Intelligent Traction Inverter in Next Generation Electric Vehicles: The Health Monitoring of Silicon-Carbide Power Modules
Carmelo Pino, Alessandro Sitta, Giulia Castagnolo, Angelo Messina, S. Coffa, Michele Calabretta, Fabio Scotti, Angelo Genovese, Vincenzo Piuri, Concetto Spampinato, Francesco Rundo
Abstract
In automotive and industrial domains, the “health monitoring” or “condition monitoring” of electronic devices is gradually playing a key role in manufacturing processes and innovation roadmaps. The concept of health monitoring is often related to the so-called “residual lifetime” of the monitored system. In this work, the authors have designed a deep learning system for the health monitoring of power devices in Silicon Carbide (SiC) technology used in the Traction Inverter Systems of the latest generation electric cars. A Temporal Fusion Transformer embedding such layers of Temporal Convolutional Network with a Multi-Head Attention block for the robust lifetime assessment of SiC power devices, is proposed. Specifically, the designed system predicts such future samples of the ON-state voltage between drain and source of the low-side part of the SiC power module <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$Vds_{LS}$</tex-math></inline-formula> , in half-bridge configuration. Extensive literature confirmed that the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$Vds_{LS}$</tex-math></inline-formula> signal can be efficiently used as a robust predictive device-degradation marker. Through the learning of the temporal feature relationships at different scales and the intelligent selection of relevant input features, the proposed solution will discard unnecessary input dynamics building a multi-step predictive model of the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$Vds_{LS}$</tex-math></inline-formula> signal, significantly more performing than the existing state-of-the-art architectures. The proposed deep pipeline has been tested on several ACEPACK <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TM</sup> DRIVE SiC power modules delivered by STMicroelectronics, with an average error of about <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$0.2\%$</tex-math></inline-formula> , confirming the effectiveness of the proposed system.