| Keywords [eng] |
Anomaly detection, maritime traffic patterns, Bi- direction LSTM, prediction interval, confidence interval, conformal prediction region, ellipsoidal prediction region, mutivessel data analysis |
| Abstract [eng] |
This research focuses on the comparative evaluation of anomaly detection techniques for maritime vessel traffic patterns, using multivessel data analysis. To establish a foundation for understanding vessel movement behavior and trajectory patterns, a Bidirectional Long ShortTerm Memory (BiLSTM) model was employed to predict changes in latitude (∆lat) and longitude (∆long). Four anomaly detection techniques Prediction Interval, Confidence Interval, Conformal Prediction Region and Ellipsoidal Prediction Region were evaluated and compared. Bringing novelty to the maritime domain, this study highlights the strengths and tradeoffs of these techniques in capturing anomalous patterns in vessel movement. Results showed that Conformal Prediction Region achieved the highest coverage probability (94.80%), reliably identifying normal vessel patterns and flagging deviations as anomalies. The Ellipsoidal Prediction Region, with a coverage probability of 86.46%, provided a complementary geometric evaluation aligned with spatial vessel dynamics, despite its higher computational demands. This work advances maritime traffic analysis by offering novel insights into effective anomaly detection strategies, with potential applications in enhancing maritime safety, security, and operational efficiency. |
