An Intent-Based Autonomous Operation System of Edge Device Fleets

Abstract

The present invention relates to an operating system for an edge device fleet, and more specifically, to an efficient operating system for an edge device fleet applicable to various industrial fields such as intelligent transportation systems, construction site safety management, smart retail, and smart factories, which can increase the efficiency of operating a large-scale edge device fleet by converting natural language input by a user into an intent-based formalized policy, automatically distributing AI model applications required for the edge device fleet, monitoring and verifying the status of new AI model applications, and automatically switching them, and also enabling active response to changes in user intent and continuous monitoring of the performance of new AI model applications so that verification of performance degradation, automatic switching, and redeployment can be performed quickly and autonomously.

Description

Technology Field

[0001] The present invention relates to an operating system for an edge device fleet, and more specifically, to an efficient operating system for an edge device fleet applicable to various industrial fields such as intelligent transportation systems, construction site safety management, smart retail, and smart factories, which can increase the efficiency of operating a large-scale edge device fleet by converting natural language input by a user into an intent-based formalized policy, automatically distributing AI model applications required for the edge device fleet, monitoring and verifying the status of new AI model applications, and automatically switching them, and also enabling active response to changes in user intent and continuous monitoring of the performance of new AI model applications so that verification of performance degradation, automatic switching, and redeployment can be performed quickly and autonomously. Background Technology

[0002] Looking at the field of video surveillance using CCTV, traditional video surveillance architectures relied on streaming raw video data from cameras to a central cloud or data center for processing and analysis. This model had serious drawbacks, such as high network bandwidth consumption, increased operating costs, and unacceptable latency (often 500-700 ms) for real-time mission-critical applications like threat detection. Furthermore, the reliance on a stable internet connection created a single point of failure, which could render the entire system inoperable in the event of a network outage.

[0003] Therefore, the emergence of edge computing to address this has partially resolved these issues by moving AI processing to the devices themselves. This has reduced latency to less than 50ms and enabled autonomous operation even during network outages. However, current edge AI systems are generally managed through imperative, manual processes. Consequently, administrators must manually configure each device or group of devices to deploy or update applications; this is time-consuming, error-prone, and does not scale effectively for large fleets. Furthermore, the application lifecycle is static, presenting a limitation where, once deployed, an application remains in that state until further manual intervention.

[0004] <Patent Literature>

[0005] Korean Patent No. 10-1935006 (Registered Dec. 27, 2018) "Following agent included in a multi-agent system and method for controlling the same"

[0006] In the case of the prior art disclosed in the above <Patent Document>, the concept of collecting, storing, and managing agent states is disclosed in the concept of checking state information (output signals, etc.) for each agent (leader / neighbor agent) included in a multi-agent system and deriving control inputs of follower agents based thereon, but this is merely monitoring and managing agent states.

[0007] Another important aspect when deploying edge AI models is managing performance over time. Model performance degrades due to "model drift" or "conceptual drift," where the data encountered in the actual operating environment differs from the training data. Consequently, there is a growing need for monitoring the performance of such edge AI models in operating systems and enabling autonomous functions. The problem to be solved

[0008] The present invention has been devised to solve the above-mentioned problems,

[0009] The objective of the present invention is to provide an autonomous operation system for an intent-based edge device fleet that can increase the efficiency of large-scale edge device fleet operations by converting natural language input by a user into an intent-based formalized policy, automatically distributing AI model applications required for the edge device fleet, monitoring and verifying the status of new AI model applications, and automatically switching them.

[0010] Another objective of the present invention is to provide an autonomous operation system for an intent-based edge device fleet that can actively respond to changes in user intent and continuously monitor the performance of new AI model applications so that verification of performance degradation, automatic switching, and redeployment can be performed quickly and autonomously.

[0011] Another objective of the present invention is to provide an efficient operating system for an edge device fleet applicable to various industrial fields such as intelligent transportation systems, construction site safety management, smart retail, and smart factories. means of solving the problem

[0012] The present invention is implemented by an embodiment having the following configuration to achieve the aforementioned objective.

[0013] According to one embodiment of the present invention, the autonomous operation system of an intent-based edge device fleet according to the present invention comprises: an edge device fleet; an INTENT classification module that tokenizes or labels natural language from voice or text input through a user interface, finally extracts INTENT (intention) or slots through an AI language model, converts them into a structured policy, and provides them; an orchestration unit that performs management, such as the deployment or deletion of an AI model application on the edge device fleet, by comparing a desired state based on the policy of the INTENT classification module with the current state of the edge device fleet; and a loop feedback control unit that monitors the state of an AI model application running on the edge device fleet and, when performance degradation is detected, verifies a new candidate AI model application and automatically switches to it.

[0014] According to another embodiment of the present invention, the orchestration unit in the present invention is characterized by comprising: a state analysis module that compares and analyzes differences by comparing a desired state with the current state of an edge device fleet defined by a policy presented in the INTENT classification module; and an orchestration execution module that maps and deploys or deletes AI model applications for each edge device based on the differences in the state analysis module.

[0015] According to another embodiment of the present invention, the state analysis module of the present invention comprises an individual function differentiation module that differentiates the execution function for each edge device according to a desired state for an edge device fleet defined by a policy presented by the INTENT classification module, and a difference information provision module that compares and analyzes the current state of each edge device with the execution function differentiated by the individual function differentiation module and presents difference information; and the orchestration execution module comprises an application mapping module that maps a specific AI model application to the execution function for each edge device presented by the individual function differentiation module, an application scheduling module that schedules the distribution of the AI ​​model application mapped for each edge device so that the AI ​​model application is provided and executed sequentially, and an INTENT-linked application redeployment module that analyzes the difference regarding the changed function for each edge device when the INTENT (intention) or policy for the edge device fleet is changed, deletes unnecessary AI model applications, and redeploys replacement AI model applications for execution.

[0016] According to another embodiment of the present invention, the loop feedback control unit of the present invention is characterized by comprising: a monitoring agent module that monitors the status of an AI model application running on an edge device fleet; a performance degradation detection module that collects performance indicators of each AI model application and detects performance degradation based on changes in the distribution of performance data; and a verification switching module that verifies a new AI model application to replace the AI ​​model application with degraded performance on the edge device fleet and enables automatic switching.

[0017] According to another embodiment of the present invention, the performance degradation detection module of the present invention comprises: an input / output data collection module that collects metadata regarding the input and output of each AI model application through the monitoring agent module; a first detection module that detects performance degradation of each AI model application by assigning a first-1 weight to the occurrence and degree of an inflection point by comparing the distribution of the input / output metadata of each AI model application with a reference distribution; a second detection module that detects performance degradation by synthesizing a value calculated by assigning a first-2 weight to the state information of the preceding AI model application and a first-3 weight to the existence and number of the subsequent AI model application based on the correlation between edge devices or AI model applications, and a third detection module that finally analyzes the performance degradation of each AI model application by integrating and analyzing the results of the first detection module and the second detection module; and the verification conversion module [requires] the AI ​​model application with degraded performance A candidate model presentation module that presents a candidate for a new AI model application to replace; a first verification module that deploys and executes the new AI model application on a portion of the edge device fleet and evaluates its performance by comparing it with the existing AI model application; an application switching module that deploys and executes the new AI model application on the entire edge device fleet if performance is recognized through the first verification module; and, if performance degradation is detected after monitoring the execution performance of the newly deployed new AI model application, a value calculated by assigning a 2-1 weight to the number of new AI model applications where performance degradation was detected, and a value calculated by assigning a 2-2 weight to the degree of performance degradation,It is characterized by including a second verification module that evaluates whether to roll back to an existing version by synthesizing values ​​calculated by assigning second and third weights to the existence and number of subsequent linked AI model applications of an AI model application in which performance degradation has been detected, and a rollback execution module that performs a review of new AI model application candidates to replace the edge device fleet after rolling back to the existing version of the AI ​​model application according to the result of the second verification module. Effects of the invention

[0018] The present invention can achieve the following effects through the combination and usage relationship of the embodiments described above and the configuration described below.

[0019] The present invention has the effect of increasing the efficiency of large-scale edge device fleet operations by converting natural language input by a user into an intent-based structured policy, automatically distributing AI model applications required for an edge device fleet, monitoring and verifying the status of new AI model applications, and automatically switching them.

[0020] The present invention has the effect of enabling active response to changes in user intent and continuously monitoring the performance of new AI model applications so that verification of performance degradation, automatic switching, and redeployment can be performed quickly and autonomously.

[0021] The present invention provides an efficient operating system for an edge device fleet applicable to various industrial fields such as intelligent transportation systems, construction site safety management, smart retail, and smart factories. Brief explanation of the drawing

[0022] FIG. 1 is a configuration diagram of an operating system according to an embodiment of the present invention. Figure 2 is a detailed configuration diagram of the orchestration section Figure 3 is a detailed configuration diagram of the loop feedback control unit. Specific details for implementing the invention

[0023] Preferred embodiments of an autonomous operating system for an intent-based edge device fleet according to the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that identical components in the drawings are represented by identical reference numerals wherever possible. Unless otherwise specifically defined, all terms in this specification have the same general meaning as understood by a person skilled in the art to which the present invention pertains, and if there is a conflict with the meaning of a term used in this specification, the definition used in this specification shall prevail. Throughout the specification, when a part is described as "comprising" a component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "...part," "...module," etc., described in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware, software, or a combination of hardware and software.

[0025] Referring to FIG. 1 and the like, an autonomous operation system for an intent-based edge device fleet according to an embodiment of the present invention comprises: an edge device fleet (10); an INTENT classification module (30) that tokenizes or labels natural language from voice or text input through a user interface, and finally extracts INTENT (intention) or slots through an AI language model and converts them into a standardized policy and provides them; an orchestration unit (40) that performs management, such as the deployment or deletion of an AI model application on the edge device fleet (10), by comparing a desired state based on the policy of the INTENT classification module (30) with the current state of the edge device fleet (10); and a loop feedback control unit (50) that monitors the state of an AI model application running on the edge device fleet (10), and if performance degradation is detected, verifies a new candidate AI model application and automatically switches to it. As previously explained, the system of the present invention relates to an operating system for an edge device fleet. More specifically, when a user inputs natural language, it converts it into an intent-based formalized policy to automatically distribute AI model applications required for the edge device fleet (10), monitor and verify the status of new AI model applications, and automatically switch to increase the efficiency of operating a large-scale edge device fleet (10). Additionally, it can actively respond to changes in the user's intent and continuously monitor the performance of new AI model applications so that verification of performance degradation, automatic switching, and redeployment can be performed quickly and autonomously. Thus, it presents an efficient operating system for an edge device fleet (10) applicable to various industrial fields such as intelligent transportation systems, construction site safety management, smart retail, and smart factories. The main components and functions for this purpose will be described in detail below.

[0026] For reference, the system of the present invention is a system in which the distribution, execution, and control of various AI model algorithms are performed through a server unit capable of transmitting and receiving information and processing information, and the information transmission and control process is executed in conjunction with a plurality of edge devices that are the target of control, namely an edge device fleet (10). In addition, the system of the present invention provides a mobile / web-based GUI that can combine traditional UI elements such as button clicks and form field inputs with natural language commands. For example, a user can transmit commands through a voice recording button or a text input window in the chat interface within the app, and can check the command history and results in a dashboard format. Furthermore, the same system functions can be utilized by deploying it as a plugin to messenger platforms such as Slack, Microsoft Teams, and Telegram using a chatbot plugin method, by directly inputting natural language commands into the chat window or sending voice messages. Additionally, the system of the present invention can be implemented as a lightweight version to operate in edge environments where network connectivity is limited, and can provide an offline mode to maintain basic functions even in the event of network failure or disconnection.

[0027] The edge device fleet (10) described above refers to a cluster or group of multiple edge devices. For example, in the case of an intelligent traffic control system, it may be a group of multiple smart CCTVs (such as CCTVs equipped with AI functions) installed on each road subject to control, and in the case of a smart factory, it may be a group of edge CCTV devices installed in each process for quality monitoring of each process. Each edge device within this fleet may include a processor (CPU / GPU / NPU), volatile memory (RAM), and non-volatile storage space sufficient to store an operating system and various containerized applications. Additionally, network connectivity for transmitting and receiving information is provided through a high-bandwidth, low-latency communication module such as 5G or Wi-Fi 6, ensuring a stable connection with a central platform (system, INTENT classification module (30), orchestration unit (40), loop feedback control unit (50), etc., to be described later). AI model applications for function implementation can be deployed and executed on each edge device.

[0028] The above INTENT classification module (30) is configured to tokenize or label natural language from voice or text input through a user interface, and finally extract INTENT (intention) or slot through an AI language model and convert it into a standardized policy to provide. Natural language refers to language in the form of speech or text that users use in their daily lives. Natural language processing, or NLP, analyzes the meaning of such natural language so that it can be processed by a computer or algorithm. In particular, in this invention, natural language input through a user interface can be tokenized (a process of dividing into meaningful small units) or labeled (a process of assigning semantic information to tokenized data) to finally extract INTENT (intention) or slot through an AI language model. NLP technologies such as STT (Speech-To-Text), NER (Named Entity Recognition), and SRL (Semantic Role Labeling) can be utilized. Additionally, to enhance multilingual usability including Korean and English, a LangID (Language Identification) module that automatically identifies languages ​​can be additionally utilized. Here, 'INTENT' refers to the intention (purpose, for example, the purpose of 'optimizing traffic flow at an intersection, classifying vehicles by type, and measuring traffic density' in an intelligent traffic control system) that the user intends to achieve through natural language input, and 'Slot' refers to the parameters required for performing INTENT. As for the AI ​​language model used in the INTENT classification module (30), deep learning-based language models may be utilized, such as large-scale language models (LLM) like Few-shot Learning or BERT (Bidirectional Encoder Representations from Transformers) based models.Finally, the INTENT classification module (30) can identify entities, behaviors, and constraints included in the intent based on the extracted INTENT (intent) or slot, convert them into a structured policy that can be read by a device or application, and provide it. (For example, it can generate and provide an official policy expressed in a structured data format such as YAML or JSON).

[0029] The above orchestration unit (40) is configured to perform management such as the deployment and deletion of AI model applications in the edge device fleet (10) by comparing the desired state based on the policy of the INTENT classification module (30) with the current state of the edge device fleet (10). That is, when a user inputs natural language, it converts it into an intent-based formalized policy and performs a function that can increase the efficiency of large-scale edge device fleet operation by implementing autonomous functions such as automatically deploying and deleting AI model applications required in the edge device fleet (10). To this end, the above orchestration unit (40) may more specifically include a state analysis module (410) that compares and analyzes the differences by comparing the desired state with the current state of the edge device fleet (10) for the edge device fleet (10) defined by the policy presented by the INTENT classification module (30), and an orchestration execution module (420) that maps and deploys or deletes AI model applications for each edge device based on the differences in the state analysis module (410).

[0030] The state analysis module (410) is configured to compare and analyze differences by comparing the desired state and the current state of the edge device fleet (10) for the edge device fleet (10) defined by the policy presented in the INTENT classification module (30). To this end, the state analysis module (410) may more specifically include an individual function differentiation module (411) that distinguishes the performance functions of each edge device according to the desired state for the edge device fleet (10) defined by the policy presented in the INTENT classification module (30), and a difference information provision module (412) that presents difference information by comparing and analyzing the current state of each edge device with the performance functions distinguished by the individual function differentiation module (411).

[0031] The above individual function differentiation module (411) is configured to differentiate the functions to be performed by each edge device according to the desired state for the edge device fleet (10) defined by the policy presented in the INTENT classification module (30), that is, it performs the function of differentiating the functions to be performed by each edge device of the edge device fleet (10) in order to implement (execute) the policy presented based on the INTENT (intention) extracted from the natural language input by the user. For example, if the natural language entered by the user is "identify a queue of 5 or more people at the checkout counter during peak hours, track customer movement to create a heatmap, and detect the depletion of stock of a specific display shelf," the individual function differentiation module (411) distinguishes and presents the functions to be performed (identifying the queue, creating a movement tracking heatmap, and detecting the depletion of stock of a specific serial) for each camera (edge ​​device) located at the checkout counter, the camera (edge ​​device) located on the ceiling of the store, and the camera (edge ​​device) located on the display shelf in the edge device fleet (10).

[0032] The difference information providing module (412) is configured to present difference information by comparing and analyzing the current state of each edge device with the functions to be performed identified by the individual function identification module (411). That is, the individual function identification module (411) described above identifies and presents the functions to be performed for each edge device of the edge device fleet (10) in order to implement (execute) the policy presented based on the user's INTENT (intention). At this time, the difference information providing module (412) can present the difference information by analyzing what difference there is between the current state of each edge device and the functions to be performed (presented through the individual function identification module (411)), that is, what functions need to be reinforced.

[0033] The above orchestration execution module (420) is configured to map and distribute or delete AI model applications for each edge device based on differences in the state analysis module (410). To this end, the above orchestration execution module (420) may more specifically include an application mapping module (421) that maps specific AI model applications to the execution functions for each edge device presented in the individual function differentiation module (411), an application scheduling module (422) that schedules the distribution of AI model applications mapped for each edge device so that AI model applications are provided and executed sequentially, and an INTENT-linked application redeployment module (423) that analyzes the differences in the changed functions for each edge device when the INTENT (intention) or policy for the edge device fleet (10) is changed, deletes unnecessary AI model applications, and redeploys replacement AI model applications to execute.

[0034] The application mapping module (421) is configured to map specific AI model applications to the functions to be performed by each edge device as presented in the individual function differentiation module (411). That is, the individual function differentiation module (411) described above differentiates the functions to be performed by each edge device of the edge device fleet (10) in order to implement (execute) the policy presented based on the INTENT (intention), and the difference information provision module (412) analyzes and presents what difference there is between the current state of each edge device and the functions to be performed (presented through the individual function differentiation module (411)), that is, what functions need to be reinforced. Accordingly, the application mapping module (421) maps specific AI model applications to be reinforced for the implementation of the functions to be performed by each edge device based on the information presented by the difference information provision module (412). For example, it can be implemented in a form that derives and matches specific application (software) candidates that need to be reinforced to implement functions specific to each edge device to specific software resources, such as containerized applications.

[0035] The above application scheduling module (422) is configured to schedule the distribution of AI model applications mapped to each edge device so that the AI ​​model applications are provided and executed sequentially. Since the functions to be performed for each edge device of the edge device fleet (10) are distinguished and their respective relationships are formed, the distribution of AI model applications mapped to each edge device of the edge device fleet (10) must also be scheduled and distributed sequentially. To this end, the above application scheduling module (422) performs the function of executing the distribution according to sequential scheduling. For example, the above application scheduling module (422) can schedule the distribution to target edge devices using KubeEdge or a similar technology.

[0036] The above INTENT-linked application redeployment module (423) is configured such that when the INTENT (intention) or policy for the edge device fleet (10) is changed, it analyzes the differences in the changed functions for each edge device, deletes unnecessary AI model applications, and redeploys replacement AI model applications to execute them. In other words, since the user's request for the system, i.e., the intention, can change in various ways depending on the situation, when the user's INTENT (intention) or policy for the system's edge device fleet (10) is changed, the above INTENT-linked application redeployment module (423) immediately analyzes the differences in the changed functions for each edge device in conjunction with that change, deletes unnecessary AI model applications, and redeploys replacement AI model applications to execute them. Thus, when the user inputs only the 'intention (what),' the system itself determines 'how' and autonomously optimizes the system execution. To this end, the INTENT-linked application redistribution module (423) can map specific AI model applications that need to be reinforced to change the function to be performed by each edge device based on the changed function information (according to the changed intention), delete unnecessary AI model applications, and redistribute replacement AI model applications to be executed.

[0037] The loop feedback control unit (50) is configured to monitor the status of an AI model application running on an edge device fleet (10) and, when performance degradation is detected, verify a new candidate AI model application and automatically switch to it. That is, the loop feedback control unit (50) can perform the function of continuously monitoring the performance of a new AI model application newly deployed and running on an edge device fleet (10) so that verification of performance degradation, automatic switching, and redeployment can be performed quickly and autonomously, thereby increasing the operational efficiency of a large-scale edge device fleet (10). To this end, the loop feedback control unit (50) may more specifically include a monitoring agent module (510) that monitors the status of an AI model application running on an edge device fleet (10), a performance degradation detection module (520) that collects performance indicators of each AI model application and detects performance degradation based on changes in the distribution of performance data, and a verification conversion module (530) that verifies a new AI model application to replace the AI ​​model application with degraded performance on the edge device fleet (10) and enables automatic conversion.

[0038] The above monitoring agent module (510) is configured to monitor the status of AI model applications running on the edge device fleet (10), that is, the above monitoring agent module (510) can perform the function of monitoring the status information of AI model applications running on each edge device of the edge device fleet (10) in real time and providing the information.

[0039] The above performance degradation detection module (520) is configured to collect performance indicators of each AI model application and detect performance degradation based on changes in the distribution of performance data. To this end, the above performance degradation detection module (520) more specifically comprises: an input / output data collection module (521) that collects metadata regarding the input and output of each AI model application through the monitoring agent module (510); a first detection module (522) that detects performance degradation of each AI model application by assigning a first-1 weight to the occurrence and degree of an inflection point by comparing the distribution of the input / output metadata of each AI model application with a reference distribution; and a value that detects performance degradation by synthesizing the value calculated by assigning a first-2 weight to the state information of the preceding AI model application and the value calculated by assigning a first-3 weight to the existence and number of the subsequent AI model application, based on the correlation between edge devices or AI model applications, for the subsequent linked AI model application of the AI ​​model application where performance degradation is expected. It may include a second detection module (523) and a third detection module (524) that integrates and analyzes the results of the first detection module (522) and the second detection module (523) to finally analyze the performance degradation of each AI model application.

[0040] The above input / output data collection module (521) is configured to collect metadata regarding the input and output of each AI model application based on information provided through the monitoring agent module (510). By collecting, analyzing, and providing metadata regarding the status of each AI model application running on each edge device, particularly regarding the input and output (e.g., a prediction confidence score based on input and output values), it can provide various status information that can be used as basic data to detect performance degradation of each AI model application described later.

[0041] The first detection module (522) is configured to detect performance degradation of each AI model application by comparing the distribution of input / output metadata of each AI model application with a reference distribution and assigning a first-1 weight to the occurrence and degree of an inflection point. That is, the first detection module (522) is basically based on the comparison of the distribution of input / output metadata of each AI model application with a reference distribution. The reference distribution represents a standard value based on input and output values ​​that serve as the standard for performance during the execution of the AI ​​model application. The reference distribution can be continuously updated through the analysis of accumulated big data on the actual execution results of the AI ​​model application. In comparing the distribution of metadata regarding inputs and outputs collected during the actual execution of the AI ​​model application with such a reference distribution, specifically by assigning a first-1 weight to the occurrence and degree of inflection points relative to the reference distribution in the actual measured metadata distribution, information regarding whether or not inflection points have occurred or the number of times they have occurred, as well as the degree of difference at the inflection points, are all reflected as major weights, thereby enabling them to be utilized as key indicators for detecting performance degradation of each AI model application.

[0042] The second detection module (523) is configured to detect performance degradation by combining the value calculated by assigning a first- and second weight to the state information of the preceding AI model application and the value calculated by assigning a first- and third weight to the existence and number of the subsequent AI model application, based on the relationship between edge devices or AI model applications, to the subsequent linked AI model application of the AI ​​model application where performance degradation is expected. That is, the second detection module is characterized by enabling the relationship between each edge device included in the edge device fleet (10) or between each AI model application executed thereon to be utilized (applied) for detecting and predicting performance degradation for each AI model application, thereby enabling more accurate and efficient detection of performance degradation for AI model applications. To this end, in the process of detecting performance degradation of a specific AI model application, the second detection module (523) calculates a value by assigning a first- and second weight to the state information of an AI model application that is related to the AI ​​model application in the preceding end and assigning a first- and third weight to the information regarding the existence and number of AI model applications related to the subsequent end, and combines this value so that it can be reflected in the evaluation of performance degradation of the AI ​​model application.

[0043] The third detection module (524) is configured to finally analyze the performance degradation of each AI model application by integrating and analyzing the results of the first detection module (522) and the second detection module (523). That is, the third detection module (524) can accurately measure (detect) and provide whether performance degradation has occurred for each AI model application by combining the value calculated by applying a first-1 weight to the information of the first detection module (522) derived by assigning a first-2 weight to the status information of the preceding AI model application based on the relationship between AI model applications provided by the second detection module (523), and the value calculated by assigning a first-3 weight to the existence and number of the subsequent AI model applications.

[0044] The verification conversion module (530) is configured to verify a new AI model application that replaces a performance-degraded AI model application in an edge device fleet (10) and to automatically convert it. To this end, the verification conversion module (530) more specifically comprises: a candidate model presentation module (531) that presents a candidate for a new AI model application to replace a performance-degraded AI model application; a first verification module (532) that distributes and executes the new AI model application to a portion of the edge device fleet (10) and evaluates its performance by comparing it with the existing AI model application; an application conversion module (533) that distributes and executes the new AI model application to the entire edge device fleet (10) if the performance is recognized through the first verification module (532); and, if performance degradation is detected after monitoring the execution performance of the newly distributed AI model application, a value calculated by applying a 2-1 weight to the number of new AI model applications with detected performance degradation, a value calculated by applying a 2-2 weight to the degree of performance degradation, and the AI ​​model with detected performance degradation It may include a second verification module (534) that evaluates whether to roll back to the existing version by synthesizing values ​​calculated by assigning second and third weights to the existence and number of subsequent linked AI model applications of the application, and a rollback execution module () that performs a review of new AI model application candidates to replace the edge device fleet after rolling back to the existing version of the AI ​​model application according to the result of the second verification module (534).

[0045] The above candidate model presentation module (531) is configured to present a new AI model application candidate that replaces an AI model application with degraded performance, and can present new AI model application candidate(s) that can replace an AI model application identified as having degraded performance through the performance degradation detection module (520) described above.

[0046] The first verification module (532) is configured to deploy and execute a new AI model application to a portion of the edge device fleet (10) and evaluate its performance by comparing it with an existing AI model application. That is, the first verification module (532) first deploys and executes the new AI model application to only a portion (for example, about 10%) of the new AI model applications for each edge device (identified as having degraded performance) presented through the candidate model presentation module (531) described above, rather than all of them collectively, and then evaluates its performance by comparing it with an existing AI model application. This allows for the prior reliability of the information regarding the presented new AI model application candidates to be evaluated. After such evaluation (verification), the new AI model application can be deployed to the entire edge device fleet (10) through the application conversion module (533) described later, thereby increasing the overall system stability and reliability.

[0047] The above application switching module (533) is configured to distribute and execute a new AI model application to the entire edge device fleet (10) when performance is recognized through the above first verification module (532). After distributing and executing the new AI model application to some of the edge device fleet (10) through the above first verification module (532) described above and undergoing a verified prior reliability evaluation, the above application switching module (533) distributes and executes the new AI model application to the entire edge device fleet (10).

[0048] The second verification module (534) is configured to evaluate whether to roll back to the existing version by combining the value calculated by applying a second-1 weight to the number of new AI model applications where performance degradation is detected, applying a second-2 weight to the degree of performance degradation, and applying a second-3 weight to the existence and number of subsequent linked AI model applications of the AI ​​model applications where performance degradation is detected, after monitoring the execution performance of the newly deployed new AI model application and detecting performance degradation. That is, the second verification module (534) is configured to perform monitoring and evaluation functions to enable an immediate automated rollback if the result is not improved compared to the existing model or is rather degraded, after the execution performance of the newly deployed new AI model application is continuously monitored immediately after the new AI model application is deployed and executed across the entire edge device fleet (10) through the application switching module (533). To this end, the second verification module (534) can evaluate whether to roll back to the existing version by combining the value calculated by applying a 2-1 weight to the number of new AI model applications where performance degradation is detected during the monitoring process of the execution performance of new AI model applications, the value calculated by applying a 2-2 weight to the degree of performance degradation, and the value calculated by applying a 2-3 weight to the existence and number of subsequent linked AI model applications of the AI ​​model applications where performance degradation is detected. This allows for the evaluation to additionally reflect the relationship between the application or edge device, particularly the degree of impact on subsequent linked AI model applications that are most significantly affected by the performance degradation of the application, in addition to reflecting the evaluation of the new AI model application itself where performance degradation is detected, thereby increasing the overall system stability and reliability.

[0049] The above rollback execution module () is configured to perform a review of a new AI model application candidate to replace the existing version of the AI ​​model application for the edge device fleet after rolling back according to the result of the second verification module (534). That is, when a rollback to the existing model is determined according to the result of the second verification module (534), the rollback execution module () not only executes a rollback to the existing version of the AI ​​model application for the edge device fleet, but also immediately afterwards performs a review of a new AI model application candidate that can replace the existing model, thereby enabling the existing model (evaluated as having degraded performance) to be quickly replaced with a new application.

[0051] Although the applicant has described various embodiments of the present invention above, such embodiments are merely examples of implementing the technical concept of the present invention, and any modification or alteration that implements the technical concept of the present invention should be interpreted as falling within the scope of the present invention. Explanation of the symbols

[0052] 10: Edge Device Fleet 30: INTENT Classification Module 40: Orchestration Department 410: State Analysis Module 411: Individual Function Differentiation Module 412: Difference Information Provision Module 420: Orchestration Execution Module 421: Application Mapping Module 422: Application Scheduling Module 423: INTENT-integrated Application Redeployment Module 50: Loop Feedback Control Unit 510: Monitoring Agent Module 520: Performance degradation detection module 521: Input / output data collection module 522: 1st detection module 523: 2nd detection module 524: 3rd detection module 530: Verification Transformation Module 531: Candidate Model Presentation Module 532: Verification Module 1 533: Application Transition Module 534: Verification Module 2

Claims

Claim 1 delete Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 Edge device fleet; an INTENT classification module that tokenizes or labels natural language from voice or text input through a user interface, finally extracts INTENTs or slots through an AI language model, converts them into a standardized policy, and provides them; an orchestration unit that compares a desired state based on the policy of the INTENT classification module with the current state of the edge device fleet, and performs management such as the deployment or deletion of AI model applications on the edge device fleet; It includes a loop feedback control unit that monitors the status of AI model applications running on an edge device fleet and, if performance degradation is detected, verifies a new candidate AI model application and automatically switches to it; wherein the orchestration unit includes a status analysis module that compares and analyzes differences by comparing a desired state with the current state of an edge device fleet defined by a policy presented by the INTENT classification module, and an orchestration execution module that maps and deploys or deletes AI model applications per edge device based on the differences from the status analysis module; wherein the loop feedback control unit includes a monitoring agent module that monitors the status of AI model applications running on an edge device fleet, a performance degradation detection module that collects performance indicators of each AI model application and detects performance degradation based on changes in the distribution of performance data, and a verification switching module that verifies a new AI model application to replace the AI ​​model application with degraded performance on the edge device fleet and automatically switches to it; and wherein the performance degradation detection module, through the monitoring agent module, metadata regarding the input and output of each AI model application A data collection module for collecting input / output data, a first detection module for detecting performance degradation of each AI model application by assigning a first-1 weight to the occurrence and degree of an inflection point by comparing the distribution of input / output metadata of each AI model application with a reference distribution, andBased on the correlation between edge devices or AI model applications, the second detection module detects performance degradation by synthesizing a value calculated by assigning first and second weights to the state information of the preceding AI model application and assigning first and third weights to the existence and number of the subsequent AI model application to the value calculated for the subsequent linked AI model application of an AI model application expected to experience performance degradation, and a third detection module that finally analyzes the performance degradation of each AI model application by integrating and analyzing the results of the first and second detection modules. The verification conversion module includes a candidate model presentation module that presents a candidate for a new AI model application to replace the AI ​​model application with degraded performance, a first verification module that distributes and executes the new AI model application on a portion of the edge device fleet and evaluates its performance by comparing it with the existing AI model application, an application conversion module that distributes and executes the new AI model application on the entire edge device fleet if performance is recognized through the first verification module, and if performance degradation is detected after monitoring the execution performance of the newly distributed AI model application, the performance degradation An autonomous operation system for an intent-based edge device fleet, characterized by comprising: a second verification module that evaluates whether to roll back to an existing version by synthesizing a value calculated by applying a second-1 weight to the number of detected new AI model applications, a value calculated by applying a second-2 weight to the degree of performance degradation, and a value calculated by applying a second-3 weight to the existence and number of subsequent linked AI model applications of the AI ​​model application with detected performance degradation; and a rollback execution module that performs a review of new AI model application candidates to be replaced after rolling back to the existing version of the AI ​​model application for the edge device fleet according to the evaluation result of the second verification module.

Images