Semi-Supervised AI Analytics for Prognostic Health Indicator Learning from Acoustic and Guided-Wave Sensor Streams
Main article
Abstract
Structural health monitoring of composite components used in aerospace, civil, and industrial machinery requires reliable prognostic health indicators (HIs) that track degradation continuously from initial healthy conditions to end-of-life. Constructing such HIs is exceptionally challenging because ground-truth damage labels are almost universally unavailable, damage mechanisms in heterogeneous materials are stochastic and multi-scale, and single-modality sensing inherently limits either temporal or spatial diagnostic resolution. This paper proposes a semi-supervised AI analytics framework that integrates passive acoustic emission (AE) sensor streams with active guided-wave (GW) interrogation data to derive high-quality fused HIs under severe label scarcity. The framework embeds three prognostic criteria—monotonicity (Mo), prognosability (Pr), and trendability (Tr)—directly into the learning process through physics-informed proxy targets and criteria-driven regularization, eliminating the need for manually annotated health trajectories. Dedicated intra-modality pipelines handle AE and GW separately before a late-fusion meta-learner combines unimodal HIs into a single prognostic index. Bayesian hyperparameter optimization and ensemble averaging stabilize the learned trajectories across random initializations. Strict leave-one-out cross-validation on a run-to-failure fatigue dataset of carbon-fiber-reinforced composite panels confirms that the proposed fused HI consistently achieves a cohort-level Fitness score above 2.70 (out of 3.0), representing improvements of 12–18 percentage points over single-modality baselines. Among the evaluated inter-modality fusion regressors, Gaussian process regression yields the highest test Fitness of 2.65 ± 0.10, demonstrating that complementary temporal (AE) and spatial (GW) information can be effectively unified within a semi-supervised prognostic learning paradigm.
