An unsupervised methodology for online drift detection in multivariate industrial datasets

Slight deviations in the evolution of measured parameters of industrial machinery or processes can signal performance degradations and upcoming failures. Therefore, the timely and accurate detection of these drifts is important, yet complicated by the fact that industrial datasets are often multivariate in nature, inherently dynamic and often noisy. In this paper, a robust drift detection approach is proposed that extends a semi-parametric log-likelihood detector with adaptive windowing, allowing to dynamically adapt to the newly incoming data over time. It is shown that the approach is more accurate and can strongly reduce the computation time when compared to non-adaptive approaches, while achieving a similar detection delay. When evaluated on an industrial data set, the methodology can compete with offline drift detection methods.