심층 트랜스포머 네트워크 기반의 인공지능 기술을 이용한 무기체계 실시간 이상 감지 기술
Real-time Anomaly Detection for Weapon Systems using Deep Transformer Networks based on AI
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study explores the applicability of AI(Artificial Intelligence) -based anomaly detection technologies in the testing and operation of weapon systems by implementing an anomaly detection mechanism for projectiles using AI technology. Traditional methods, such as LSTM(Long Short-Term Memory)-based autoencoders commonly used for anomaly detection, often suffer from overfitting to normal data and fail to adequately capture the complex relationships inherent in multivariate time-series data. To address these limitations, this study utilizes the TranAD(Transformer-based Anomaly Detection) model that is based on the transformer architecture widely used in the AI field. TranAD combines transformer structures with adversarial training to robustly and effectively learn generalized features from normal data, thereby accommodating both long- and short-term temporal patterns. This study utilizes radar observation data and predicts the trajectories of projectiles for training and experimentation. By analyzing information such as attitude, position, and velocity from the observation data, the model detects abnormal states in the projectile behavior. Additionally, the discrepancy between the predicted and the observed trajectories is incorporated as an auxiliary input variable to enhance the generalization of the model across various trajectories.
Abstract
This study explores the applicability of AI(Artificial Intelligence) -based anomaly detection technologies in the testing and operation of weapon systems by implementing an anomaly detection mechanism for projectiles using AI technology. Traditional methods, such as LSTM(Long Short-Term Memory)-based autoencoders commonly used for anomaly detection, often suffer from overfitting to normal data and fail to adequately capture the complex relationships inherent in multivariate time-series data. To address these limitations, this study utilizes the TranAD(Transformer-based Anomaly Detection) model that is based on the transformer architecture widely used in the AI field. TranAD combines transformer structures with adversarial training to robustly and effectively learn generalized features from normal data, thereby accommodating both long- and short-term temporal patterns. This study utilizes radar observation data and predicts the trajectories of projectiles for training and experimentation. By analyzing information such as attitude, position, and velocity from the observation data, the model detects abnormal states in the projectile behavior. Additionally, the discrepancy between the predicted and the observed trajectories is incorporated as an auxiliary input variable to enhance the generalization of the model across various trajectories.