Maritime Anomaly Detection
0 likes84 views
The document discusses the importance of maritime anomaly detection for safe navigation and preventing issues such as collisions, illegal fishing, and piracy. It highlights the challenges faced due to unreliable AIS data and the lack of annotated datasets for effective anomaly detection. Various machine learning approaches, including supervised, semi-supervised, and unsupervised methods, are explored for identifying anomalies in maritime activities.
Maritime Anomaly Detection
- 1. Maritime Anomaly Detection Zahra Sadeghi
- 2. Motivation and importance • Safe and secure maritime navigation by identifying suspicious activities • Marine transportation protection • preventing dangerous situations: • Collision • illegal fishing • Smuggling • Pollution • piracy AIS data from marine traffic around the ports of Seattle and Vancouver 2 Maritime Anomaly Detection - By: Zahra Sadeghi
- 3. Time series analysis of AIS data • AIS is an automated tracking and monitoring system used by marine vessels • It is facilitated by the continuous transmission of data with other nearby vessels, as well as AIS base stations and satellites • By analyzing the time series of space-time coordination, we can assess the situation and maintain a situation awareness • A time series is a collection of a sequential series of transmitted data points (AIS messages) measured at successive points over time. • The study of these data points, in order to extract meaningful feature, behavior, characteristics, and statistics, is called time series analysis. 3
- 4. Anomaly and outlier detection • Anomaly detection deals with identifying unlikely and rare events. • Finding observations that do not fit the typical/normal statistical distribution of a dataset. • Movement behavior deviation from other vessels of the same type 4 Maritime Anomaly Detection - By: Zahra Sadeghi
- 5. Challenges • AIS data is unreliable, noisy and inaccurate • AIS information not updated in a timely manner • long gaps between messages • AIS Transmitters and satellite receivers' noise • Vessel operators have to input codes to their AIS by hand (erroneous manual input) • AIS signals can be easily spoofed and manipulated by attackers or parties willing to obscure their locations • There is a lack of well-known anomalous AIS situations 5 Maritime Anomaly Detection - By: Zahra Sadeghi
- 6. Lack of annotated public dataset • labeling sequence data for the task of anomaly detection is an expensive manual task. • abundance of unknown and undefined anomalous events • lack of well-studied anomalous AIS situations to represent a reliable ground truth. • common ML approaches require training a model on an annotated dataset for learning the distinction between groups of normal and abnormal data 6 Maritime Anomaly Detection - By: Zahra Sadeghi
- 7. Anomalous AIS behavior Deviation from standard route Unexpected activity Unexpected port arrival Close approach Zone entry 7 Maritime Anomaly Detection - By: Zahra Sadeghi
- 8. ML techniques for anomaly detection • supervised anomaly detection – modeling both the normal and anomalous behaviour. • it requires labeled data. • includes classification-based methods. • semi-supervised anomaly detection – modeling just one type of behaviour • the model could be incrementally trained, as new instances appear. • Normal behavior learning • unsupervised anomaly detection – searching for anomalies with no previous knowledge of the data. • analogical to clustering, where similar instances are grouped into clusters, based on some similarity measure (distance, density, …) • the assumption is that anomalies are well separated from the rest of the data 8 Maritime Anomaly Detection - By: Zahra Sadeghi
- 9. Prediction-based methods • forecasting the next time step • Autoregression models (ARIMA/SARIMA) • Pros • No labeled data is required. • Cons • sensitive to parameter selection • Poor performance for long trajectories 9 Maritime Anomaly Detection - By: Zahra Sadeghi
- 10. Clustering-based methods • Data that doesn't fit well to clusters are considered as anomaly/outlier • K-means, spectral clustering, hierarchical clustering, ... • Pros: • Can be applied in unsupervised way • Cons: • Hyperparameter tuning • Is not optimized for finding anomalies 10 Maritime Anomaly Detection - By: Zahra Sadeghi
- 11. Network-based methods • traditional Machine Learning algorithms are not capable of finding efficient answers due to unpredictability and complexity of maritime navigation • Supervised and unsupervised approaches • Sequential models (RNN, LSTM, AE) • Pros • Learn high-level and complex features automatically • Automatic feature engineering and self-learning capabilities • High performance, efficiency and accuracy • Cons • Requires large amount of data • Computationally expensive 11 Maritime Anomaly Detection - By: Zahra Sadeghi
- 12. Q&A 12