Files
Abstract
A highly accurate modeling of power electronic devices is of considerable significanceto the health management and maintenance of power converter equipment. Thereliability-based predictions that are based on theoretical approaches are inefficientdue to the unique unit-to-unit condition and the complexity of advanced power elec-tronics systems. Classical data-driven approaches have been proposed to model thereliability of power transistors in the design, fabrication, and maintenance process.Kalman Filter (KF) [1], Particle Filter (PF) [2], and Bayesian calibration [3] are threeexamples of classical time series modeling and prediction techniques; however, the ac-quired model is not valid or scalable for the same underlying technologies because ofthe different fabrication variants and loads.On the algorithm side, I present a scalable real-time solution based on a deeplearning algorithm, Temporal Convolutional Network (TCN), to enable high accuratereliability modeling of power converter devices. The proposed method constructs anaccumulated degradation knowledge of different power devices and predict the agingprocess of a completely new and unknown device. In order to facilitate real-timepredictions, I leverage a vertical co-design framework, DeepDive [4], for power effi-cient implementation of Deep Neural Networks (DNN) on edge FPGAs. We outfitDeepDive with additional components in order to support the unique characteristicsof the TCN architecture, such as dilated convolutions and zero padding. Our exper-iments on MOSFET power transistors dataset provided by NASA [2] confirm thatdeep learning methods can improve the overall accuracy of aging prediction by 34.2%and 6.6% for four transistors over classical approaches and deep learning approaches,respectively. The customized hardware solution provided by the augmented DeepDive framework reduces the latency by 8.4×, 5×, and 2.1×, over Jetson Nano highand low power modes and Ulta96 Processing Subsystem (PS), respectively. In termsof energy efficiency, my solution reduces energy consumption by 94.4%, 93%, and 32.6%, in comparison with the high and low power modes of the Jetson Nano, andthe Ultra96 processing system (PS), respectively.