Reservoir computing approaches to unsupervised concept drift detection in dynamical systems

While the assumption that dynamical systems are stationary is common for modeling purposes, in reality, this is rarely the case. Rather, these systems can change over time, a phenomenon referred to as concept drift in the modeling community. While there exist numerous statistics-based methods for concept drift detection on stochastic processes, approaches leveraging nonlinear time series analysis (NTSA) are rarer but seeing increased focus in cases where the processes are deterministic. In this work, we propose a novel approach to unsupervised concept drift detection in dynamical systems utilizing the embedding offered by a reservoir computing (RC) model. This approach is inspired by the performance of RC on supervised classification tasks that indicates a strong ability to characterize dynamical systems. We assess this method on a number of synthetic drifting data streams from dynamical systems as well as an experimental case concerning faulty ball bearing. Our results suggest that the RC based methods are able to generally outperform the existing NTSA methods across the test cases. We conclude our work with some comments regarding real-time implementation and the impact of hyper-parameters on the proposed algorithm.