Multi-Label Classification for AIS Data Anomaly Detection Using Wavelet Transform

Abstract

Thanks to the Automatic Identification System (AIS), ships and other maritime equipment are able to communicate with each other, for example, by sending information about their position. This solution allows for early collision detection when two or more ships are on a collision course. In the newer version of AIS, a satellite infrastructure is used to extend the communication range. Unfortunately, satellite AIS deals with so-called packet collision effect: since there is a problem with synchronizing AIS data coming from multiple terrestrial areas, a single satellite may receive several AIS messages at the same time and be unable to correctly process them, causing the data to get lost or garbled. In this article, a machine learning based framework for detecting the incorrect AIS data is presented. In this approach, after the first stage (clustering), a dedicated anomaly detection algorithm searches for damaged AIS messages and conducts multi-label classification (with Random Forest and wavelet transform) to decide which fields of such message requires further correction. The results of measuring the effectiveness of the proposed approach using real AIS data are presented.

Reconstructing Damaged Data in AIS and Other Telecommunications Systems: A Survey

“…We find the results presented in given references promising (the proposed algorithm managed to find 90%-95% messages and 58%-74% of their fields with 2 bits artificially damaged [Szarmach and Czarnowski 2023]). Therefore, we will continue our work on developing this framework in the future.…”

“…At first, we take a look at the inside of the aforementioned 1-element clusters [Szarmach and Czarnowski 2022]. Since the origin of such messages is unknown (as they are not related to any other messages to form a trajectory cluster, and moreover, we cannot be certain if its MMSI field is correct or not), we use the K Nearest Neighbour algorithm [Altman 1992] to decide which other cluster seems the most similar to that message.…”

“…The damage can be recognised by a sudden value change, thus, the differences calculated should be high. The pre-trained classifiers (such as Random Forest [Ho 1998] or XGboost [Chen and Guestrin 2016], one for each field) analyse the differences obtained and decide whether to consider them (and their corresponding AIS message field) anomalous or not -see this method's performance results in [Szarmach and Czarnowski 2022]. Then, we examine the proper, multi-element clusters.…”

“…For static fields, we use an Isolation Forest [Liu, Ting and Zhou 2008] to find anomalous values directly. We believe that this method deals with finding damage even when the control sum field fails to do so (such as GPS drift) -see [Szarmach and Czarnowski 2023] for more details and results.…”