Physics-informed Markov chains for remaining useful life prediction of wire bonds in power electronic modules

This paper presents a new approach to estimate the remaining useful life of a power electronic module where failure is caused by degradation in the wire bonds. The novelty of this work is that estimation is given for each loading cycle as opposed to estimating only the number of cycles to failure. A direct consequence is that one can make predictions on variable loading profiles using the proposed method, whereas classical solutions assume periodic loading, which limits their applicability. Experimental data of failure tests are used alongside finite element simulation to mechanically describe the state of the power module at each cycle. Using these mechanical quantities, we iteratively infer how the degradation evolves using Markov chains until failure. A first machine learning algorithm is used to establish a relationship between the degradation and the health indicator, and a second algorithm is used as a surrogate model for finite element simulations to drastically reduce computational time. Results show high extrapolation and interpolation capabilities of the obtained model, meaning that precise predictions can be obtained from experimental data where loading conditions are significantly different from realistic conditions. • A remaining useful life estimation pipeline is proposed for real life applications. • The pipeline is tested for difficult tasks such as extrapolation and interpolation. • Machine learning is used to leverage knowledge from physics. • Previously ambiguous relationships are established using this coupling. • The pipeline is probabilistic accounting for variance and quantifying uncertainty.