Spatial-Temporal Customizable Topology Graph Networks Combined with LSTM for Power Device RUL Prediction

Abstract:
Reliably predicting the remaining useful life (RUL) of power semiconductor devices remains a major challenge. This study performed SiC MOSFET power module power cycling aging tests to simulate device degradation process. A graph convolutional neural network hybrid framework integrating long short-term memory (GCN-LSTM) with customizable graph topological data structures was developed for remaining useful life (RUL) prediction. on-state voltage (Vds), junction temperature swing (DeltaT), maximum junction temperature (Tj), and thermal resistance (Rth), are employed as key degradation precursor parameters, and five different graph topology time series are constructed. Experimental results demonstrate that the proposed model can provide reliable RUL predictions, with the capability to dynamically update predictions during aging process.