Multi-Task Machine Learning Framework for Turbine Prognostics
A scientific machine learning framework jointly predicts turbine gas temperature (TGTU), Delta Turbine Gas Temperature (DTGT), and Remaining Useful Life (RUL) with quantified uncertainty. The model uses a shared sequence encoder with convolutional layers, residual bidirectional LSTM, and attention pooling, feeding task-specific heads for probabilistic regression and survival analysis. Designed for heterogeneous fleet data, it supports risk-aware maintenance decisions.
Key facts
- Framework jointly predicts TGTU, DTGT, and RUL
- Shared encoder uses convolutional front-end with residual bidirectional LSTM and attention pooling
- Task-specific heads include mean-variance estimation and survival head
- Designed for heterogeneous and non-stationary fleet data
- Provides prediction intervals with empirical coverage evaluation
- Supports risk-aware maintenance decisions
- Published on arXiv with ID 2605.30593
- Multi-task scientific machine learning approach
Entities
Institutions
- arXiv