A Continual Learning Approach for Failure Prediction under Non-Stationary Conditions: Application to Condition Monitoring Data Streams

Accurate forecasting of Remaining Useful Life (RUL) is crucial for predictive maintenance (PdM), permitting prompt actions that decrease downtime and maintenance expenses. Yet, conventional RUL estimation techniques often struggle to adjust to changing operational conditions and data drift, restricting their use in dynamic industrial settings. This research presents a continual learning framework designed for these non-stationary conditions, where models progressively learn from new data while retaining previously gained insights. By tackling issues such as catastrophic forgetting, the proposed technique enhances robustness and adaptability, rendering it applicable to real-world scenarios. The framework’s efficacy is validated using realistic industrial scenario datasets, like AI4I and N-CMAPSS, which include varied scenarios such as manufacturing and aerospace operations. Experimental findings showcase significant enhancements in prediction accuracy and model resilience over traditional methods. Incorporating deep neural network architectures, such as Multi-Layer Perceptrons (MLP) and Convolutional Neural Networks (CNN), within the continual learning paradigm, optimizes performance for dynamic data shifts. This study advances RUL estimation methods by offering a scalable and practical solution for PdM in sectors where reliability and efficiency are paramount. Its capacity to adapt to evolving operational conditions and provide actionable insights bolsters improved maintenance strategies, curtailing unexpected disruptions and operational costs. The proposed framework is particularly beneficial for sectors like manufacturing, energy, and aerospace, where non-stationary data and evolving failure modes present significant challenges.