Method for creating artificial intelligence learning dataset in an intelligent system

The method and device automate data collection in intelligent systems by using a learning data profile and management unit, addressing the inefficiency of manual collector development and enhancing reusability and efficiency in data collection for UAVs.

KR103047195B1Active Publication Date: 2026-11-11AGENCY FOR DEFENSE DEV
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
KR1020220117110
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2022-09-16
Publication Date
2026-11-11
Estimated Expiration
2042-09-16

AI Technical Summary

Technical Problem

Existing methods for constructing training datasets in intelligent systems, particularly for unmanned aerial vehicles (UAVs), require manual development of data collectors for each mission, limiting reusability and efficiency in data collection.

Method used

A method and device for constructing an artificial intelligence learning dataset using a learning data profile and management unit that automates data collection, allowing for the reuse of existing methods by referencing dependencies and collecting data based on specified profiles.

Benefits of technology

Enables automated data collection and reuse of existing methods across varying environments, reducing the need for manual development of new collectors and enhancing the efficiency of data collection for intelligent systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112022097456559-PAT00007_ABST
    Figure 112022097456559-PAT00007_ABST
Patent Text Reader

Abstract

A method for constructing an artificial intelligence learning dataset in an intelligent system according to an embodiment of the present invention comprises: a step of generating situational data through an intelligent framework; a step of determining target data to be used as learning data for machine learning among the situational data; a step of obtaining raw data by collecting the determined target data; a step of processing the raw data into source data by setting the format or size suitable for machine learning and removing duplicates; a step of generating labeled data by performing data labeling on the source data; and a step of constructing a learning dataset including the source data and the labeled data. The determined target data is data determined among the situational data based on items included in the selected learning data profile, wherein when the situational data is generated, a learning data profile is automatically selected through a learning data management unit based on the generated situational data, and the raw data is acquired by automatically collecting and storing the determined target data through the learning data management unit when the target data is determined. Accordingly, the reusability of the data collection method according to the data collection environment can be secured and data collection can be automated.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to a method for constructing an artificial intelligence learning dataset in an intelligent system, and more specifically, to a method for constructing an artificial intelligence learning dataset in an intelligent system that secures the reusability of a data collection method according to the data collection environment and automates data collection by collecting learning data using a learning data profile and a learning data management unit. Background Technology

[0002] With the rapid advancement of machine learning technology in recent years, it is being applied across various fields. In particular, there is a growing trend of research utilizing machine learning to implement specific functions in areas that require the organic integration of multiple technologies, such as intelligent systems that perceive changing environments and execute selected actions through decision-making. Representative applications of machine learning in intelligent systems include research on distinguishing people through voice or images, recognizing facial expressions, and extracting emotions. Furthermore, research is also being conducted on policy learning required for meta-planning of task management functions to support decision-making capabilities, as well as reinforcement learning for decision-making functions. To utilize machine learning in the functions that constitute intelligent systems in this manner, the construction of training data is essential.

[0003] In acquiring data for building training datasets, necessary data is obtained by directly generating it or securing it from existing systems. However, since various types of data are generated depending on the environment in which data is to be acquired, data collection methods specialized for each environment are required.

[0004] Meanwhile, regarding intelligent systems for controlling autonomous mission-performing unmanned aerial vehicles (UAVs), since UAVs must be able to perform improved missions based on their past mission records, it is necessary to store mission execution information for learning and reflect the results of the learning model in decision-making functions in real time. To this end, it is necessary to collect training data produced by the intelligent system, and since the data produced varies depending on the mission performed by the UAV, data collection specialized for each mission is required. However, if a training data collector is developed for each mission, there is a limitation in that the data collector must be developed manually every time the target data changes.

[0005] Accordingly, for the collection of training data produced within an intelligent system, there is a need for a method that reuses existing collection methods for data commonly generated by mission, and enables the collection of desired data through a specification describing a list of collection targets without the need to develop a separate data collector for data generated differently by mission. Prior art literature

[65535] U.S. Registered Patent No. 11,263,188 U.S. Published Patent No. 2019-0220698 U.S. Published Patent No. 2021-0304057 The problem to be solved

[0006] The present invention aims to solve the problems of the prior art as described above by providing a method for constructing an artificial intelligence learning dataset in an intelligent system that automates data collection and ensures the reusability of the data collection method according to the data collection environment through the collection of learning data utilizing a learning data profile and a learning data management unit. means of solving the problem

[0007] A method for constructing an artificial intelligence learning dataset in an intelligent system according to the first feature of the present invention comprises: a step of generating situation data through an intelligent framework; a step of determining target data to be used as learning data for machine learning among the situation data; a step of collecting the determined target data to obtain raw data; a step of processing the raw data into source data by setting the format or size suitable for machine learning and removing duplicates; a step of generating labeled data by performing data labeling on the source data; and a step of constructing a learning dataset including the source data and the labeled data. The determined target data is data determined among the situation data based on items included in the selected learning data profile, wherein when the situation data is generated, a learning data profile is automatically selected through a learning data management unit based on the generated situation data, and the raw data is acquired by automatically collecting and storing the determined target data through the learning data management unit when the target data is determined.

[0008] An artificial intelligence learning dataset construction device in an intelligent system according to the second feature of the present invention comprises: a learning data management unit; a communication unit for transmitting and receiving signals; a memory; and a processor for functionally controlling the transmitting and receiving unit and the memory. The processor comprises: an operation of generating situation data through an intelligent framework; an operation of determining target data to be used as learning data for machine learning among the situation data; an operation of collecting the determined target data to obtain raw data; an operation of processing the raw data into source data by setting the format or size suitable for machine learning and removing duplicates; an operation of generating labeled data by performing data labeling on the source data; and an operation of constructing a learning dataset including the source data and the labeled data. The determined target data is data determined among the situation data based on items included in the selected learning data profile, wherein when the situation data is generated, a learning data profile is automatically selected through the learning data management unit based on the generated situation data, and the raw data is acquired by automatically collecting and storing the determined target data through the learning data management unit when the target data is determined.

[0009] A program of instructions that can be executed by a digital processing device is tangibly implemented to provide for the construction of an artificial intelligence learning dataset in an intelligent system according to the third feature of the present invention, and a recording medium readable by a digital processing device is characterized by having a program recorded thereon for executing a method for constructing an artificial intelligence learning dataset in an intelligent system according to the first feature of the present invention on a computer. Effects of the invention

[0010] The method for constructing an artificial intelligence learning dataset in an intelligent system according to an embodiment of the present invention provides the following effects.

[0011] Through the "Dependencies" item, which indicates the reference relationship between a training data profile and other profiles, the training data management unit can automatically collect related data by referencing profiles described in "Dependencies" even if they are not directly described in the profile, thereby enabling the reuse of existing data collection methods in new environments.

[0012] By utilizing the training data management unit to automatically select a training data profile based on context data generated through an intelligent framework, and automatically collecting desired data based on the selected profile, it is possible to collect desired data in new environments without developing a separate data collector. Brief explanation of the drawing

[0013] FIG. 1a is a diagram illustrating an intelligent software framework for an autonomous mission of an unmanned aerial vehicle according to a disclosed embodiment. FIG. 1b is a table showing the protocol of a common knowledge repository included in an intelligent software framework for autonomous missions of unmanned aerial vehicles according to a disclosed embodiment. FIG. 2a is a table illustrating the configuration of a learning data profile according to a disclosed embodiment. FIG. 2b is a drawing showing an exemplary description of some items of a learning data profile according to a disclosed embodiment. FIG. 3 is a conceptual structural diagram illustrating a learning data management unit according to a disclosed embodiment. FIG. 4 is a block diagram conceptually illustrating an artificial intelligence learning dataset construction device in an intelligent system according to a disclosed embodiment. FIG. 5 is a flowchart illustrating a method for constructing an artificial intelligence learning dataset in an intelligent system according to a disclosed embodiment. FIG. 6a is a table summarizing data generated through the performance of a basic mission by an unmanned aerial vehicle according to a disclosed embodiment. FIG. 6b is a diagram showing a learning data profile configured exemplarily to reflect a threat situation of an unmanned aerial vehicle according to one disclosed embodiment. FIG. 7a is a table summarizing data generated through the performance of an unmanned aerial vehicle tracking mission according to a disclosed embodiment. FIG. 7b is a diagram showing an exemplary training data profile configured to reflect the movement path of an unmanned aerial vehicle when tracking a target according to one disclosed embodiment. FIG. 8a is a table summarizing data generated through the performance of an unmanned aerial vehicle strike mission according to a disclosed embodiment. FIG. 8b is a diagram showing an exemplary learning data profile configured to reflect the result of a strike when an unmanned aerial vehicle strikes a target according to one disclosed embodiment. FIG. 9 is a table showing the number of knowledge generated and the number of knowledge collected per mission according to a disclosed embodiment. Specific details for implementing the invention

[0014] The present invention will be described in detail below with reference to the embodiments and drawings. However, the following description is not intended to limit the present invention to specific embodiments, and detailed descriptions of related known technologies are omitted if it is determined that such detailed descriptions may obscure the essence of the present invention.

[0015] There are conventional technologies for automatically collecting data. First, CREME is a framework developed for constructing training datasets for machine learning used in intrusion detection systems for networks such as IoT. It executes attack scenarios modeling various network attacks based on user settings, collects network packets or system logs generated during the attack, extracts features from the collected data, assigns labels, and evaluates the datasets composed of these elements. However, additional system development is required when new data needs to be collected, and since it targets data collection within the network domain, it is not suitable for data collection in intelligent systems.

[0016] AUTODATA is a system designed to automate the construction and labeling of datasets for deep learning. It is a reinforcement learning-based automated data retrieval system that utilizes information existing on the web to automatically build datasets for deep learning and provide them to AutoML. However, since the training data this system intends to collect is limited to web-based data and the type of training model to be utilized is fixed to AutoML, it is impossible to construct datasets for various training models to collect training data for intelligent systems.

[0017] FIG. 1a is a diagram illustrating an intelligent software framework for an autonomous mission of an unmanned aerial vehicle according to a disclosed embodiment.

[0018] In the disclosed embodiment, the method for constructing an artificial intelligence learning dataset in an intelligent system can be used more specifically in an intelligent software framework for unmanned aerial vehicle control, and an intelligent system for an unmanned aerial vehicle autonomous mission is described below.

[0019] The operating environment of an unmanned aerial vehicle (UAV) contains various information required for the mission, and the execution of appropriate actions in response to environmental changes is required. Standard operational guidelines, such as field manuals, exist for the actions performed. A characteristic of UAV missions—such as reconnaissance and tracking—is that the actual actions performed remain similar regardless of the environment. Furthermore, the UAV control system must be able to intelligently utilize internal modules, such as cameras and GPS, by integrating them, and must be able to improve mission execution capabilities based on past mission records.

[0020] Due to the characteristics of the unmanned aerial vehicle (UAV) operating environment described above, an intelligent software framework for UAV control has the following requirements. For mission execution, it must be able to understand various environmental information through intelligent components such as perception, reasoning, and memory; select appropriate actions in response to real-time environmental changes; and convert standard operational guidelines into knowledge that can be applied to the intelligent software framework. Furthermore, it must be able to utilize existing missions when developing similar missions and possess scalability for mission development. Support for the integration and reuse of various modules within the UAV is required, mission execution information must be stored for use in learning, and the learned results must be reflected in real-time decision-making.

[0021] The intelligent framework supports the UAV in understanding the real-time environment by providing real-time situational reasoning capabilities based on information acquired from sensors or perception functions grounded in situational reasoning rules. Furthermore, it selects appropriate tasks in response to changes in the real-time environment through a task management function that implements a procedural reasoning system. Additionally, all data used is converted into knowledge through ontologies, and an architecture is applied that enables the application and execution of mission packages composed of necessary situations and tasks, allowing for the reuse or extension of previously developed missions. Moreover, it provides external connectivity interfaces for linking with various sensors and actuators mounted on the UAV, as well as a knowledge repository for sharing or storing knowledge generated during mission execution.

[0022] A mission package, consisting of operational and situational knowledge required for an unmanned aerial vehicle to perform a mission, can implement the mission by referencing other mission packages through meta-information. Accordingly, this provides a method to refine or specify the composition of a mission by reusing the knowledge information constituting an existing mission to expand the knowledge required for performing a new mission. In other words, this means that the knowledge information constituting a specific mission is utilized identically in the new mission that expands upon it, and that such knowledge can be generated directly during mission execution.

[0023] The knowledge repository of the intelligent software framework consists of a knowledge query interface that supports CRUD functions and an event-driven interface, which is an asynchronous interface that performs real-time event generation and delivery. The protocols for performing each function are as illustrated in FIG. 1b.

[0024] FIG. 1b is a table showing the protocol of a common knowledge repository included in an intelligent software framework for autonomous missions of unmanned aerial vehicles according to a disclosed embodiment.

[0025] The protocol includes assert for storing knowledge, match and retrieve for querying, retract for removing, and update for modifying, as well as subscribe and notify for generating and delivering real-time events. The subscribe protocol takes as arguments a rule consisting of the event generation condition and the knowledge to be received when the event is generated. When a subscription to the knowledge store is made using the above rule, if the event generation condition is met, new knowledge is generated in the knowledge store and delivered via the notify protocol.

[0026] The intelligent framework utilizes the aforementioned knowledge repository to employ a learning model through the following process. First, raw data necessary for learning is acquired from data related to the environment obtained through sensors and high-level context-related knowledge data generated by the intelligent framework through such data acquisition. This raw data is then processed into a format or size suitable for machine learning to obtain source data. After assigning labeling data to the processed source data, learning is performed. The model trained in this manner is utilized during decision-making by the intelligent framework; once an action is determined through this decision, it causes a change in the environment. This change is detected by sensors, processed back into the form of knowledge, and the aforementioned process is repeated in real-time while the unmanned aerial vehicle performs its mission.

[0027] In order to perform decision-making using a learning model, training data must be collected in real time. Accordingly, the method for constructing an artificial intelligence training dataset in an intelligent system according to the disclosed embodiment can be implemented through a profile that describes the data to be collected and a framework that collects data within the intelligent system through said profile.

[0028] FIG. 2a is a table illustrating the configuration of a learning data profile according to a disclosed embodiment.

[0029] In the disclosed embodiment, a learning data profile and a learning data management unit that automatically collects learning data using the learning data profile may be used to construct a dataset utilizing data generated by an intelligent system, and the intelligent system may be an intelligent system for controlling an unmanned aerial vehicle.

[0030] A learning data profile is a specification describing the data to be collected, and can specify the target to be collected and the storage format of the target among the data generated through the performance of a task (i.e., data reflecting a real-time environment or situation, in this disclosure the term 'situational data' is used in the above sense).

[0031] The training data management unit automatically collects and stores data when specified data is generated in the intelligent system based on the training data profile. For reference, the training data management unit may also be referred to as a 'training data manager' or an 'automatic training data collection framework.' It should be noted that as long as it functions to automatically collect desired data based on the training data profile, the technical characteristics described above are not affected by these differences in terminology.

[0032] Again, the training data profile is a specification describing the data to be collected, consisting of items that specify the task in which the data is generated and the data to be collected, as well as when, where, and in what format the collected data will be stored.

[0033] Referring to FIG. 2a, in the disclosed embodiment, the items of the learning data profile may include "ProfileID", "MissionID", "Dependencies", "Predicates", "Template", "Trigger", and "Storage", and the contents of each of the items may, more specifically, be a profile ID, a mission for data collection, another profile to be referenced, knowledge to be included in the data to be collected, a data storage format, a time of data collection, and a data storage location.

[0034] FIG. 2b is a drawing showing an exemplary description of some items of a learning data profile according to a disclosed embodiment.

[0035] Referring to the top example of Fig. 2b, "Predicates" is an item for specifying data to be collected from among the data generated by the intelligent system. Knowledge containing variables may be described here.

[0036] Referring to the example of the stoppage in Fig. 2b, "Template" is an item for specifying the format of the training data. The data format described here is of JSON type, where the key can be defined directly by the profile developer, and the value can be a variable included in predicates.

[0037] Referring to the bottom example of Fig. 2b, there are two types of "Trigger": event and cyclic. The event type consists of items for specifying specific data and can be used when data is to be collected upon the generation of such data. The cyclic type consists of items for describing the period for collecting data and can be used when data is to be collected at regular intervals.

[0038] "Dependencies" (not shown) is an item used to indicate reference relationships with other profiles. Based on these reference relationships, the training data management unit described later can automatically collect related data by referencing profiles described in Dependencies, even if they are not directly described in the selected profile. Accordingly, existing collection methods can be reused even when collecting data in a new environment.

[0039] FIG. 3 is a conceptual structural diagram illustrating a learning data management unit according to a disclosed embodiment.

[0040] In the disclosed embodiment, the training data management unit selects a suitable profile according to the mission being performed by the unmanned aerial vehicle by referring to the MissionID described in the training data profile. Subsequently, it collects suitable data based on the items described in the profile. If a profile is described in the Dependencies item, data for that profile is also collected simultaneously.

[0041] If the type of the Trigger described in the selected profile is event, the Predicates described in the Trigger are subscribed to the public knowledge repository. Through this, when target data is generated in the public knowledge repository, the training data management unit receives a notification and collects the related data. If the type of the Trigger is cyclic, data is retrieved at regular intervals according to the period described in the Trigger.

[0042] During data collection, knowledge described in Predicates is requested from the public knowledge repository via the retrieve protocol, and the public knowledge repository returns knowledge that matches the requested knowledge. If there are variables with the same name in the knowledge described in Predicates or in the predicates described in Triggers, knowledge is retrieved where all corresponding variables contain the same value. Training data is created by utilizing the data included in the collected knowledge and the data format described in Template, and the training data is generated through mapping between the variables included in Predicates and the variables included in Template. The generated training data is stored in the repository based on the information described in the Storage item.

[0043] In this way, by utilizing the training data management unit, desired data can be automatically collected through the addition of training data profiles, without the need to implement changes to the data collection function depending on the collection environment and training model.

[0044] FIG. 4 is a block diagram conceptually illustrating an artificial intelligence learning dataset construction device in an intelligent system according to a disclosed embodiment.

[0045] The device (400) for constructing an artificial intelligence learning dataset in an intelligent system may include an electronic device including an AI module capable of performing AI processing or a server including said AI module, and may include a processor (410), memory (450) and / or a communication unit (470). The device (400) is a computing device capable of learning a neural network and may be implemented as various electronic devices such as a server, a desktop PC, a laptop PC, a tablet PC, etc.

[0046] The processor (410) can train a neural network using a program stored in memory (450). In particular, the processor (410) can train a neural network for recognizing data related to a real-time environment or situation. Here, the neural network for recognizing data related to a real-time environment or situation can be designed to simulate the structure of the human brain on a computer and may include multiple network nodes having weights that simulate the neurons of a human neural network. The multiple network nodes can exchange data according to their respective connection relationships to simulate the synaptic activity of neurons that transmit and receive signals through synapses. Here, the neural network may include a deep learning model that has been developed from a neural network model.

[0047] Meanwhile, a processor that performs the functions described above may be a general-purpose processor (e.g., CPU), but may also be an AI-dedicated processor for artificial intelligence learning (e.g., GPU).

[0048] The memory (450) can store various programs and data required for the operation of the artificial intelligence learning dataset building device (400) in an intelligent system. The memory (450) can be implemented as non-volatile memory, volatile memory, flash memory, hard disk drive (HDD), or solid-state drive (SDD). The memory (450) is accessed by the processor (410), and the processor (410) can perform reading / writing / modification / deletion / updating of data. Additionally, the memory (450) can store a neural network model (e.g., machine learning model (460)) generated through a learning algorithm for data classification / recognition according to the disclosed embodiment.

[0049] Meanwhile, the processor (410) may include a data learning unit (420) that learns a neural network for data classification / recognition. The data learning unit (420) can learn criteria regarding which training data to use to determine data classification / recognition, and how to classify and recognize data using the training data. The data learning unit (420) can learn a machine learning model by acquiring training data to be used for learning and applying the acquired training data to the machine learning model.

[0050] The data learning unit (420) may be manufactured in the form of at least one hardware chip and mounted on the device (400). For example, the data learning unit (420) may be manufactured in the form of a dedicated hardware chip for artificial intelligence (AI), or may be manufactured as part of a general-purpose processor (CPU) or a graphics-dedicated processor (GPU) and mounted on the device (400). Additionally, the data learning unit (420) may be implemented as a software module. If implemented as a software module (or a program module containing instructions), the software module may be stored on a non-transitory computer-readable media. In this case, at least one software module may be provided by an operating system (OS) or by an application.

[0051] The data learning unit (420) may include a learning data acquisition unit (430) and a model learning unit (440). The learning data acquisition unit (430) may acquire learning data required for a neural network model for classifying and recognizing data. For example, the learning data acquisition unit (430) may acquire, as learning data, data related to an environment or situation to be input into a neural network model, or data collected and processed based on a profile.

[0052] The model learning unit (440) can learn to have a judgment criterion regarding how a neural network model classifies a predetermined data using the acquired learning data. At this time, the model learning unit (440) can train the neural network model through supervised learning, which uses at least a portion of the learning data as a judgment criterion; through unsupervised learning, which discovers a judgment criterion by learning on its own using the learning data without supervision; or through reinforcement learning, which uses feedback on whether the result of the situation judgment based on learning is correct. Additionally, the neural network model can be trained using a learning algorithm including error back-propagation or gradient descent.

[0053] When the neural network model is trained, the model training unit (440) can store the trained neural network model in memory (450). The model training unit (440) may also store the trained neural network model in the memory of a server connected to the device (400) via a wired or wireless network.

[0054] The data learning unit (420) may include a training data preprocessing unit (not shown) and a training data selection unit (not shown) to improve the analysis results of the constructed model or to save resources or time required to create the constructed model. Additionally, the data learning unit (420) may further include a model evaluation unit (not shown) to improve the analysis results of the neural network model.

[0055] The learning data preprocessing unit can preprocess the acquired data so that the acquired data can be used for learning to make situational judgments, and the learning data selection unit can select the data necessary for learning from the learning data acquired by the learning data acquisition unit (430) or the learning data preprocessed by the preprocessing unit. The model evaluation unit inputs evaluation data into the neural network model, and if the analysis result output from the evaluation data does not satisfy a predetermined standard, it can cause the model learning unit (440) to learn again. In this case, the evaluation data may be predefined data for evaluating the recognition model.

[0056] The communication unit (470) can transmit the AI ​​processing results by the processor (410) to an external electronic device. Here, the external electronic device may be defined as a terminal or a client. Additionally, the device (400) may be implemented via a server or a network.

[0057] Meanwhile, the artificial intelligence learning dataset building device (400) in the intelligent system illustrated in FIG. 4 was described by functionally separating it into a processor (410), memory (450), and communication unit (470), but the aforementioned components may be integrated into a single module and referred to as an AI module.

[0058] Additionally, in the disclosed embodiment, the artificial intelligence learning dataset building device (400) in the intelligent system may include a learning data management unit (not shown), and the learning data management unit may select a learning data profile by reflecting a situation recognized by the intelligent system to automate the process of collecting data related to an environment or situation, collect data from a knowledge repository based thereon to generate learning data, and store the generated learning data in a memory (450), etc.

[0059] FIG. 5 is a flowchart illustrating a method for constructing an artificial intelligence learning dataset in an intelligent system according to a disclosed embodiment.

[0060] In the disclosed embodiment, the method for constructing an artificial intelligence learning dataset in an intelligent system may include the steps of: generating situational data through an intelligent framework (510); determining target data to be used as machine learning learning data among the situational data (520); collecting the determined target data to obtain raw data (530); processing the raw data into source data by setting the format or size suitable for machine learning and removing duplicates (540); generating labeled data by performing data labeling on the source data (550); and constructing a learning dataset including the source data and the labeled data (560).

[0061] In step 520, the determination of target data can be performed by automatically selecting a training data profile through the training data management unit based on the generated situation data when situation data is generated, and determining the target data among the situation data based on the items included in the selected training data profile.

[0062] In step 530, the acquisition of raw data, that is, the collection of target data, can be performed by the learning data management unit automatically collecting and storing the confirmed target data when the target data is confirmed.

[0063] Through the learning data management unit, a learning data profile is selected based on data related to the real-time environment or situation (in other words, high-level situational knowledge data), and the data to be collected is collected and stored based on the specifications of the profile, thereby enabling the acquisition of raw data necessary for learning, and this process can be performed automatically.

[0064] FIG. 6a is a table summarizing data generated through the performance of a basic mission of an unmanned aerial vehicle according to one disclosed embodiment, and FIG. 6b is a diagram showing a learning data profile exemplarily configured to reflect a threat situation of an unmanned aerial vehicle according to one disclosed embodiment.

[0065] The basic mission of an unmanned aerial vehicle (UAV) is to fly over a specific area using modules mounted on the UAV, and the modules of the UAV used in the experiment of the present disclosure are GPS, camera, dynamic pressure sensor, static pressure sensor, and geomagnetic sensor. The table in FIG. 6a summarizes the data generated through the performance of the basic mission.

[0066] During the operation of a mission, "GPS Time" is generated by default, representing the duration of the mission in GPS time. As situational information for the UAV, "ownship position" is generated, indicating the UAV's current location via latitude, longitude, altitude, and elevation above sea level; "ownship velocity" is generated, indicating the UAV's speed in NED format; "flight direction" is generated, indicating the flight direction; and "camera gimbal" is generated, indicating the UAV's camera angle. Additionally, as environmental information for the UAV, "Airdata" is generated, representing meteorological data such as temperature, pressure, and wind speed around the UAV; "designated position" is generated, indicating the UAV's destination via altitude, latitude, and longitude; "object recognized" is generated, indicating objects recognized by the UAV; and "current danger" is generated, indicating threat information threatening the UAV.

[0067] In the basic mission experiment, the frequency of threat occurrence by mission area was learned from the above data, and based on this, threats were prevented in advance in areas with a high probability of occurrence. Therefore, when an unmanned aerial vehicle is threatened, data collection regarding the corresponding threat type, time and location of the threat is required, so a learning data profile as shown in Fig. 6b was constructed.

[0068] FIG. 7a is a table summarizing data generated through the performance of an unmanned aerial vehicle tracking mission according to one disclosed embodiment, and FIG. 7b is a diagram showing an exemplary learning data profile configured by reflecting the movement path when an unmanned aerial vehicle tracks a target according to one disclosed embodiment.

[0069] Experiments related to UAV tracking and strike missions were designed to verify whether existing profiles can be reused when collecting new types of data, provided that relevant data has already been designated as collection targets in the existing profiles. To this end, tracking and strike missions were configured as extensions of the basic missions. The experiments were structured so that when collecting new types of data utilizing newly generated data from each mission, the profiles defined in the basic missions were referenced to define the new profiles and collect the data.

[0070] A tracking mission is a task to track a specific target when it is discovered during flight. Therefore, compared to the basic mission, additional data is generated, such as "tracking target," which corresponds to prior knowledge about the target the UAV is to track; "tracking target perceived," which occurs when the UAV recognizes the tracking target; and "target tracked," which indicates the target the UAV is currently tracking. Figure 7a summarizes the data generated through the execution of the tracking mission.

[0071] In the tracking mission experiment, when the unmanned aerial vehicle (UAV) tracks a target, a tracking path is collected based on the UAV's position, and this is used to learn a movement path according to the type of target, so that tracking can continue by predicting the movement path when the tracked target is temporarily lost. Therefore, since data collection such as the type of target, the UAV's position, and time is required when the UAV is tracking a target, a learning data profile as shown in Fig. 7b was constructed.

[0072] The collected data utilizes data already being collected in the profile defined in the existing Fig. 6b. Since the necessary data is already being collected in the existing profile, the profile was constructed by utilizing dependencies to reuse parts of the existing profile.

[0073] FIG. 8a is a table summarizing data generated through the performance of an unmanned aerial vehicle strike mission according to one disclosed embodiment, and FIG. 8b is a diagram showing an exemplary learning data profile configured by reflecting the strike results when an unmanned aerial vehicle strikes a target according to one disclosed embodiment.

[0074] A strike mission is a mission to strike a specific target when it is discovered during flight. Therefore, compared to a basic mission, data such as "attack target," which corresponds to prior knowledge about the target to be tracked by the unmanned aerial vehicle; "attack target perceived," which occurs when the target is recognized; "target distance," which indicates the distance to the target; and "attack result," which indicates whether the strike was successful, are generated. Figure 8a summarizes the data generated through the execution of a tracking mission.

[0075] In the strike mission experiment, data on the success of a strike was collected based on the distance to the target, the type of target, the location of the drone, and weather conditions when the drone strikes a target, and this was used to predict the success rate of the strike to increase the hit rate. Therefore, since it is required to collect data such as the location of the drone, the type of target, the distance to the target, and weather data when the drone strikes, a training data profile as shown in Fig. 8b was constructed.

[0076] The collected data utilizes data already collected in the profile defined in the existing Fig. 6b. Accordingly, the profile was constructed by partially reusing the existing profile through dependencies.

[0077] FIG. 9 is a table showing the number of knowledge generated and the number of knowledge collected per mission according to a disclosed embodiment.

[0078] Each mission was performed for 30 minutes, and the number of knowledge generated and the number of knowledge collected per mission are as shown in Fig. 9. Through the experiment of the present disclosure, it can be confirmed that information specified in the learning data profile is automatically collected during mission execution, and it can be confirmed that the specified information is collected normally even when the existing profile is reused and applied to a new mission.

[0079] As described above, the method for constructing an artificial intelligence learning dataset in an intelligent system according to the disclosed embodiment collects learning data by utilizing a learning data profile and a learning data management unit when collecting data in a new environment. More specifically, by automatically selecting a learning data profile by reflecting real-time conditions through the learning data management unit and automatically collecting desired data based thereon, it is possible to collect desired data without developing a separate data collector. Furthermore, by automatically collecting related data by referencing other profiles through the Dependencies item of the learning data profile, it has the advantage of being able to reuse existing data collection methods.

[0080] Meanwhile, embodiments of the present invention can be implemented as computer-readable code on a computer-readable recording medium. A computer-readable recording medium includes all types of recording devices in which data that can be read by a computer system is stored.

[0081] Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage devices, and also include implementation in the form of carrier waves (e.g., transmission over the Internet). Furthermore, computer-readable recording media may be distributed across networked computer systems, allowing computer-readable code to be stored and executed in a distributed manner. Additionally, functional programs, codes, and code segments for implementing the present invention can be easily inferred by programmers in the art to which the present invention belongs.

[0082] Since various variations may be made with the configurations and methods described and illustrated in this specification without departing from the scope of the invention, all matters included in the above detailed description or illustrated in the accompanying drawings are exemplary and are not intended to limit the invention. Accordingly, the scope of the invention is not limited by the exemplary embodiments described above, but should be determined only by the following claims and their equivalents. Explanation of the symbols

[0083] 410: Processor 420: Data learning unit 430: Training Data Acquisition Unit 440: Model Training Unit 450: Memory 460: Machine learning model 470: Communications Department

Claims

Claim 1 A method for constructing an artificial intelligence learning dataset in an intelligent system comprises: a step of generating situational data including time information, location information, speed information, flight direction information, environmental information, and recognition object information of an unmanned aerial vehicle through an intelligent framework; a step of determining target data to be used as training data for machine learning from the situational data; a step of obtaining raw data by collecting the determined target data; a step of processing the raw data into source data by setting the format or size suitable for machine learning and removing duplicates; and a step of generating labeled data by performing data labeling on the source data. A method for constructing an artificial intelligence learning dataset comprising: a step of constructing a learning dataset including the source data and the labeled data; wherein the confirmed target data is data confirmed from the situation data based on the knowledge (Predicates) items to be included in the data to be included in the selected learning data profile, and the raw data is acquired by automatically collecting and storing the confirmed target data through the learning data management unit when the target data is confirmed and a trigger condition set in the learning data profile is satisfied. Claim 2 A method for constructing an artificial intelligence learning dataset according to claim 1, characterized in that the learning data profile is a specification describing the target data in relation to collection. Claim 3 A method for constructing an artificial intelligence learning dataset according to claim 1, characterized in that the contents of the items included in the learning data profile include a profile ID, a task for data collection, another profile to be referenced, knowledge to be included in the data to be collected, a data storage format, a time of data collection, and a data storage location. Claim 4 A method for constructing an artificial intelligence learning dataset according to paragraph 3, wherein the learning data management unit automatically collects the confirmed target data based on the contents of the items included in the learning data profile, and also collects data associated with the profile described in the items of the other referenced profile. Claim 5 A method for constructing an artificial intelligence learning dataset according to claim 3, wherein the item regarding the data collection time of the learning data profile includes a first type for collecting data at the time of data generation and a second type for collecting data at predetermined intervals. Claim 6 A method for constructing an artificial intelligence learning dataset according to claim 5, wherein the learning data management unit automatically collects and stores the confirmed target data based on the contents of the items included in the learning data profile, selects the learning data profile based on the task for collecting the data, and collects data associated with other profiles other than the selected learning data profile based on other referenced profiles, collects the target data when it is created in a public knowledge repository in the case of the first type based on the data collection time, and collects the data according to the cycle described in the item contents regarding the data collection time in the case of the second type, generates learning data through mapping between variables included in the knowledge item containing the data to be collected and variables included in the data storage format item using the data included in the collected knowledge based on the knowledge containing the data to be collected, and stores the generated learning data based on the data storage location. Claim 7 An artificial intelligence learning dataset construction device in an intelligent system comprises: a learning data management unit; a communication unit for transmitting and receiving signals; a memory; and a processor for functionally controlling the communication unit and the memory. The processor comprises: an operation of generating situational data including time information, location information, speed information, flight direction information, environment information, and recognition object information of an unmanned aerial vehicle through an intelligent framework; an operation of determining target data to be used as learning data for machine learning among the situational data; an operation of collecting the determined target data to obtain raw data; an operation of processing the raw data into source data by setting the format or size suitable for machine learning and removing duplicates; and an operation of generating labeled data by performing data labeling on the source data. An artificial intelligence learning dataset construction device comprising: an operation of constructing a learning dataset including the source data and the labeled data; wherein the confirmed target data is data confirmed among the situation data based on the knowledge (Predicates) items to be included in the data to be included in the selected learning data profile, and the raw data is acquired by automatically collecting and storing the confirmed target data through the learning data management unit when the target data is confirmed and the trigger condition set in the learning data profile is satisfied. Claim 8 An artificial intelligence learning dataset construction device according to claim 7, characterized in that the contents of the items included in the learning data profile include a profile ID, a task for data collection, another profile to be referenced, knowledge to be included in the data to be collected, a data storage format, a time of data collection, and a data storage location. Claim 9 An artificial intelligence learning dataset construction device according to claim 8, wherein the learning data management unit automatically collects the confirmed target data based on the contents of the items included in the learning data profile, and also collects data associated with the profile described in the items of the other referenced profile. Claim 10 A computer-readable recording medium that is readable by a digital processing device and has a program of instructions that can be executed by a digital processing device to provide for the construction of an artificial intelligence learning dataset in an intelligent system, wherein the program of instructions that can be executed by a digital processing device is tangibly implemented, and the program for executing the method of any one of claims 1 to 6 is recorded on the computer.

Citation Information

Patent Citations

  • Artificial device and method for controlling the same

    KR1020190106948A

  • Cognition system of object in ship's drawing using a deep-learning

    KR102274081B1

  • Generation and management of an artificial intelligence (AI) model documentation throughout its life cycle

    US11263188B2

  • Automated collection of machine learning training data

    US20190220698A1

  • Dynamically applying machine learning models from profiling received data

    US20210304057A1