Physics-informed Machine Learning for Prognostics and Health Management: Foundations, Advances and Prospects

Prognostics and Health Management (PHM) underpins condition-based predictive maintenance for safety-critical industrial assets. However, purely data-driven approaches face persistent challenges in generalization and interpretability, whose limitations are exacerbated due to domain shifts and label scarcity in real-world deployments. Physics-Informed Machine Learning (PIML) fills these gaps by embedding mechanistic priors into the learning pipeline, utilizing inductive biases to regularize ill-posed inference and enhance out-of-distribution robustness. This review systematizes the field's latest evolution—spanning component- to subsystem-level diagnostics and prognostics—through a framework defined by two orthogonal design axes. Along the physics-integration axis, we extend the generic input-architecture-loss classifications into a fine-grained taxonomy, exemplified by functional subcategories such as signal-conditioned inputs and constrained representation learning. Along the data-driven axis, we examine the data-driven backbones ranging from shallow to deep and discriminative to generative/representation learning frameworks; crucially, we consolidate cross-cutting learning strategies including transfer learning and uncertainty quantification, to bridge the gap between theoretical formulation and practical industrial viability. The review concludes with a strategic roadmap toward the next generation of robust, auditable, and decision-grade PIML4PHM systems.