Abstract
The proliferation of Industrial Internet of Things (IIoT) devices has generated massive volumes of multivariate time series data, yet the effective utilization of this data remains hindered by the scarcity of high-quality annotations and the ubiquity of non-stationary environments. A critical challenge in this domain is sensor drift—a gradual, often imperceptible deviation in sensor readings caused by aging, thermal degradation, or fouling—which severely degrades the performance of data-driven monitoring systems trained on historical data. While Self-Supervised Learning (SSL) has emerged as a potent paradigm for learning transferable representations without labels, standard SSL methods often conflate semantic system states with nuisance drift factors, leading to non-robust embeddings. This paper proposes a novel framework, Drift-Resilient Contrastive Coding (DRCC), designed to disentangle genuine system dynamics from sensor artifacts. We introduce a specialized drift-injection augmentation strategy and a dual-stream architecture that explicitly isolates drift components using orthogonal regularization. By maximizing mutual information between temporally coherent views while minimizing the retention of drift-specific features, DRCC learns representations that are invariant to long-term sensor degradation. Extensive experiments on the Tennessee Eastman Process and real-world wind turbine datasets demonstrate that our approach outperforms state-of-the-art self-supervised baselines in downstream classification and anomaly detection tasks, particularly under conditions of severe simulated and natural sensor drift.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2025 Wei Zhang (Author)