Hybrid modeling for remaining useful life prediction in power module prognosis

The ever-growing use of power electronic modules, caused by the shift towards sustainable technologies, raises the issue of remaining useful life (RUL) estimation. Purely physics-based models fall short, as physical knowledge proved to be insufficient to fully and accurately describe the evolution of the module's lifetime. simulation-based physical models also suffer from high computational costs in the case of complex problems. Meanwhile, classical data-driven models require high volume and good quality of data. This is difficult to meet when the data generation process is slow or expensive, which is the case for problems similar to ours. In this paper, we present an efficient hybrid model for remaining useful life prediction of wire bonds in power electronic modules. It uses Paris' law as its physical basis, alongside data from previous experiments. We first apply an interpolation-based transformation to said data, allowing us to distinguish different phases in the module's lifetime. We then define a prediction scheme for each phase to forecast the module's lifetime. Physical knowledge is incorporated in an adaptive way, so as to use it only where it is valid and in accordance with the experimental data. This allows us to leverage the benefits of physical knowledge as well as data to form a comprehensive model that can be extrapolated for predictions on unseen test conditions.