Scientific Machine Learning for Engine Health Management and Remaining Useful Life Prediction

Researchers present a multi-task machine learning framework for predicting turbine remaining useful life (RUL) and thermal indicators with quantified uncertainty. The system combines convolutional neural networks with bidirectional LSTMs to handle heterogeneous real-world fleet data and provides prediction intervals rather than point estimates, enabling risk-aware maintenance decisions.

This research addresses a critical challenge in industrial asset management: predicting when engines will fail before catastrophic breakdown occurs. Traditional forecasting methods struggle with real-world fleet data that are heterogeneous and non-stationary, making point predictions unreliable for high-stakes maintenance decisions. The proposed framework advances the state of prognostics by treating engine health management as a multi-task learning problem, simultaneously predicting turbine gas temperature metrics and remaining useful life while quantifying prediction uncertainty through intervals.

The technical innovation lies in the architecture's integration of convolutional feature extraction with residual bidirectional LSTM layers and attention mechanisms. Rather than forcing practitioners to adopt a one-size-fits-all model, the framework includes tunable parameters aligned with organizational policies and proprietary maintenance criteria. This flexibility is crucial for industrial adoption, where operational procedures vary significantly across organizations and regulatory environments.

For the aerospace and power generation sectors, this approach reduces maintenance costs by enabling condition-based rather than time-based scheduling while simultaneously lowering failure risk through uncertainty quantification. The stratified evaluation by flight phase and maintenance segment reveals that prediction quality varies by operational context—a finding that encourages more granular, context-aware deployment strategies.

The emphasis on prediction interval coverage probability (PICP) and coverage-width criterion (CWC) metrics represents a maturation in how the industry evaluates prognostic systems. Rather than optimizing accuracy alone, organizations can now trade off confidence levels against interval width, aligning model behavior with their risk tolerance. This probabilistic framework positions machine learning as genuinely practical for safety-critical systems.

• →Multi-task learning architecture jointly predicts turbine gas temperature and remaining useful life with quantified uncertainty intervals.
• →Framework designed with practitioner-facing tunable parameters to align deployment with organizational maintenance policies and proprietary criteria.
• →Prediction performance evaluated using both point metrics (MAE) and interval metrics (PICP, MPIW, CWC) with stratification by operational context.
• →Addresses real-world challenges of heterogeneous, non-stationary fleet data that traditional point-prediction methods cannot reliably handle.
• →Enables condition-based maintenance scheduling in aerospace and power generation, reducing costs while managing failure risk through probabilistic forecasting.