The applications of semiconductor power electronic devices, including power and RF devices, in industry have stringent requirements on their reliability. Power devices are subject to various types of failure mechanisms under various stressors. Prognostics and health management (PHM) allows detecting signs of failures, providing warnings of failures in advance, and performing condition-based maintenance. There is a pressing need to develop a robust prognostic model to detect anomalous behavior and predict the lifetime of devices that can be applicable to different types of power transistors. In the present dissertation, a comprehensive prognostic model

for remaining useful life (RUL) prediction of semiconductor power electronic devices is developed. The model consists of an anomaly detection module and a RUL prediction module including a non-linear system process model describing the evolution of parametric degradation, and a measurement model. The anomaly detection module uses principal component analysis (PCA) for dimensionality reduction and feature extraction, as well as k-means clustering to establish baseline clusters in the feature space. The novel singular-value-weighted distance (SVWD) is developed as the distance measure in the feature space, based on which Fisher criterion (FC) is used for anomaly probability calculation. The system process model incorporates variables concerning loading conditions and physics-of-failure of devices, and uses particle filter (PF) approach for process model training and RUL prediction. For PF methodology, a novel resampling technique, called MHA-replacement resampling, is developed to solve the particle degeneracy in classic PF techniques and sample impoverishment in traditional resampling techniques. The developed prognostic model is first implemented on IGBT modules for validation. It was reported that the module package of power transistors was susceptible to various types of fatigue-related failure modes due to coefficient of thermal expansion (CTE) mismatches under temperature/power cycles introducing thermomechanical stresses. The physics-of-failure "driving variable" is derived from Paris equation. The model is validated on several time-series IGBT module degradation data under power cycles from literature sources, based on SIR particle filter for RUL prediction with good accuracy. Then the model is implemented on GaN HEMTs, a representative of wide-bandgap semiconductor power devices. GaN HEMTs are susceptible to degradation mechanisms such as ohmic contact inter-diffusion that leads to voiding in the field plate at high temperature

under RF accelerated life tests (ALTs). The time-series data of the physics-of-failure "driving variable" is obtained from diffusion computation in Mathematica with the temperature prole coming from COMSOL thermal simulation. The RUL prediction results based on SIR lter are also satisfactory for GaN HEMTs. For each type of device, the new resampling technique is validated through performance benchmarking against state-of-the-art resampling techniques. Another reliability threat for GaN HEMTs, especially in aerospace and nuclear applications, is the degradation due to radiation effect on the device performance. Gamma radiation has been found to lead to generation of defects in AlGaN/GaN layers, which form complexes acting as carrier traps, thus reducing carrier density and current. EPC GaN HEMTs are irradiated under a wide range of Gamma ray doses and critical DC characteristics are recorded before and after radiation to quantify their shifts during the irradiation. Future work needed to allow implementation of the developed prognostic model for

RUL estimation is proposed.