit (1010) transmits the preprocessing results to the rule-based analysis module, transmits the rule-based analysis results to the RAG-based information retrieval unit, and controls the input of the retrieved reference context and quantitative analysis results to the local LLM-based comprehensive analysis unit. Subsequently, the control unit (1010) transmits the output results of the local LLM to the report generation unit to control the generation of a military analysis report.

[0251] The control unit (1010) also performs control functions for error and exception situations. For example, if format errors in simulation result data, omission of analysis steps, failure of RAG search, or degradation of local LLM output occur, the control unit (1010) can control the request for reprocessing, re-search, or re-analysis. In addition, the control unit (1010) can ensure the reproducibility and reliability of the analysis results by controlling the generation of logs and audit records for the entire analysis process.

[0252] Meanwhile, the control unit (1010) can control the transmission of the analysis report and analysis results generated by the military analysis report generation device (920) to a third-party device (930) through the communication interface unit (1000), and may also be configured to process a re-summary request, a specific section regeneration request, or a re-analysis request received from a user.

[0253] In this way, the control unit (1010) can be implemented in hardware as an integrated circuit including a processor and memory, and functionally performs the role of comprehensively controlling the overall analysis flow of the military analysis report generating device (920) and the operation of each component.

[0254] The military analysis report generation unit (1020) of the military analysis report generation device (920) according to another embodiment of the present invention is a functional component for performing automatic analysis of the results of defense modeling and simulation (M&S) and generating the results in the form of a military analysis report.

[0255] The military analysis report generation unit (1020) performs a military analysis procedure including maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using simulation result data received through the communication interface unit (1000) under the control of the control unit (1010). To this end, the military analysis report generation unit (1020) may be configured to include functions for preprocessing of simulation result data, rule-based analysis, reference knowledge search, generation of analysis statements based on a large language model, and report assembly.

[0256] Specifically, when simulation result data is provided in the form of log data, the military analysis report generation unit (1020) may be configured to structure the log data to enable time alignment, missing value processing, and event classification, and to calculate quantitative parameters as analysis reference values ​​that numerically express the state or performance of maneuver, detection, engagement, and damage assessment. These quantitative parameters may consist of Time-To-Detect (TTD), detection range, Time-To-Kill (TTK), relative approach speed, course change amount, damage level, etc.

[0257] Additionally, the military analysis report generation unit (1020) may be configured to perform functional rule-based analysis for maneuver analysis, detection performance analysis, engagement result analysis, and damage assessment analysis using the above quantitative parameters. At this time, the rule-based analysis may be performed using conditions defined by a tactical model (Doctrine), rules of engagement (ROE), sensor model, or weapon system operation rules.

[0258] The military analysis report generation unit (1020) may also include a function for searching external reference knowledge documents, including military doctrines, rules of engagement, sensor models, or simulated logic analysis documents corresponding to the rule-based analysis results, using a search augmentation generation (RAG) method. To this end, the military analysis report generation unit (1020) may construct a reference context used for inputting a large language model by utilizing a vector index that converts and stores the reference knowledge documents into embedding vectors, and by searching for related paragraphs based on similarity based on quantitative analysis results and metadata.

[0259] Furthermore, the military analysis report generation unit (1020) may be configured to use a local large language model (LLM) operating in a closed network environment to integrate the quantitative analysis results and reference contexts retrieved via RAG, and to generate an analysis report according to a military bulleted style and a predefined report structure. At this time, the military analysis report generation unit (1020) may be configured so that the output of the local large language model is controlled to use the quantitative analysis results as they are without altering them.

[0260] Additionally, the military analysis report generation unit (1020) can be configured to automatically assemble the output results of the local large language model according to the standard defense analysis report structure. For example, the analysis text can be classified and assigned into multiple sections including a simulation overview, maneuver analysis results, detection effect analysis, engagement result analysis, damage assessment analysis, and comprehensive judgment and operational implications, and a final report can be generated by automatically inserting graphs and tables based on quantitative analysis values.

[0261] In addition, the military analysis report generation unit (1020) may be configured to automatically derive corrective measures, improvement plans, or courses of action (CoA) based on the analysis results, and to include each course of action in the report by assigning a priority or recommendation grade.

[0262] The military analysis report generation unit (1020) can output the finally generated military analysis report in PDF, DOCX, or TXT format, and the generated report can be provided through a third-party device (930) or an internal user interface under the control of the control unit (1010).

[0263] In this way, the military analysis report generation unit (1020) is a core component that automatically generates a military analysis report on the results of defense M&S simulations by integrally performing automatic analysis based on defense M&S simulation result data, combination of reference knowledge, generation of description based on a large language model, and assembly of reports.

[0264] The storage unit (1030) of the military analysis report generation device (920) is a component for storing and managing data, intermediate results, and operational information used in the process of generating a military analysis report. The storage unit (1030) may be configured to support temporary storage, caching, and high-speed access to data generated during the operation of the military analysis report generation unit (1020) under the control of the control unit (1010). The storage unit (1030) may be implemented in the form of, for example, volatile memory (RAM), non-volatile memory (NVRAM), flash memory, or a high-speed storage device.

[0265] Specifically, the storage unit (1030) can store the following data. First, the simulation result data or log data of the defense M&S simulation device (900) received through the communication interface unit (1000) can be stored. Second, quantitative parameters, analysis reference values, event classification results, and metadata calculated during the rule-based analysis process can be stored. Third, search queries generated during the RAG-based information retrieval process, searched reference contexts, temporary embedding results, and similarity scores can be stored. Fourth, intermediate analysis statements, reliability scores, section-by-section output results, and post-processing results generated during the local large language model (LLM)-based comprehensive analysis process can be stored. Fifth, template information used during the report generation process, section mapping results, and temporary data required for generating graphs and tables can be stored.

[0266] Meanwhile, the storage unit (1030) according to the present invention is configured to store and manage data generated during operation or temporary data required during the analysis process as described above, and permanent data requiring long-term storage may be configured to be stored and managed in the storage area (920a) of the military analysis report generation system shown in FIG. 9. That is, permanent data such as reference knowledge data such as military doctrine documents, rules of engagement (ROE), sensor models, and simulation logic analysis reports, vector indices composed of embedding vectors, final military analysis reports completed, audit logs, and analysis history may be stored and managed in the storage area (920a) located outside the military analysis report generation device (920) or in a higher layer. The storage unit (1030) may be configured to refer to the permanent data stored in the storage area (920a) or transfer data generated after analysis completion to the storage area (920a) according to the control of the control unit (1010) when necessary.

[0267] In this way, the storage unit (1030) is responsible for storing temporary and operational data to ensure efficiency and speed of analysis processing within the military analysis report generating device (920), and can be configured with a dual storage structure in which long-term storage and system-level data management are delegated to the DB (920a) of FIG. 9.

[0268] In addition to the above, the communication interface unit (1000), control unit (1010), military analysis report generation unit (1020), and storage unit (1030) of FIG. 10 can perform various operations, and other details have been sufficiently explained above, so they will be replaced by those details.

[0269] Meanwhile, the communication interface unit (1000), control unit (1010), military analysis report generation unit (1020), and storage unit (1030) of FIG. 10 according to an embodiment of the present invention are composed of hardware modules that are physically separated from each other, but each module may store software for performing the above operations internally and execute it. However, since the software is a set of software modules and each module can be formed as hardware, the configuration will not be specifically limited to software or hardware. For example, the storage unit (1130) may be storage or memory which is hardware. However, since it is also possible to store information in a software repository, the above content will not be specifically limited.

[0270] In addition, as another embodiment of the present invention, the control unit (1010) may include a CPU and memory and may be formed as a single chip. The CPU includes a control circuit, an arithmetic logic unit (ALU), an instruction interpretation unit, and a registry, and the memory may include RAM. The control circuit may perform control operations, the arithmetic logic unit may perform operations on binary bit information, and the instruction interpretation unit may include an interpreter or a compiler to perform operations that convert high-level language into machine language and machine language into high-level language, and the registry may be involved in software data storage. According to the above configuration, for example, at the beginning of the operation of the military analysis report generation device (920), the program stored in the military analysis report generation unit (1020) may be copied and loaded into memory, i.e., RAM, and then executed, thereby rapidly increasing the data calculation processing speed. In the case of a deep learning model, it may be loaded into GPU memory instead of RAM and executed by using the GPU to accelerate the execution speed.

[0271] Figure 11 is a flowchart showing the operation process of the military analysis report generating device of Figure 9.

[0272] For convenience of explanation, referring to FIG. 11 together with FIG. 9, a military analysis report generating device (920) according to an embodiment of the present invention receives simulation result data generated as a result of a simulation execution process or execution from a defense modeling and simulation (M&S) device (900) (S1100).

[0273] Specifically, the military analysis report generating device (920) receives simulation result data generated during the simulation process or as a result of the simulation in the defense modeling and simulation (M&S) device through a communication interface unit. The simulation result data may be provided in the form of simulation log data, event logs, state change records, sensor output data, engagement history data, or damage assessment related data.

[0274] The simulation result data may be provided in the form of log data including timestamps, and may include platform identifiers, target identifiers, location information, velocity information, course information, sensor status information, weapon operation information, and damage status information. Additionally, the simulation result data may be provided in JSON, CSV, or a similar structured data format.

[0275] The military analysis report generating device (920) can verify the validity of the data format, the existence of required fields, and the inclusion of simulation scenario metadata for the received simulation result data, and initializes the input data so that automatic analysis in subsequent steps is possible according to the verification results.

[0276] In this way, step S1100 is a step in which a military analysis report generating device (920) receives simulation result data generated from a defense modeling and simulation (M&S) device, thereby securing basic data for preprocessing, rule-based analysis, search augmentation generation (RAG)-based reference context search, and local large language model-based analysis report generation performed in subsequent steps.

[0277] In addition, the military analysis report generation device (920) generates a quantitative analysis result by performing a functional rule-based analysis of maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the received simulation result data, and searches for a reference context in a search augmentation generation (RAG) manner from a storage unit where an external reference knowledge document, including military doctrine, rules of engagement (ROE), sensor model, or simulation logic analysis document corresponding to the quantitative analysis result, is stored as an embedding vector using a vector index, and inputs the quantitative analysis result and the reference context into a local large language model operating in a closed network environment to automatically generate a military analysis report on the defense M&S simulation result (S1110).

[0278] Specifically, the military analysis report generating device (920) generates quantitative analysis results by performing functional rule-based analysis for maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the simulation result data received in step S1100. At this time, the rule-based analysis can be performed using conditions defined by a tactical model (Doctrine), rules of engagement (ROE), sensor model, or simulation logic, and quantitative parameters that numerically express the state or performance of each analysis step, such as Time-To-Detect (TTD), detection range, Time-To-Kill (TTK), relative approach speed, course change amount, damage level, etc., can be calculated.

[0279] Subsequently, the military analysis report generation device (920) searches for external reference knowledge documents corresponding to the quantitative analysis results using a search augmentation generation (RAG) method. To this end, a search query based on the quantitative analysis results and metadata is generated using a vector index in which external reference knowledge documents, including military doctrine, rules of engagement (ROE), sensor models, or simulation logic analysis documents, are stored as embedding vectors, and a reference context is constructed by searching for related paragraphs based on similarity from a storage unit in which the vector index is stored.

[0280] The military analysis report generation device (920) inputs the generated quantitative analysis results and the reference context into a local large language model (LLM) operating in a closed network environment. At this time, the local large language model is controlled to use the quantitative analysis results as they are without modification, and is controlled to generate an analysis statement according to a military modified style and a predefined report structure based on the military doctrine, rules of engagement, and sensor model included in the reference context.

[0281] Furthermore, the military analysis report generating device (920) automatically generates a military analysis report including a simulation overview, maneuver analysis results, detection effect analysis, engagement result analysis, damage assessment analysis, comprehensive judgment, and operational implications using the output results of a local large language model.

[0282] As such, step S1110 is a key step that quantitatively analyzes the received defense M&S simulation result data based on rules, combines the reference context obtained via Search Augmented Generation (RAG) with a local large language model, and automatically generates a military analysis report on the defense M&S simulation results.

[0283] Meanwhile, the military analysis report generation device (920) can perform additional subsequent operations to ensure the reliability of the analysis results, improve the completeness of the report, and enhance operational convenience even after the military analysis report is automatically generated in step S1110.

[0284] Specifically, the military analysis report generation device (920) can verify whether the analysis text generated by the local Large Language Model (LLM) matches the quantitative analysis results, and control the device to perform re-analysis or regeneration if distortion or omission of quantitative values ​​occurs. In addition, if the reference context searched by the Search Augmentation Generation (RAG) method is insufficient or does not meet the similarity criteria, the device can control the device to perform a re-search by changing the search conditions or to apply alternative reference documents.

[0285] Additionally, the military analysis report generating device (920) can rearrange the generated military analysis report according to the standard defense analysis report structure and refine sentence connections and expressions so that a logical flow between each section is maintained. In this process, a graph or table based on quantitative analysis values ​​can be automatically generated and inserted into the report, and source information and reliability index of the retrieved reference documents can be included as additional information in the report.

[0286] In addition, the military analysis report generating device (920) can generate logs and audit records for the entire analysis process and store the input data used in the analysis, rule-based analysis results, retrieved reference contexts, versions of local large language models, and output results so as to be traceable. These audit records can be used for re-analysis, review by other analysts, or verification of research reports.

[0287] Finally, the military analysis report generating device (920) provides the completed military analysis report to an analyst or manager through a user interface or internal file system, and can process requests for resummarization of the report, regeneration of specific analysis sections, or additional analysis as needed.

[0288] Accordingly, the driving method of the military analysis report generation device according to the present invention is performed primarily through steps S1100 and S1110, and thereafter, a plurality of subsequent operations are continuously performed to improve analysis quality, ensure reliability, and ensure operational convenience, thereby completing a series of processing procedures for automatically generating and managing military analysis reports on defense M&S simulation results.

[0289] In addition to the above, the military analysis report generating device (920) of FIG. 9 according to the embodiment of the present invention can perform various operations, and other detailed information has been sufficiently explained above, so it will be replaced by that information.

[0290] Although all components constituting an embodiment of the present invention have been described as being combined or operating as a single unit, the present invention is not necessarily limited to such an embodiment. That is, within the scope of the purpose of the present invention, all components may be selectively combined and operated in one or more ways. Furthermore, while all components may each be implemented as a single independent piece of hardware, they may also be implemented as a computer program having a program module that performs some or all of the combined functions on one or more pieces of hardware by selectively combining some or all of the components. The codes and code segments constituting the computer program can be easily inferred by those skilled in the art of the present invention. An embodiment of the present invention may be implemented by storing such a computer program on a non-transitory computer-readable medium, reading it, and executing it by a computer.

[0291] Here, a non-transient readable recording medium refers to a medium that stores data semi-permanently and can be read by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, the programs described above may be stored and provided on non-transient readable recording media such as CDs, DVDs, hard disks, Blu-ray discs, USBs, memory cards, and ROMs.

[0292] Although embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention may be implemented in other specific forms without changing the technical concept or essential features thereof. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. Explanation of the symbols

[0293] 100 : Defense M&S Simulation Result Automatic Analysis and Report Generation System 200: Data Collection and Preprocessing Unit 210: Data Input Unit 220 : Preprocessing Engine 230 : Metatagging Unit 300: Rule-based analysis module 310: Startup analysis unit 320: Detection Analysis Unit 330: Engagement Analysis Unit 340 : Damage Assessment Analysis Department 400 : RAG-based information retrieval and context reinforcement unit 410 : Embedding engine 420 : Vector index 430 : Search query generation unit 440 : Similarity search unit 450 : Contextual Reinforcement Section 500 : Local LLM-based Comprehensive Analysis Section 510 : System Prompt Generator 520 : LLM Input Configurator 530 : Local LLM Inferer 540 : LLM Output Postprocessor 600 : Report Generation Unit 610 : Section Mapper 620 : Evidence and Reliability Organizer 630 : Policy Analysis & Generator 640: Report assembler and printer 700, 1010: Control unit 710: Flow Controller 720: Data Routing Manager 730: Error and Exception Handlers 740: Log and Audit Record Manager 800: User Interface (UI) 810: Result display area by clause (function) 820: Area for displaying grounds and reliability 830: Area for suggesting corrective action 840 : Simulation Results Summary and Report Output Area 900: Defense M&S Simulation Device 910: Communication Network 920 : Military analysis report generator 930 : Third-party device 1000: Communication Interface Unit 1020: Military Analysis Report Generation Unit 1030: Storage

Claims

Claim 1 A military analysis report generating device comprising: a communication interface unit for receiving simulation result data generated during or as a result of a simulation execution process from a defense modeling and simulation (M&S) device; and a control unit that generates quantitative analysis results by performing functional rule-based analysis of maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the simulation result data received through the communication interface unit, and searches for a reference context from a storage unit using a search augmentation generation (RAG) method using a vector index in which an external reference knowledge document, including military doctrine, rules of engagement (ROE), sensor model, or simulation logic analysis document corresponding to the quantitative analysis results, is stored as an embedding vector, and inputs the quantitative analysis results and the reference context into a local large language model operating in a closed network environment to automatically generate a military analysis report on defense M&S simulation results. Claim 2 In claim 1, the control unit comprises: a data collection and preprocessing unit that, when simulation results generated from a defense modeling and simulation (M&S) device are provided in the form of log data, receives said log data, performs time alignment, missing value processing, and event classification, and calculates quantitative parameters as analysis reference values ​​that numerically express the state or performance of maneuver, detection, engagement, and damage assessment; a rule-based analysis unit that generates quantitative analysis results by performing functional rule-based analysis on maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using said preprocessed data; a RAG-based information retrieval and context reinforcement unit that constructs a reference context for LLM input by searching for paragraphs associated with said quantitative analysis results based on similarity, based on a vector index in which external reference knowledge documents including military doctrine, rules of engagement (ROE), sensor models, and simulation logic analysis documents are stored as embedding vectors; a local LLM-based comprehensive analysis unit that integrates said quantitative analysis results and said reference context using a local Large Language Model (LLM) operating in a closed network environment and generates an analysis statement with a military modified sentence structure; and said comprehensive A military analysis report generating device comprising: a report generating unit that automatically configures the analysis statement generated by the analysis unit according to the standard defense analysis report template structure of simulated overview, maneuver analysis, detection effect, engagement result, damage assessment, and comprehensive judgment to generate a final report. Claim 3 A military analysis report generating device according to paragraph 2, wherein the data collection and preprocessing unit performs timestamp correction and common timeline alignment of simulation logs, interpolation of missing values ​​of coordinates, speed, and course, rule-based classification of maneuver, detection, engagement, and damage assessment events, calculation of quantitative parameters of Time to Kill (TTD), Time to Kill (TTK), detection range, and relative approach speed, and assignment of metadata including platform, scenario, event type, and doctrine-based tags. Claim 4 In paragraph 2, the rule-based analysis unit comprises determining a maneuver pattern using a change in course, a rate of change in speed, a change in distance and bearing, evaluating detection performance based on sensor signal-to-noise ratio (SNR), detection probability (Pd), and false alarm probability (Pfa), determining interception success based on engagement initiation conditions and kill radius, and calculating damage levels (Level 0~3) and remaining combat power, a military analysis report generating device. Claim 5 In paragraph 2, the RAG-based information retrieval and context reinforcement unit subdivides military doctrine, rules of engagement, and sensor model documents into paragraph units and converts them into embedding vectors, stores the vectors in an approximate nearest neighbor search (ANN)-based vector index, generates a search query vector based on quantitative analysis results to search for related paragraphs using a cosine similarity-based Top-k method, and configures the searched paragraphs and source information into context blocks for LLM input, a military analysis report generation device. Claim 6 A military analysis report generation device according to paragraph 2, wherein the local LLM-based comprehensive analysis unit provides control instructions to the local LLM, including military writing style, analysis stage structure, rules for direct use of quantitative values, and rules for prioritizing RAG grounds, through a system prompt generator; performs local LLM inference by configuring a quantitative analysis result block, a RAG search context block, and a scenario metadata block as integrated inputs; and provides the output analysis statement to the report generation unit after performing duplicate removal, quantitative value matching verification, and reliability score calculation in a post-processor. Claim 7 A military analysis report generating device according to paragraph 2, wherein the report generating unit automatically assigns LLM output statements to six sections corresponding to the simulation overview, maneuver analysis, detection analysis, engagement analysis, damage assessment, and comprehensive judgment of the standard defense analysis report structure, organizes RAG citations, source information, and reliability index to include them as additional information, and automatically inserts graphs and tables based on quantitative analysis values ​​to output a final report in PDF, DOCX, or TXT format. Claim 8 A military analysis report generating device according to claim 1, wherein the control unit performs control of the execution order of each module configured internally, data structure conversion and routing, re-search or application of alternative literature when RAG search is insufficient, request for regeneration when LLM output is omitted or quantitative value distortion occurs, and generation of an audit record including a model version, input data hash, similarity score, and confidence score. Claim 9 A method for operating a military analysis report generation device, comprising: a step in which a communication interface unit receives simulation result data generated during or as a result of a simulation execution process from a defense modeling and simulation (M&S) device; and a step in which a control unit performs functional rule-based analysis of maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the simulation result data received through the communication interface unit to generate a quantitative analysis result, and searches for a reference context in a search augmentation generation (RAG) manner from a storage unit storing a vector index using a vector index in which an external reference knowledge document, including a military doctrine, rules of engagement (ROE), sensor model, or simulation logic analysis document corresponding to the quantitative analysis result, is stored as an embedding vector, and inputs the quantitative analysis result and the reference context into a local large language model operating in a closed network environment to automatically generate a military analysis report on the defense M&S simulation result. Claim 10 A military analysis report generation system comprising: a defense modeling and simulation (M&S) device that provides simulation result data generated during or as a result of a simulation execution process; and a military analysis report generation device that generates quantitative analysis results by performing functional rule-based analysis of maneuver analysis, detection analysis, engagement analysis, and damage assessment analysis using the simulation result data provided from the defense modeling and simulation (M&S) device, searches for a reference context from a storage unit using a search augmentation generation (RAG) method using a vector index in which an external reference knowledge document, including military doctrine, rules of engagement (ROE), sensor model, or simulation logic analysis document corresponding to the quantitative analysis results, is stored as an embedding vector, and inputs the quantitative analysis results and the reference context into a local large language model (LLM) operating in a closed network environment to automatically generate a military analysis report on defense M&S simulation results.

Images