Platform for Identification and Research of Anomalies based on AI - using Multi-Varied Methods – PIRANIA-MMV

This proposal follows the PIRANIA-MS project submitted in 2021 to ANR-ASTRID.
As a result of the committee’s evaluation report, comments were taken into account for this new submission.
As a consequence, our proposal has been reviewed by tightening our ambitions. We have worked closely together to limit the ambitions to achievable objectives in the field of expertise of TéSA and INP Toulouse laboratories.

As a result, PIRANIA-MMV proposes to adapt state-of-the-art anomaly detection methods to provide decision support mechanisms to operational people by determining the probability of anomalies in a tactical maritime scenario. The project aims at developing methods based on signal processing technics combined with Artificial Intelligence (Machine Learning, Neural Networks) and modeling for detecting association anomalies of various sensor measures and ship platform trajectory anomalies.
The scientific issue is thus linked to the theme of data processing and exploitation and the one related to the processing of massive data from heterogeneous sensors.


The strategy proposed in this project is to jointly use Radar and AIS time series to detect anomalies in vessel paths. This detection of anomalies can be done at the level of the associations of the Radar and AIS time series or at the level of parameters estimated from the ship’s trajectories. In both cases, we propose to develop new anomaly detection methods adapted to the joint processing of Radar and AIS data:

Association anomalies: We propose to modify existing anomaly detection methods based on the LoOP (Local Outlier Probabilities) method or the One-class SVM method to take heterogeneous Radar and AIS data into account. We also propose to analyse the potential of dictionary learning methods for detecting anomalies in Radar and AIS times series. These methods have been used successfully for the analysis of satellite telemetry.

Trajectory anomalies: The idea is to use an anomaly detection algorithm such as the One-class SVM algorithm to detect abnormal trajectories even if the boat has not completed its full journey. The innovative part will be here to determine the parameters adapted to the monitoring of ship trajectories.

For both kinds of anomalies, the use of active learning methods will be also considered to take advantage of a possible user feedback that would confirm the normal or abnormal character of some trajectories.