Prognostics and Health Monitoring: Case Study of a Light Rail Vehicle Power Converter Assembly

Electronic systems form the backbone of today's fast-moving and highly connected world. The reliability of electronics is consequently relevant, especially when electronics are used in safety-critical applications such as public transportation. Worn-out electronics can disrupt public transportation or even lead to catastrophic accidents. With the help of Prognostic and Health Monitoring (PHM), electronic failures can be anticipated in advance and predictive maintenance can be planned. However, the availability of sufficient data for health status prediction often is a challenge. In this article, we therefore present a cascade of machine learning (ML)-models for fatigue prediction based on insufficient data. As the use case, the application of a PHM approach to fleets of Light Rail Vehicles (LRVs) was chosen, and more specifically to solder joints in power converter assembly (PCA) used in these LRVs. Thermal loads are one of the major causes of solder fatigue. However, this load history is not directly available in most cases. Therefore, the first objective of the study is to predict thermal loads that the PCA experiences via virtual sensing techniques. Three prediction models a linear model, a multilayer perceptron (MLP), and a long shortterm memory (LSTM) network were developed to estimate the temperature based on the input current and ambient air temperature. Both the MLP and LSTM models demonstrated high prediction accuracy with coefficient of determination (R2) greater than 0.9, making them suitable for further use for fatigue estimation. The second objective of the study is to develop a solder fatigue estimation model. With the help of a virtual twin in the form of a physical model (finite element simulation), two sets of solder fatigue data were generated: one based on a synthetic temperature profile, and the other based on acquired mission profile loads from the LRV. An MLP was trained and successfully demonstrated to predict fatigue evolution in the solder joint under thermal loads with an averaged error less than 5%. Overall, this research provided two major PHM methodologies: virtual sensing for temperature prediction and a fatigue prediction model for solder joints in the PCA. The cascade of them enables a fatigue prediction for LRV based on insufficient data.