Reservoir computing approaches to unsupervised concept drift detection in dynamical systems.

Area of Science:

• Complex Systems
• Time Series Analysis
• Machine Learning

Background:

• Dynamical systems often change over time (concept drift), challenging stationary modeling assumptions.
• Existing concept drift detection methods are primarily statistics-based, with fewer options for deterministic processes.
• Nonlinear time series analysis (NTSA) is gaining traction for deterministic systems but remains less explored for concept drift.

Purpose of the Study:

• To propose a novel unsupervised approach for concept drift detection in dynamical systems.
• To leverage the embedding capabilities of reservoir computing (RC) for characterizing system dynamics.
• To evaluate the proposed RC-based method against existing NTSA techniques.

Main Methods:

• Utilized reservoir computing (RC) models to generate embeddings of dynamical system data.
• Developed an unsupervised concept drift detection algorithm based on RC embeddings.
• Assessed the method on synthetic drifting data streams from various dynamical systems.
• Validated the approach on an experimental dataset of faulty ball bearing vibrations.

Main Results:

• The proposed RC-based method demonstrated superior performance in detecting concept drift compared to traditional NTSA methods.
• RC embeddings effectively captured the underlying dynamics of the systems, facilitating drift detection.
• The method showed robustness across different synthetic datasets and a real-world experimental case.

Conclusions:

• Reservoir computing offers a powerful framework for unsupervised concept drift detection in dynamical systems.
• The proposed RC-based approach provides a promising alternative to existing NTSA methods, especially for deterministic systems.
• Further research can explore real-time implementation and hyper-parameter optimization for practical applications.