Contrastive Representation Learning for Non-Stationary Time Series

Abstract

Industrial IoT sensor streams are inherently non-stationary due to machine wear, shifting load
conditions, and variations in environmental dynamics. Traditional models often fail to maintain
performance under such evolving patterns because they assume stable statistical characteristics over
time. This work presents a contrastive representation learning framework designed to generate robust
latent embeddings that remain informative across drift, transient shifts, and operational mode changes.
By pairing segments of sensor data based on temporal and contextual similarity, the model learns to
separate meaningful machine-state variations from noise while preserving continuity in normal
machine evolution. The resulting representations improve anomaly detection reliability, enable early
fault identification, and provide interpretable state trajectories for maintenance decision-making. The
approach offers a scalable and adaptable foundation for intelligent monitoring in complex industrial
environments.