Data-Efficient Reliability Assessment Using Machine Learning : Application to Lifetime Estimation of Power Electronic Modules

The global shift towards sustainable energy and transportation relies heavily on the development of power electronic modules. Assessing the reliability of these components is thus a critical requirement to ensure the success and safety of this transition. This thesis addresses the challenge of predicting the lifetime of power electronic modules to enable effective predictive maintenance strategies, specifically focusing on bond wire degradation in IGBT modules. Accurate lifetime estimation is often hindered by two major research gaps: the domain shift between accelerated laboratory tests and realistic operating conditions and the inherent structural complexity of the power modules. To bridge these gaps, this work proposes a set of methodologies that incorporate physical knowledge with machine learning techniques through three key contributions:First, to alleviate the computational burden of finite element simulations, we develop high-fidelity surrogate models. By achieving a prediction speed-up on the order of 10⁶, these surrogates enable cycle-by-cycle damage estimation.Second, we introduce a probabilistic Physics-Informed Markov Chain framework for damage modeling. By transforming sparse, time-dependent degradation curves into a dense state space defined by mechanical descriptors, we create overlaps between tests conducted under different conditions. This allows the model to transfer knowledge from accelerated tests to previously unseen scenarios, enabling robust and uncertainty-aware predictions.Third, we address the challenge of data scarcity and the prohibitive duration of realistic experiments by developing BEDTime, a few-shot time series forecasting model designed for physical phenomena exhibiting temporal warping. BEDTime utilizes Dynamic Time Warping to align partially observed signals from slow realistic experiments with complete reference trajectories from accelerated tests. By leveraging shape similarities, the model extrapolates the trajectory of a long-term experiment using only a small initial segment of data. This drastically reduces the need for exhaustive run-to-failure tests, offering significant time gains.Ultimately, this thesis demonstrates the synergetic use of physical knowledge and machine learning to ensure physical consistency, computational efficiency, and high predictive power to model complex physical phenomena. The proposed methods contribute not only to the reliability of power electronics but also offer generalizable tools for predictive maintenance in other fields characterized by data scarcity and complex degradation dynamics.