Reproduce “Comparative analysis and evaluation of ageing forecasting methods for semiconductor devices in online health monitoring” on OpenResearch

About this paper

Adrian Villalobos, Iban Barrutia, Rafael Pena-Alzola, Tomislav Dragicevic, Jose I. Aizpurua

Researchers from Mondragon University and collaborating institutions conducted a comparative evaluation of seven methods for MOSFET aging forecasting, finding that Temporal Fusion Transformers (TFTs) provide superior long-term prediction accuracy by integrating future covariates and offer interpretable insights into degradation turning points, outperforming classical and neural network models in extended horizons.

Key insights

• Classical machine learning and statistical models like Ensemble Extreme Learning Machines (E-ELM) and ARIMA provided superior accuracy with lower computational cost for short-term aging forecasts.
• Temporal Fusion Transformers (TFTs) demonstrated significantly better performance for long-term aging forecasting, effectively capturing complex temporal dependencies and benefiting from the integration of expected future covariates.
• The interpretability of TFTs through their attention mechanism allowed for the identification of 'key aging turning points' in the R_DS_ON signal, correlating with accelerated degradation phases or emerging failure mechanisms.

Summary by alphaXiv.