Deep Learning and Machine Life: A Critical Empirical Perspective — Introducing Phase-Shift Degradation Analysis in Prognostics

Abstract

Deep learning has become predominant for predicting degradation in industrial systems, widely valued for its modelling power and adaptability. However, its performance is often assessed on the basis of complete degradation cycles, under the assumption that all stages equally represent the system's health trajectory. In reality, this assumption fails to capture the actual behavior of many systems, where stabilization phases or temporary recoveries driven by control algorithms, feedback mechanisms, or inherent resilience interrupt the monotonic degradation trend. These phases do not necessarily indicate improved health but are often treated as such during training, leading to mislabeling, overfitting, and ultimately, reduced interpretability. In this work, we challenge the conventional end-to-end training paradigm by presenting three illustrative case studies involving simple degradation, nonideal deterioration, and performance plateau with partial recovery. These examples reveal how deep models, when trained indiscriminately on full-cycle data, can produce misleading predictions, particularly in regions where the system's behavior deviates from expected degradation. We demonstrate how commonly used performance metrics may fail to detect these issues and how models may be inadvertently guided toward false predictive confidence. Rather than discarding deep learning as ineffective in such cases, this work encourages a context-aware training strategy that accounts for the complexity of real-world degradation and addresses the interpretive gaps caused by overly generalized models. It also proposes opening a new topic, called “Phase-Shift Degradation Analysis,” to encourage further research in this field. Our goal is to initiate a broader conversation on responsible deployment of predictive models in critical monitoring applications.