Self-supervised learning can estimate engine health from operational sensor data without labeled health states, but traditional Bayesian filters remain competitive—revealing that this inverse problem is fundamentally harder than direct prediction.
This paper tackles turbofan engine health estimation as an inverse problem using machine learning. The authors create a realistic dataset with maintenance events and usage changes, then benchmark steady-state models, Bayesian filters, and self-supervised learning approaches to recover component health from sensor data without ground truth labels.