Machine learning for noisy multivariate time series classification: a comparison and practical evaluation

Abstract

Multivariate Time Series Classification (MTSC) is a complex problem that has seen great advances in recent years from the application of state-of-the-art machine learning techniques. However, there is still a need for a thorough evaluation of the effect of signal noise in the classification performance of MTSC techniques. To this end, in this paper, we evaluate three current and effective MTSC classifiers – DDTW, ROCKET and InceptionTime – and propose their use in a real-world classification problem: the detection of mooring line failure in offshore platforms. We show that all of them feature state-of-the-art accuracy, with ROCKET presenting very good results, and InceptionTime being marginally more accurate and resilient to noise.