Ai workflow-based multimodal integrated control method and intelligent resource management system for tco reduction

Abstract

An AI Workflow-based multimodal integrated control method and an intelligent resource management system for reducing TCO are disclosed. An intelligent resource management system for reducing Total Cost of Ownership (TCO) according to one embodiment may include: a multimodal data integration processing engine that preprocesses multimodal data collected in real time into a form that can be processed by a Large Language Model (LLM); an LLM-based situation analysis engine that generates a situation-adaptive workflow based on situation information interpreted through relationship analysis using a Large Language Model for the preprocessed multimodal data; and a 3D topology visualization engine that visualizes an infrastructure operating environment based on the generated situation-adaptive workflow.

Description

Technology Field

[0001] The following description concerns technology for managing resources by integrating and monitoring multimodal data. Background Technology

[0003] Current infrastructure management systems have major limitations, including inefficiency due to distributed management frameworks, lack of responsiveness due to static workflows, and analysis based on a single modality, which leads to reduced operational efficiency and increased costs.

[0004] I will now explain the inefficiencies resulting from the decentralized management system. Various infrastructure environments, including cloud, on-premises, and edge, are managed individually. Since data integration between systems does not occur in real-time, it is difficult to accurately assess the current status, and end-to-end management is impossible due to a lack of interoperability between systems. Consequently, costs increase due to redundant management tools and personnel (increased operating costs), a burden arises from having to check multiple systems when a problem occurs (delayed response time), and over-provisioning occurs because integrated resource optimization is difficult (inefficient resource utilization).

[0005] Furthermore, we will discuss the lack of responsiveness caused by static workflows. Since work is performed only according to predefined workflows, there is a lack of flexibility. It is difficult to apply situationally optimized processes, leading to persistent inefficient workflows. Automated responses to exceptions are limited. Consequently, incorporating new situations or requirements requires significant time and cost (delayed response to change), unnecessary work occurs due to processes unsuitable for the situation (reduced operational efficiency), and prompt action is difficult when unexpected problems arise (limited failure response).

[0006] Furthermore, we will explain the limitations of single-modality-based analysis. Data collection and analysis are limited and centered on text logs; it fails to process various forms of data, such as images and audio, in an integrated manner, making it difficult to accurately analyze the causes of complex failure situations. Consequently, it is difficult to accurately grasp the situation using only a single type of data (reduced analysis accuracy), accurate predictions are restricted due to the lack of multidimensional data analysis (limitations in prediction performance), and decision-making uncertainty increases due to limited information (decision delay).

[0007] To overcome these limitations, it is necessary to develop a multimodal integrated management system based on AI Workflow. The problem to be solved

[0009] The present invention aims to provide an integrated solution that autonomously and intelligently manages an enterprise's IT infrastructure by combining a multimodal LLM and an AI Workflow engine. In particular, the present invention focuses on maximizing the efficiency and stability of infrastructure operations through the integrated analysis of various forms of data, the generation of context-aware workflows, and intuitive 3D visualization. The ultimate goal is to reduce the management costs of complex IT infrastructure, enable failure prevention and rapid response, and improve the work efficiency of operations personnel. means of solving the problem

[0011] An intelligent resource management system for reducing Total Cost of Ownership (TCO) may include: a multimodal data integration processing engine that preprocesses multimodal data collected in real time into a form that can be processed by a Large Language Model (LLM); an LLM-based context analysis engine that generates a context-adaptive workflow based on context information interpreted through relationship analysis using a Large Language Model for the preprocessed multimodal data; and a 3D topology visualization engine that visualizes the infrastructure operating environment based on the generated context-adaptive workflow.

[0012] The above multimodal data integration processing engine includes collecting multimodal data including log data, metric performance indicators, image data, and voice data, and performing preprocessing on the collected multimodal data, wherein the log data includes system logs, application logs, and security logs, and the metric performance indicators include performance indicators including CPU, memory, disk, and network, and the image data may be collected through CCTV and thermal imaging cameras, and the voice data may be collected through acoustic sensors.

[0013] The above multimodal data integration processing engine can dynamically configure a pipeline for preprocessing for each data type in the collected multimodal data, perform preprocessing including data cleaning, normalization, and feature extraction on the collected multimodal data through the configured preprocessing pipeline, adjust parameters of the preprocessing pipeline through quality inspection of the preprocessed data, and generate metadata by reprocessing the preprocessed data through the preprocessing pipeline with adjusted parameters.

[0014] The above multimodal data integration processing engine includes extracting each embedding vector from the multimodal data for which the above preprocessing has been performed, and can generate text embeddings from text data in the multimodal data through a language model, generate image embeddings from image data through an image processing model, and generate speech embeddings from speech data through a speech recognition model.

[0015] The above multimodal data integration processing engine can index each of the extracted embedding vectors and store them in an integrated vector database, and support searching through similarity calculation for multimodal queries in the integrated vector database.

[0016] The above LLM-based situation analysis engine can identify situational information of the current system through correlation analysis between time, space, and events from the stream of the preprocessed multimodal data, and can track the causal relationships of related events through dependency mapping between system components.

[0017] The above LLM-based situation analysis engine can derive diagnoses and measures regarding the identified current system situation information based on rules required for infrastructure operation, detect abnormal event patterns through sequence monitoring of events occurring throughout the current system to predict the possibility of failure, and generate warning information based on the predicted possibility of failure.

[0018] The above LLM-based situation analysis engine can obtain pattern matching results from a stream of the preprocessed multimodal data through a deep learning-based time-series pattern recognition model that has learned normal or abnormal patterns of past infrastructure operation data, and can predict system state and performance changes based on the obtained pattern matching results to support response information.

[0019] The above LLM-based situation analysis engine detects anomalies based on the acquired pattern matching results, generates a recovery policy based on the characteristics of the detected anomalies and the system state, sequentially executes an automated recovery procedure according to the generated recovery policy, and can utilize the executed recovery procedure and the detected anomalies as training data for failure pattern analysis and recovery strategy improvement.

[0020] The above LLM-based situation analysis engine can predict future resource demand through the analysis of resource usage patterns, establish a capacity plan based on business requirements and system constraints based on the predicted future resource demand, and evaluate the risk level of the current system based on the established capacity plan.

[0021] The above LLM-based situation analysis engine analyzes the type of LLM model, processing priority, and resource requirements according to an LLM processing request, monitors the status of currently available computing nodes in real time through the analysis, establishes a task allocation plan based on the monitored status of currently available computing nodes, searches for a pre-loaded cache for LLM models requested above a preset standard according to the established task allocation plan, and adjusts the scheduling policy by analyzing the processing history and performance metrics of the loaded LLM models using the searched cache.

[0022] The above 3D topology visualization engine can generate a topology by mapping infrastructure components and relationship information of infrastructure components into a three-dimensional space, and can generate a three-dimensional object based on configuration information obtained through the mapping.

[0023] The above 3D topology visualization engine can process interaction commands using user interaction events and visually represent system performance data and status information on the generated topology.

[0024] The above 3D topology visualization engine provides a visualization view based on situation information and event data, visually represents the event data and system notifications, and can control system state changes over time.

[0025] An infrastructure integrated control method performed by an intelligent resource management system for reducing the Total Cost of Ownership (TCO) may include: a step of preprocessing multimodal data collected in real time into a form that can be processed by a Large Language Model (LLM); a step of generating a context-adaptive workflow based on context information interpreted through relationship analysis using the Large Language Model for the preprocessed multimodal data; and a step of visualizing the infrastructure operating environment based on the generated context-adaptive workflow. Effects of the invention

[0027] 1. Technological advancement

[0028] Multimodal LLM-based Integrated Analysis: A comprehensive understanding of infrastructure status is possible through the integrated analysis of various data forms, such as text, images, and audio. Correlation analysis of multimodal data enables the identification of complex issues that are difficult to detect through single-modality analysis. Domain-specific LLM allows for the accurate grasp and processing of the professional context within the infrastructure management domain.

[0029] 2. Operational Efficiency

[0030] AI Workflow-based Automation: Manual intervention by operations personnel can be minimized through the creation of context-aware, dynamic workflows. Failure prevention and downtime can be minimized through predictive maintenance based on pattern learning. Operating costs can be reduced and system performance maximized through automated resource optimization.

[0031] 3. Scalability

[0032] Cloud-Native Architecture: A microservices-based architecture enables independent scaling of system components. Container-based deployment allows for consistent operation across diverse environments. API-based standard interfaces enable flexible integration of new features and services.

[0033] 4. User Experience

[0034] Intuitive 3D Topology Visualization: Complex infrastructure configurations can be intuitively grasped through real-time 3D topology maps. Interactive analysis tools allow users to easily explore and analyze the information they need. Context-based auto-focus enables immediate recognition and response to critical events.

[0035] 5. Security

[0036] Integrated Security Management: Security threats can be identified early through anomaly detection via multimodal data analysis. New and unknown types of security threats can be detected through AI-based pattern analysis. A consistent level of security can be maintained through the application of automated security policies. Brief explanation of the drawing

[0038] FIG. 1 is a drawing for explaining an intelligent resource management system in one embodiment. FIG. 2 is a diagram illustrating a multimodal data integration processing engine in one embodiment. FIG. 3 is a diagram illustrating an LLM-based situation analysis engine in one embodiment. FIG. 4 is a diagram illustrating a 3D topology visualization engine in one embodiment. FIG. 5 is a diagram illustrating the entire process of a core module in one embodiment. FIG. 6 is a diagram illustrating the process of a core module and a peripheral module in one embodiment. FIGS. 7 to 9 are drawings for explaining the system architecture and cloud architecture in one embodiment. FIG. 10 is a diagram illustrating the preprocessing optimization process of multimodal data in one embodiment. FIG. 11 is a diagram illustrating a distributed LLM processing process in one embodiment. FIG. 12 is a diagram illustrating an intelligent fault prediction and recovery process in one embodiment. Specific details for implementing the invention

[0039] Hereinafter, embodiments will be described in detail with reference to the attached drawings.

[0041] In the embodiments, the intelligent resource management system can establish an integrated infrastructure monitoring system. The intelligent resource management system can establish a system for collecting and integrating various forms of data, such as text, images, and voice, in real time. The intelligent resource management system can accurately identify and interpret complex system states by utilizing multimodal LLM. The intelligent resource management system can enhance management efficiency by integrating distributed monitoring tools into a single platform.

[0042] Furthermore, the Intelligent Resource Management System can realize intelligent operational automation. Through an AI Workflow engine, the Intelligent Resource Management System can automatically generate and execute operational processes optimized for specific situations. By learning patterns from historical operational data, the Intelligent Resource Management System can predict failures and take preventive measures. By automating repetitive operational tasks, the Intelligent Resource Management System can minimize human error and improve operational efficiency.

[0043] Furthermore, the Intelligent Resource Management System can establish a proactive failure management framework. Through multimodal data analysis, the Intelligent Resource Management System can detect failure precursors at an early stage. Using predictive models, the System can identify potential failure risks and assess their impact. Furthermore, through automated response processes, the System can take preemptive measures before a failure occurs.

[0044] Furthermore, the Intelligent Resource Management System can provide an intuitive operational environment. It enables easy identification of complex infrastructure configurations through 3D topology-based visualization. It supports rapid decision-making by operations personnel through interactive analysis tools. The Intelligent Resource Management System enables immediate recognition of critical events through context-aware automatic focusing.

[0045] Furthermore, the Intelligent Resource Management System can implement cost-effective resource management. It can establish optimal capacity plans through AI-based resource usage prediction. The Intelligent Resource Management System can maximize the utilization of system resources through dynamic resource allocation. It can prevent unnecessary resource waste through automated cost optimization policies.

[0046] FIG. 1 is a drawing for explaining an intelligent resource management system in one embodiment.

[0047] The intelligent resource management system may include a multimodal data integration processing engine, an LLM-based situation analysis engine, and a 3D topology visualization engine.

[0048] The multimodal data integration processing engine may include a data collector, a multimodal preprocessor, a vector embedding engine, and an integrated vector DB.

[0049] The data collector can collect various forms of infrastructure data in real time and perform initial filtering. The data collector can collect multimodal data in parallel through distributed data collection agents. The data collector can provide a standardized collection interface through connectors specific to each data source. The data collector can perform real-time data validation and deduplication.

[0050] Data collectors may include log collectors, metrics collectors, and image / audio collectors.

[0051] The log collector can collect system and application logs in real time. The log collector can convert logs into structured data through parsers for each log format. The log collector can perform filtering and priority assignment by log level.

[0052] The metric collector can periodically collect system performance indicators. The metric collector can provide a plugin structure for custom metric collection. The metric collector can perform normalization and scaling of the collected metrics.

[0053] The image / audio collector can collect monitoring camera and audio sensor data. The image / audio collector can perform compression and format conversion of media data including video and / or audio. The image / audio collector can perform metadata extraction and indexing.

[0054] The multimodal preprocessor can transform collected multimodal data into a form suitable for LLM processing and normalize it. The multimodal preprocessor can implement preprocessing pipelines specialized for each data type. The multimodal preprocessor can support real-time stream processing and batch processing in parallel. The multimodal preprocessor can automatically perform data quality verification and correction.

[0055] A multimodal preprocessor may include a text preprocessor, an image preprocessor, and a speech preprocessor.

[0056] The text preprocessor can perform text normalization through natural language processing. The text preprocessor can perform text correction using domain-specific term dictionaries. The text preprocessor can perform text classification and keyword extraction.

[0057] The image preprocessor can perform image noise removal and quality improvement. The image preprocessor can perform object detection and segmentation. The image preprocessor can perform image feature extraction and metadata generation.

[0058] A speech preprocessor can perform speech signal noise removal and quality improvement. A speech preprocessor can perform speech-to-text conversion and feature extraction. A speech preprocessor can perform speech event classification and segmentation.

[0059] A vector embedding engine can transform various forms of data into an integrated vector space. It can apply embedding models specialized for each modality. It can perform dimensionality reduction for the efficient processing of high-dimensional vectors. It can perform alignment and integration of multimodal embeddings.

[0060] The vector embedding engine may include a text embedding processor, an image embedding processor, and a speech embedding processor.

[0061] The text embedding handler can generate BERT-based text embeddings. The text embedding handler can apply domain-specific embedding models. The text embedding handler can perform context-based dynamic embeddings.

[0062] The image embedding processor can generate Vision Transformer-based image embeddings. The image embedding processor can perform object feature-based embeddings. The image embedding processor can extract visual context information.

[0063] The speech embedding processor can generate Wav2Vec-based speech embeddings. The speech embedding processor can perform speech feature vectorization. The speech embedding processor can extract speech event patterns.

[0064] The integrated vector database can support the efficient storage and retrieval of multimodal embedding vectors. The integrated vector database can implement a high-performance vector indexing structure. The integrated vector database can support real-time similarity search. The integrated vector database can perform time-series-based vector data management.

[0065] An integrated vector database may include a vector indexer, a search engine, and a data manager.

[0066] Vector indexers can perform efficient indexing of high-dimensional vectors. Vector indexers can implement approximate nearest neighbor search. Vector indexers can support dynamic index updates.

[0067] Search engines can perform multimodal query processing. Search engines can perform real-time similarity calculations. Search engines can support context-based search.

[0068] The data manager can manage the lifecycle of vector data. The data manager can perform storage space optimization. The data manager can provide backup and recovery functions.

[0069] An LLM-based situation interpretation engine may include a context processor, a pattern analyzer, and a prediction engine.

[0070] A context processor can analyze the contextual relationships of multimodal data and comprehensively interpret the situation. It can construct a situational context by analyzing spatiotemporal data relationships. It can enhance the accuracy of situational interpretation by applying domain knowledge-based rules. Furthermore, it can identify the situational context by tracking the causal relationships of real-time event streams.

[0071] A context processor may include a context analyzer, a rule engine, and an event tracker.

[0072] The context analyzer can analyze the correlations of multimodal data. The context analyzer can construct time-based event sequences. The context analyzer can map spatial associations.

[0073] The rule engine can manage domain-specific rules. The rule engine can perform rule-based inference. The rule engine can support dynamic rule updates.

[0074] The event tracker can monitor real-time event flow. The event tracker can perform event chain analysis. The event tracker can detect anomaly patterns.

[0075] The pattern analyzer can extract and learn meaningful patterns from infrastructure operational data. It can generate a learning model by analyzing multimodal patterns in historical failure data. The pattern analyzer can perform matching of learned patterns from real-time data streams. The pattern analyzer can improve the model through automatic learning when new patterns are discovered.

[0076] The pattern analyzer may include a pattern learner, a real-time matcher, and a model optimizer.

[0077] The pattern learner can extract patterns based on historical data. The pattern learner can train a pattern classification model. The pattern learner can continuously evaluate model performance.

[0078] The real-time matcher can perform pattern matching of stream data. The real-time matcher can perform similarity-based pattern identification. The real-time matcher can evaluate the reliability of the matching results.

[0079] The model optimizer can monitor model performance. The model optimizer can perform automatic retraining. The model optimizer can manage model versions.

[0080] The prediction engine can support preemptive responses by predicting changes in system status and performance. The prediction engine can predict performance degradation through time-series prediction models. The prediction engine can establish capacity plans based on resource usage patterns. The prediction engine can perform real-time calculations of failure probabilities.

[0081] The forecasting engine may include a time series forecaster, a capacity planner, and a risk assessor.

[0082] Time series forecasters can perform multivariate time series analysis. Time series forecasters can calculate forecast intervals by analyzing trends and seasonality.

[0083] The capacity planner can calculate the optimal capacity by analyzing resource usage patterns. Based on the calculated optimal capacity, the capacity planner can establish a scaling plan.

[0084] The risk assessor can determine response priorities by identifying risk factors through the calculation of failure probability.

[0085] A 3D topology visualization engine may include a topology generator, interactive analysis tools, and a view controller.

[0086] The topology generator can convert real-time infrastructure configuration information into a 3D visualization model. The topology generator can map logical and physical relationships between infrastructure components into 3D space. The topology generator can perform dynamic rendering of hierarchical structures. The topology generator can reflect real-time configuration changes in the topology.

[0087] The topology generator may include a configuration mapper, a 3D modeler, and a change manager.

[0088] The configuration mapper can collect infrastructure configuration information. The configuration mapper can map the logical relationships of the collected configuration information. The configuration mapper can map the physical connections of the collected configuration information.

[0089] A 3D modeler can create 3D objects, optimize layouts, and perform rendering.

[0090] The change manager can detect configuration changes, update the topology, and manage the change history of the updated topology.

[0091] Interactive analysis tools can support users' intuitive data exploration and analysis. Interactive analysis tools can provide drill-down access to detailed information. Interactive analysis tools can implement real-time metric overlay displays. Interactive analysis tools can support custom filtering and highlighting features.

[0092] Interactive analysis tools may include interaction handlers, data visualizers, and filter managers.

[0093] Interaction handlers can process user input, manage view transitions, and handle selection events.

[0094] Data visualizers can visualize metrics, display alarms, and overlay performance indicators.

[0095] The filter manager manages filter conditions to apply highlighting rules and can save view settings based on the applied highlighting rules.

[0096] The view controller can provide the optimal visualization view based on the situation. The view controller can perform automatic focusing on relevant areas when an event occurs. The view controller can provide visual notifications based on severity. The view controller can implement a playback function for state changes over time.

[0097] A view controller can include a view optimizer, a notification manager, and a timeline controller.

[0098] The view optimizer can optimize the viewpoint, adjust the zoom level, and manage the focus.

[0099] The notification manager can determine notification priorities, apply visual effects, and manage notification history.

[0100] The timeline controller can control time-based playback, manage change history, and support bookmarks.

[0101] Peripheral technology areas of an intelligent resource management system may include an integrated management interface, a monitoring system, and a data management system.

[0102] FIG. 2 is a diagram illustrating a multimodal data integration processing engine in one embodiment.

[0103] The multimodal data integration processing engine may include a data collector, a multimodal preprocessor, a vector embedding engine, and an integrated vector database (Vector DB).

[0104] Data collectors may include log collectors, metrics collectors, and image / audio collectors.

[0105] The log collector can collect various types of log data, including system, application, and security logs, in real time. It enables stable collection without data loss through customized collection agents for each log source. The log collector can automatically perform initial filtering and importance classification of collected logs.

[0106] The metrics collector can periodically collect system performance indicators such as CPU, memory, disk, and network. It supports the collection of various performance indicators, including user-defined metrics. The metrics collector can automatically perform sampling and aggregation of real-time performance data.

[0107] The image / audio collector can collect multimedia data in real time from sources such as CCTVs, thermal cameras, and acoustic sensors. The image / audio collector can perform initial quality checks and format optimization on the collected media data. The image / audio collector can implement streaming processing for the efficient transmission of large-capacity media data.

[0108] A multimodal preprocessor may include a text preprocessor, an image preprocessor, and a speech preprocessor.

[0109] Text preprocessors can transform log data into an analyzable form by structuring and normalizing it. They can extract and classify key information from text by utilizing natural language processing techniques. Furthermore, they can standardize technical terms and abbreviations by employing domain-specific dictionaries.

[0110] The image preprocessor can perform basic preprocessing such as image noise removal, quality enhancement, and resizing. It can automatically identify regions of interest through object detection and segmentation. Furthermore, it can convert images into a searchable format through image metadata extraction and feature point analysis.

[0111] Speech preprocessors can improve analysis quality by removing noise from speech signals and enhancing sound quality. They can perform speech feature vector extraction and speech-to-text conversion. Speech preprocessors can identify meaningful patterns through acoustic event detection and classification.

[0112] The vector embedding engine may include a text embedding processor, an image embedding processor, and a speech embedding processor.

[0113] The text embedding processor can convert text data into vectors by utilizing a BERT-based language model. Through domain-specific learning, the text embedding processor can generate accurate semantic representations of infrastructure-related text. The text embedding processor can capture context-dependent semantic changes through context-based dynamic embedding.

[0114] The image embedding processor can vectorize the visual features of an image through a Vision Transformer model. The image embedding processor can preserve detailed visual information through multi-scale feature extraction. The image embedding processor can generate an integrated image representation that considers the spatial relationships between objects.

[0115] The speech embedding processor can convert the features of speech data (signals) into vectors by utilizing the Wav2Vec model. The speech embedding processor can generate embeddings that consider acoustic characteristics and temporal patterns. The speech embedding processor can generate vector representations that preserve sequence information of speech events.

[0116] An integrated vector database may include a vector indexer, a search engine, and a data manager.

[0117] The vector indexer can manage index structures for the efficient storage and retrieval of high-dimensional vectors. The vector indexer can perform dynamic indexing that supports real-time vector data insertion and updates. The vector indexer can implement optimized algorithms for calculating similarity between vectors.

[0118] Search engines can process multimodal queries to retrieve highly relevant vector data. Search engines can support real-time similarity-based search and filtering. Search engines can evaluate the accuracy and reliability of search results.

[0119] The data manager can manage the lifecycle of vector data and optimize storage space. The data manager can support version control and history tracking of vector data. The data manager can provide data backup and recovery mechanisms.

[0120] FIG. 3 is a diagram illustrating an LLM-based situation analysis engine in one embodiment.

[0121] An LLM-based situation analysis engine may include a context processor, a pattern analyzer, and a prediction engine.

[0122] A context processor may include a context analyzer, a rule engine, and an event tracker.

[0123] The context analyzer can construct the context of the current system situation by analyzing correlations between time, space, and events from multimodal data streams. It can track the causal relationships of related events through dependency mapping between system components. Furthermore, the context analyzer can determine the importance and priority of the situation by utilizing domain knowledge-based ontologies.

[0124] The rule engine manages domain-specific rules necessary for infrastructure operations and can apply appropriate rules depending on the situation. Through rule-based reasoning, the rule engine can diagnose the current situation and derive necessary actions. The rule engine can evaluate the effectiveness of rules through feedback based on execution results and automatically optimize them.

[0125] The event tracker can monitor and track the sequence of events occurring throughout the system in real time. It can predict the potential for cascading failures through event chain analysis. The event tracker can generate preemptive warnings by detecting abnormal event patterns.

[0126] The pattern analyzer may include a pattern learner, a real-time matcher, and a model optimizer.

[0127] The pattern learner can automatically learn the characteristics of normal / abnormal patterns by analyzing past operational data. The pattern learner can model complex system behavior patterns through a deep learning-based time-series pattern recognition model. The pattern learner can discover new patterns and update the model through continuous learning.

[0128] The real-time matcher can identify patterns similar to learned patterns in real-time data streams. It can calculate the confidence of pattern matching results and generate threshold-based alerts. The real-time matcher can recognize complex situations through the simultaneous matching of multiple patterns.

[0129] The model optimizer can continuously monitor and evaluate the performance of a pattern recognition model. It can identify and improve model weaknesses by analyzing false positives and false negatives. The model optimizer can optimize performance by periodically retraining the model based on new data.

[0130] The prediction engine may include a Time Series Predictor, a Capacity Planner, and a Risk Assessor.

[0131] Time series forecasters can predict future values ​​of system performance indicators through multivariate time series analysis. They can construct short- and long-term forecasting models by separating trends and seasonal factors. Time series forecasters can quantify uncertainty by calculating confidence intervals for forecast results.

[0132] The capacity planner can predict future resource demand by analyzing resource usage patterns. It can establish an optimal capacity plan considering business requirements and system constraints. The capacity planner can optimize resource utilization by generating automated scaling policies.

[0133] The risk assessor can evaluate the risk level of a system by comprehensively analyzing various risk factors. It can calculate the impact and probability of occurrence of predicted anomalies. The risk assessor can determine response priorities based on risk and propose automated mitigation measures.

[0134] FIG. 4 is a diagram illustrating a 3D topology visualization engine in one embodiment.

[0135] A 3D topology visualization system may include a topology generator, an interactive analysis tool, and a view controller.

[0136] The 3D topology visualization system may include a Configuration Mapper, a 3D Modeler, and a Change Handler.

[0137] The configuration mapper can perform the function of mapping the logical and physical relationships of infrastructure components into a 3D space. The configuration mapper can automatically collect and classify infrastructure configuration information. The configuration mapper can perform hierarchical structure analysis and relationship mapping. The configuration mapper can generate and manage topology metadata. In this case, the data input to the configuration mapper consists of infrastructure configuration information and connection relationship data, while the data output to the configuration mapper is structured topology mapping data.

[0138] The 3D modeler can perform the function of creating and placing 3D objects based on mapped configuration information. The 3D modeler can create and render 3D objects for each component. The 3D modeler can optimize hierarchical layouts. The 3D modeler can perform object placement based on space efficiency. In this case, the data input to the 3D modeler is structured topology mapping data, and the data output to the 3D modeler is a rendered 3D topology model.

[0139] The change manager can perform the function of detecting infrastructure configuration changes in real time and reflecting them in the topology. The change manager can dynamically update the topology by detecting real-time configuration changes. The change manager can manage and track change history. In this process, the data input to the change manager is real-time configuration change events, and the data output to the change manager is updated topology data.

[0140] Interactive analysis tools may include an interaction handler, a data visualizer, and a filter manager.

[0141] Interaction handlers can perform the function of processing user input and generating appropriate responses. Interaction handlers can handle mouse / keyboard input and recognize and process gestures. Interaction handlers can perform viewport control and navigation. In this case, the data input to the interaction handler is user interaction events, and the data output to the interaction handler is processed interaction commands.

[0142] The data visualizer can perform the function of visually representing system performance metrics and status information on the topology. The data visualizer can display performance metric overlays. The data visualizer can visualize alarm and warning statuses. The data visualizer can process real-time data updates. In this case, the data input to the data visualizer consists of system performance data and status information, while the data output to the data visualizer is visualized overlay data.

[0143] The Filter Manager can perform the function of managing and applying user-defined filtering conditions. The Filter Manager can apply component highlighting rules by setting and managing filter conditions. The Filter Manager can save and load user-specific view settings. In this case, the data input to the Filter Manager consists of filtering conditions and highlighting rules, while the data output to the Filter Manager is the filtered topology view.

[0144] A view controller may include a view optimizer, an alert manager, and a timeline controller.

[0145] The View Optimizer can perform the function of automatically generating and managing views optimized for the current situation. It can automatically optimize the viewpoint, automatically adjust the zoom level, and optimize the focus area. In this process, the data input to the View Optimizer consists of the current situation context and key events, while the data output is the optimized view parameters.

[0146] The notification manager can perform the function of visually representing and managing important events and notifications. It can determine notification priorities, apply visual notification effects, and manage notification history. In this process, the data input into the notification manager consists of system notifications and event data, while the data output is visualized notification effects.

[0147] The Timeline Controller can perform the function of playing back and controlling system state changes over time. It can play back time-based state changes, navigate the timeline, and create and manage bookmarks. In this case, the data input to the Timeline Controller is hourly system state data, and the data output to the Timeline Controller is a timeline-based topology view.

[0148] FIG. 5 is a diagram illustrating the entire process of a core module in one embodiment.

[0149] In the embodiments, data transmission and reception operations between the data collector, multimodal preprocessor, vector embedding engine, integrated vector database, context processor, pattern analyzer, prediction engine, topology generator, and view controller will be described.

[0150] The data collector can transmit multimodal data to the multimodal preprocessor. The data collector can collect multimodal data including log data, performance metrics, image data, and voice data. The multimodal preprocessor can perform data format validation and normalization on the received multimodal data.

[0151] The multimodal preprocessor can transmit preprocessed data to the vector embedding engine. The multimodal preprocessor can transmit normalized data to the vector embedding engine. The vector embedding engine can extract modality-specific features using the data normalized through preprocessing.

[0152] The vector embedding engine can store embedding vectors for extracted modal-specific features in an integrated vector database. The integrated vector database can index the embedding vectors and store the indexed embedding vectors.

[0153] The context processor can query vector data stored in the integrated vector database through vector queries. The context processor can perform context-based searches on vector queries retrieved from the integrated vector database.

[0154] The context processor can transmit context information to the pattern analyzer. The pattern analyzer can perform pattern matching and analysis using the context data.

[0155] The pattern analyzer can transmit the pattern analysis results to the prediction engine. The prediction engine can apply the pattern analysis result data to the prediction model.

[0156] The prediction engine can transmit the analysis results to the topology generator. The topology generator can map the topology using the prediction results obtained through the prediction model.

[0157] The topology generator can send topology data to the view controller. The view controller can use the 3D topology data to perform view optimization and rendering.

[0158] The view controller can provide feedback on user interactions to the prediction engine. The prediction engine can provide feedback on prediction accuracy to the pattern analyzer. The pattern analyzer can provide feedback on pattern mapping results to the context processor. The context processor can provide feedback on context analysis results to the integrated vector database.

[0159] FIG. 6 is a diagram illustrating the process of a core module and a peripheral module in one embodiment.

[0160] In the embodiments, data transmission and reception operations between data storage, core systems, monitoring systems, API gateways, user interfaces, and administrators will be described.

[0161] The administrator can transmit system settings and policies to the core system. The core system can apply system initialization and settings based on monitoring policies and threshold settings.

[0162] Data storage can transmit historical data to the core system. The core system can use historical operational data to load data and train initial models.

[0163] The monitoring system can transmit real-time monitoring data to the core system. The core system can analyze real-time data using system metrics and performance data. The core system can transmit analysis results through the API gateway. The API gateway can convert analyzed insights into an API format and route them.

[0164] The API gateway can send UI update data to the user interface. The user interface can update the dashboard using visualization data.

[0165] The core system can notify the monitoring system of the detection of performance issues. The monitoring system can enhance monitoring using abnormal sign data. The monitoring system can transmit detailed diagnostic data to the core system. The core system can perform in-depth analysis using detailed system information.

[0166] The user interface can forward user requests through the API gateway. The API gateway can validate and transform requests based on user actions. The API gateway can forward analysis requests to the core system. The core system can execute customized analysis based on analysis parameters. The core system can store analysis results in data storage. Data storage can permanently store processed data.

[0167] Administrators can deliver policy updates to the core system. The core system can apply changed configuration information to real-time policies. The core system can report system performance to the administrator. Administrators can evaluate policy effectiveness using performance metrics. The core system can send real-time notifications. Administrators can generate user notifications based on important events.

[0168] According to the embodiments, it can be applied to smart factory operation management. For example, various data sources, such as production line equipment, IoT sensors, and control systems, can be monitored in an integrated manner. Production optimization can be performed by linking real-time quality inspection data with equipment status data. Equipment downtime can be minimized by linking with a predictive maintenance system. Through this, the utilization rate of production equipment can be improved, the quality defect rate can be reduced, and maintenance costs can be lowered.

[0169] According to the embodiments, it can be applied to semiconductor manufacturing process management. For example, real-time monitoring and anomaly detection of the wafer manufacturing process can be performed. Cleanroom environment data and manufacturing equipment data can be integrated and analyzed. Automatic adjustment of process parameters for yield optimization can be performed. Through this, manufacturing yield can be improved, process defect rates can be reduced, and energy consumption can be reduced.

[0170] According to the embodiments, it can be applied to an automotive parts production system. For example, real-time status monitoring and quality control of a multi-product production line can be performed. Integrated control of a robot automation system and inspection equipment can be implemented. Optimization linked with production planning and actual data can be performed. Through this, production lead time can be shortened, work-in-process inventory can be reduced, and worker safety accidents can be reduced.

[0171] FIGS. 7 to 9 are drawings for explaining the system architecture and cloud architecture in one embodiment.

[0172] Referring to Figure 7, the system architecture can be composed of a data source layer, a data collection layer, a processing layer, an analysis layer, and a visualization layer.

[0173] The data source layer may include production facilities, IoT sensors, CCTV, acoustic sensors, quality inspection equipment, environmental sensors, and MES / ERP.

[0174] The data collection layer may include an equipment data collector, a sensor data collector, an image / video collector, an acoustic data collector, and a system log collector.

[0175] The processing layer may include a multimodal data preprocessor, a vector embedding engine, an integrated vector database, and a real-time stream processor.

[0176] The analysis layer may include an LLM situation analysis engine, a pattern analyzer, a prediction engine, a quality analyzer, and a process optimization engine.

[0177] The visualization layer may include 3D topology visualization, real-time dashboards, notification management systems, and report generators.

[0178] Referring to Figure 8, a hybrid cloud architecture is shown, and data can be transmitted and received through a hybrid connection between the cloud and an on-premise IDC.

[0179] Referring to Figure 9, the cloud architecture is shown and may consist of an external layer, an infrastructure layer, a data service layer, a cloud service layer, and a platform service layer.

[0180] FIG. 10 is a diagram illustrating the preprocessing optimization process of multimodal data in one embodiment.

[0181] In the embodiments, the data transmission and reception operations between the Data Collector, Classification Tool, Preprocessing Pipeline, Quality Monitoring System, Optimization Engine, and Data Buffer will be described.

[0182] The multimodal data integration processing engine collects various forms of multimodal data, such as log data, image data, and voice data, through data collection devices and transmits the collected data to a classification tool. When various forms of multimodal data—including log data, image data, and voice data—are input via the classification tool, the engine can automatically detect the type of each data and classify it. The engine can then deliver the classification results by data type to the preprocessing pipeline. The engine can dynamically configure an optimized preprocessing pipeline for each classified data type. In this process, the engine can determine the structure of the pipeline by considering system resource status and processing priority. The preprocessing pipeline includes stages such as data cleaning, normalization, and feature extraction, and the processing parameters for each stage can be automatically adjusted based on data quality metrics.

[0183] The quality monitoring system continuously evaluates the quality of preprocessed data in response to requests for quality inspection and can deliver immediate feedback to the preprocessing pipeline if quality degradation is detected. The quality monitoring system evaluates the quality metrics of the preprocessed data and can transmit the quality evaluation results to the optimization engine. The optimization engine can collect and analyze performance indicators such as processing latency, resource usage, and data quality occurring during the preprocessing process in real time. Based on the collected performance indicators, the optimization engine automatically adjusts preprocessing parameters using machine learning, and in this process, can derive the optimal balance point between processing efficiency and result quality. The quality optimization engine can deliver optimized parameters based on the analysis results of the performance indicators. The multimodal data integration processing engine applies parameters to the preprocessing pipeline and can reprocess the preprocessed data according to the applied parameters.

[0184] The quality monitoring system performs quality checks on reprocessed data and can transmit the results to the optimization engine. The optimization engine can direct parameter optimization. The multimodal data integration processing engine detects specific patterns or anomalies occurring for each data source through the preprocessing pipeline, dynamically adjusts the preprocessing logic, and can automatically generate new preprocessing rules if necessary. The multimodal data integration processing engine can convert data preprocessed or reprocessed through the preprocessing pipeline into a standardized format by undergoing a quality assurance stage. During this process, metadata is automatically generated to record data history and processing information. The multimodal data integration processing engine periodically evaluates and optimizes the performance of the preprocessing pipeline, and can perform continuous improvement by utilizing the processing history and performance data accumulated during this process. Finally, the preprocessed data is converted into a form optimized for LLM processing and passed to the next stage, and detailed logs and quality reports for the entire processing process can be automatically generated.

[0185] FIG. 11 is a diagram illustrating a distributed LLM processing process in one embodiment.

[0186] In the embodiments, data transmission and reception operations between the client, request analyzer, distributed processing node manager, workload balancing scheduler, model / result cache, processing node, and performance monitoring system will be described.

[0187] Clients can request LLM processing from the request analyzer. The request analyzer analyzes processing requests to identify the types of required LLM models, processing priorities, and resource requirements, and performs initial job classification based on this. The request analyzer can transmit resource requirements to the distributed processing manager. The distributed processing manager monitors the status of currently available computing nodes in real time (e.g., CPU / GPU utilization, memory usage, processing queue status, etc.) and maintains performance profiles for each node. The distributed processing node manager checks node status and transmits information on available nodes to the workload balancing scheduler. The workload balancing scheduler can establish an optimal job allocation plan by considering the characteristics of each processing request and the current status of the nodes. In this process, the minimization of processing latency and resource utilization efficiency can be considered simultaneously. At this stage, frequently requested LLM models are pre-loaded to configure a model cache, and the cache replacement policy can be dynamically adjusted by tracking the usage frequency and memory usage of cached models for each node. Intermediate results generated during the distributed processing are stored in the distributed cache system and can be reused for similar subsequent requests, and the validity of the cached results can be periodically verified.

[0188] In the event of a cache hit, the cached result can be returned to the client. In the event of a cache miss, the task scheduler can be notified. The task scheduler can assign tasks to processing nodes. Processing nodes can load the LLM model and report the processing status to the performance monitoring system. The performance monitoring system can transmit performance metrics to the task scheduler. Processing nodes can store processing results. Errors or performance degradation occurring at each processing node are detected in real-time to automatically reassign tasks, and new processing nodes can be dynamically added or removed if necessary. Data transmission and synchronization between processing nodes are performed via optimized protocols, allowing for operation that minimizes processing latency while minimizing network bandwidth usage. The task scheduler collects and analyzes detailed processing history and performance metrics for each processing request, and based on this, can continuously improve task scheduling policies and resource allocation strategies. Regarding security, each processing node applies strict access control and data encryption, and requests containing sensitive data are processed only on specially designated security-enhanced nodes. The final processed result is delivered to the requester after quality verification, and a detailed processing report on the entire process is automatically generated and can be utilized to improve system operations.

[0189] FIG. 12 is a diagram illustrating an intelligent fault prediction and recovery process in one embodiment.

[0190] In the embodiments, data transmission and reception operations between the data collector, anomaly detector, propagation simulator, recovery policy generator, recovery executor, learning engine, and decision manager will be described.

[0191] The data collector can collect various monitoring data in real time, such as network traffic, system logs, performance metrics, and hardware sensor data. The data collector organizes the collected real-time monitoring data into an integrated data stream, which can be utilized as foundational data for anomaly detection. The data collector can transmit the real-time monitoring data to an anomaly detector.

[0192] The anomaly detector can perform real-time analysis. Through a time-series analysis engine, the anomaly detector analyzes signs of anomalies using real-time monitoring data and can transmit anomaly pattern detection alerts to a propagation simulator. The anomaly detector can evaluate potential failure possibilities by performing similarity analysis with past failure patterns, trend analysis, and seasonality analysis. The anomaly detector continuously analyzes real-time data streams through a deep learning-based anomaly detection model to detect signs of anomalies deviating from normal patterns, and can also consider the correlation of multi-sensor data during this process.

[0193] The propagation simulator is an engine that simulates fault propagation. It predicts the potential impact of detected anomalies on the entire system and identifies potential fault propagation paths through dependency graph analysis. For each detected anomaly, the simulator performs a severity assessment, determines alert levels by comparing them to predefined risk thresholds, and can generate immediate alerts if necessary. The propagation simulator can deliver impact analysis results to decision managers, who can then evaluate the severity.

[0194] The decision manager can request immediate recovery from the recovery policy generator if a situation is determined to be an emergency through severity assessment, and request preventive measures if it is determined to be a potential risk. The recovery policy generator establishes an optimal response strategy by considering the characteristics of the detected problem and the system status, and in this process, the history of successes and failures of past recovery operations can be utilized as training data. The recovery policy generator can generate a recovery policy based on learned recovery patterns and instruct the recovery executor to perform recovery operations according to the generated recovery policy.

[0195] The recovery executor sequentially executes automated recovery procedures based on the generated recovery policy, monitors the results of each step, and can switch to an alternative recovery strategy if necessary. The recovery executor can report the results of the recovery procedure execution to the decision manager. Based on the execution results, the decision manager can request an alternative policy from the recovery policy generator if the recovery procedure fails. At this time, all failure detection and recovery processes are recorded in detail to enable an audit trail, and this data can be utilized as training data for analyzing failure patterns and improving recovery strategies. The learning engine learns new failure patterns and successful recovery cases to continuously improve predictive models and recovery strategies, thereby progressively enhancing the system's failure response capabilities. Finally, a comprehensive report on failure prediction and recovery activities is generated and provided to operations personnel, which can be used to establish long-term improvement plans for enhancing system stability.

[0196] The device described above may be implemented as a hardware component, a software component, and / or a combination of a hardware component and a software component. For example, the device and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include a plurality of processing elements and / or a plurality of types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are also possible.

[0197] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0198] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0199] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0200] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 An intelligent resource management system for reducing Total Cost of Ownership (TCO), comprising: a multimodal data integration processing engine that preprocesses multimodal data collected in real time into a form that can be processed by a Large Language Model (LLM); an LLM-based context analysis engine that generates a context-adaptive workflow based on context information interpreted through relationship analysis using a Large Language Model on the preprocessed multimodal data; and a 3D topology visualization engine that visualizes an infrastructure operating environment based on the generated context-adaptive workflow, wherein the LLM-based context analysis engine identifies context information of the current system through correlation analysis between time, space, and events from a stream of the preprocessed multimodal data, and tracks causal relationships of related events through dependency mapping between system components. Claim 2 An intelligent resource management system according to claim 1, wherein the multimodal data integration processing engine collects multimodal data including log data, metric performance indicators, image data, and voice data, and performs preprocessing on the collected multimodal data, wherein the log data includes system logs, application logs, and security logs, and the metric performance indicators include performance indicators including CPU, memory, disk, and network, and wherein the image data is collected through CCTV and thermal imaging cameras, and the voice data is collected through acoustic sensors. Claim 3 In an intelligent resource management system for reducing Total Cost of Ownership (TCO), a multimodal data integration processing engine that preprocesses multimodal data collected in real time into a form that can be processed by a Large Language Model (LLM); and an LLM-based context interpretation engine that generates a context-adaptive workflow based on context information interpreted through relationship analysis using a Large Language Model for the preprocessed multimodal data; An intelligent resource management system characterized by including a 3D topology visualization engine that visualizes an infrastructure operating environment based on the generated situation-adaptive workflow, wherein the multimodal data integration processing engine collects multimodal data including log data, metric performance indicators, image data, and voice data, and performs preprocessing on the collected multimodal data, dynamically configures a pipeline for preprocessing for each data type in the collected multimodal data, performs preprocessing including data cleaning, normalization, and feature extraction on the collected multimodal data through the configured preprocessing pipeline, adjusts parameters of the preprocessing pipeline through quality inspection of the preprocessed data, and reprocesses the preprocessed data through the preprocessing pipeline with adjusted parameters to generate metadata, wherein the log data includes system logs, application logs, and security logs, the metric performance indicators include performance indicators including CPU, memory, disk, and network, the image data is collected through CCTV and thermal imaging cameras, and the voice data is collected through acoustic sensors. Claim 4 An intelligent resource management system according to claim 2 or 3, wherein the multimodal data integration processing engine comprises extracting each embedding vector from the multimodal data on which the preprocessing has been performed, and generating text embeddings from text data through a language model, generating image embeddings from image data through an image processing model, and generating speech embeddings from speech data through a speech recognition model. Claim 5 An intelligent resource management system according to claim 4, wherein the multimodal data integration processing engine indexes each of the extracted embedding vectors and stores them in an integrated vector database, and supports searching through similarity calculation for multimodal queries in the integrated vector database. Claim 6 delete Claim 7 An intelligent resource management system according to claim 1, wherein the LLM-based situation analysis engine derives diagnoses and measures regarding the situation information of the identified current system based on rules required for infrastructure operation, detects abnormal event patterns through sequence monitoring of events occurring throughout the current system to predict the possibility of failure, and generates warning information based on the predicted possibility of failure. Claim 8 An intelligent resource management system according to claim 1, wherein the LLM-based situation analysis engine obtains pattern matching results from a stream of preprocessed multimodal data through a deep learning-based time-series pattern recognition model that has learned normal or abnormal patterns of past infrastructure operation data, and predicts changes in system state and performance based on the obtained pattern matching results to support response information. Claim 9 An intelligent resource management system according to claim 8, wherein the LLM-based situation analysis engine detects abnormal signs based on the acquired pattern matching results, generates a recovery policy based on the characteristics of the detected abnormal signs and the system state, sequentially executes an automated recovery procedure according to the generated recovery policy, and utilizes the executed recovery procedure and the detected abnormal signs as learning data for failure pattern analysis and recovery strategy improvement. Claim 10 An intelligent resource management system according to claim 8, wherein the LLM-based situation analysis engine predicts future resource demand through resource usage pattern analysis, establishes a capacity plan based on business requirements and system constraints based on the predicted future resource demand, and evaluates the risk level of the current system based on the established capacity plan. Claim 11 An intelligent resource management system for reducing Total Cost of Ownership (TCO) comprises: a multimodal data integration processing engine that preprocesses multimodal data collected in real time into a form that can be processed by a Large Language Model (LLM); an LLM-based context analysis engine that generates a context-adaptive workflow based on context information interpreted through relationship analysis using a Large Language Model for the preprocessed multimodal data; and a 3D topology visualization engine that visualizes an infrastructure operating environment based on the generated context-adaptive workflow. The LLM-based context analysis engine is characterized by analyzing the type of LLM model, processing priority, and resource requirements according to an LLM processing request, monitoring the status of currently available computing nodes in real time through the analysis, establishing a task allocation plan based on the monitored status of currently available computing nodes, searching for a pre-loaded cache for LLM models requested above a preset standard according to the established task allocation plan, and adjusting a scheduling policy through the analysis of the processing history and performance metrics of the loaded LLM models using the searched cache. Claim 12 An intelligent resource management system for reducing Total Cost of Ownership (TCO), comprising: a multimodal data integration processing engine that preprocesses multimodal data collected in real time into a form that can be processed by a Large Language Model (LLM); an LLM-based context analysis engine that generates a context-adaptive workflow based on context information interpreted through relationship analysis using a Large Language Model on the preprocessed multimodal data; and a 3D topology visualization engine that visualizes an infrastructure operating environment based on the generated context-adaptive workflow, wherein the 3D topology visualization engine generates a topology by mapping infrastructure components and relationship information of infrastructure components into a three-dimensional space, and generates a three-dimensional object based on configuration information obtained through the mapping. Claim 13 An intelligent resource management system according to claim 12, wherein the 3D topology visualization engine processes interaction commands using user interaction events and visually displays system performance data and status information on the generated topology. Claim 14 An intelligent resource management system according to claim 12, wherein the 3D topology visualization engine provides a visualization view based on situation information and event data, visually represents the event data and system notifications, and controls system state changes over time. Claim 15 An infrastructure integrated control method performed by an intelligent resource management system for reducing Total Cost of Ownership (TCO), comprising: a step of preprocessing multimodal data collected in real time into a form that can be processed by a Large Language Model (LLM); a step of generating a context-adaptive workflow based on context information interpreted through relationship analysis using the Large Language Model for the preprocessed multimodal data; and a step of visualizing an infrastructure operating environment based on the generated context-adaptive workflow, wherein the generating step includes identifying context information of the current system through correlation analysis between time, space, and events from a stream of the preprocessed multimodal data, and tracing the causal relationships of related events through dependency mapping between system components.

Images