Data-Efficient Reliability Assessment Using Machine Learning: Application to Lifetime Estimation of Power Electronic Modules
Date:
This presentation represents my official PhD defense, addressing the critical challenge of predicting the lifetime of power electronic modules to enable effective predictive maintenance.
To bridge the gap between accelerated laboratory tests and realistic operating conditions, this work proposes integrating physical knowledge with machine learning across three key contributions:
- High-fidelity surrogate models: Achieving a prediction speed-up on the order of 10⁶ to enable cycle-by-cycle damage estimation.
- Physics-Informed Markov Chains: A probabilistic framework that transforms sparse, time-dependent degradation curves into a dense state space for robust, uncertainty-aware predictions.
- BEDTime: A novel few-shot time series forecasting model utilizing Dynamic Time Warping to extrapolate long-term degradation from sparse data, drastically reducing the need for exhaustive run-to-failure tests.
The defense was presented before a distinguished, multidisciplinary committee bridging power electronics, applied mathematics, and industry:
- President & Examiner: Bruno Allard (Integrated Power Management)
- Reviewers: Huai Wang (Power Electronics Reliability, AI) & Emmanuelle Abisset-Chavanne (Data-driven Simulations, Machine Learning)
- Examiner: Zeina Al-Masry (Applied Statistics, Stochastic Modeling)
- Invited Members: Sylvie Le Hegarat-Mascle (Data fusion, AI) & Emmanuel Batista (Power Systems Simulation, Alstom)