Investigation and modeling of reliability in power electronic devices under power cycling

Electronic power systems, crucial for various applications, operate in demanding environmental conditions and face different stressors, including high voltages, power and thermal cycling, and mechanical vibrations. These stressors can lead to component degradation, with consequent impacts on system performance, safety, and lifespan. Therefore, enhancing the reliability of electronic power systems is of paramount importance. Reliability assessment in the field of power electronics, especially during power and thermal cycling, necessitates a holistic approach that encompasses the analysis of failure mechanisms, the prediction of failure rates, and the development of strategies to improve system reliability. This complex task requires a profound understanding of the behavior of power semiconductor devices, thermal management techniques, and packaging technologies. The thesis primarily focuses on researching and developing reliability models related to the phenomenon of power cycling in semiconductor power devices. To achieve this goal, firstly a dedicated experimental set-up has been designed to conduct accelerated tests, enabling the initiation of the two primary failure mechanisms associated with this phenomenon: solder joint fatigue and wire bond degradation. The set-up implements two methodologies for power cycling tests: constant current and constant temperature cycling. Both methodologies are adopted in this work to calibrate analytical lifetime models, with the aim of determining their impact on the accuracy of lifetime estimation for power devices subjected to non-constant stress conditions. Furthermore, this work explores the implementation of deep learning techniques to establish a model capable of predicting the lifetime of power devices subject to various degradation mechanisms. Initial investigations involve the use of an Artificial Neural Network (ANN) to develop a non-linear static model. Experimental tests validate the accuracy and superiority of this model compared to traditional analytical approaches. In addition, a data-driven model has been developed using a bidirectional Long Short-Term Memory (bLSTM) network to predict the Remaining Useful Lifetime (RUL) of devices based on voltage degradation profiles. Attention has been focused on the impact of dataset partitioning on the model's performance, highlighting its potential for accurate predictions even with a limited amount of data.