Anomaly Detection in Regular Passenger Vessel Trajectories Using LSTM-Based Trajectory Prediction and Adaptive Statistical Techniques
본 연구는 AIS 데이터를 활용하여 운항 진행률을 반영한 이상항적 탐지 방법을 제안한다. 정기여객선의 정상 운항 특성은 운항 단계에 따라 변하므로, 이를 고려한 동적 이상탐지가 필요하다. DBSCAN(Density-Based Spatial Clustering of Applications with Noise) 기반 공간분석과 LSTM(Long Short-Term Memory) 기반 항적 예측을 통해 실제 항적과의 위치·속력·침로 편차를 산출했으며, 진행률별 통 계적 임계값을 설정하여 이탈 정도를 평가했다. 또한 육지-해양 이진지도를 적용하여 물리적 안전성을 검증했다. 결과적으로 LSTM 예 측항적과 실제항적의 평균 이격거리는 21.68 m이며, 통상적인 운항에서 위치·속력·침로의 변동범위가 진행률에 따라 유의하게 다름을 확인했다. 검증 시뮬레이션에서는 운항 이탈 시 통계적 위험단계가 명확하게 증가했으며, 육지 영역 침입 시 좌초 위험을 성공적으로 탐지했다. 본 방법은 정상 운항 특성을 반영한 이상항적 탐지 모델로서 해상교통관제의 의사결정 지원에 기여할 수 있다.
This study proposes an anomalous trajectory detection method that accounts for operational progress rates by utilizing Automatic Identification System data from regular passenger ships. Since the normal operational characteristics of regular passenger vessels vary by voyage stage, a dynamic anomaly detection model reflecting this variation is essential. Spatial analysis employing density-based spatial clustering of applications with Noise and trajectory prediction based on long short-term memory neural networks were applied to quantify deviations in position, speed over ground, and course over ground between predicted and actual trajectories. Statistical thresholds were established for each operational progress rate to evaluate the degree of deviation. Additionally, a binary spatial map distinguishing land and maritime regions was applied to validate physical safety conditions. Results indicate that the mean positional deviation between LSTM-predicted and actual trajectories was 21.68 m, with normal variation ranges for position, speed, and course showing significant differences across operational progress rates. Verification simulations demonstrated that statistical risk levels increased proportionally with trajectory deviations while grounding risk was successfully identified when trajectories entered land areas. This dynamic anomaly detection model incorporating typical operational characteristics can effectively support decision-making in maritime traffic-management systems and enhance navigational safety.