Abstract
We present a framework for deriving anomaly detection algorithms on timeseries data when the time and expression of anomalous behaviour is unknown. The framework is suited for problems in which individual machine learning paradigms cannot be directly implemented: supervised learning is not applicable due to the lack of labelled data, unsupervised learning is not effective since the normal operations are insufficiently defined and take complex and diverse forms, and deep learning risks confusing problematic behaviours for expected ones due to the possible repetitiveness of similar anomalies. The proposed approach is comprised of two phases: unsupervised discovery of anomalies, and semi-supervised construction and tuning of the anomaly detection algorithm. By leveraging data exploration methods and expert knowledge, the resulting algorithms are interpretable and detect a wide range of anomalous behaviours. The approach is applied to the early detection of wear and tear of maritime propulsion and manoeuvring machinery. We show that the final algorithm is able to detect different types of anomalies, including an actual internal leakage in a thruster which is otherwise overlooked by the present rule-based alarm system.