A Comprehensive Study of Machine Learning Algorithms for GPU based Real-time Monitoring and Lifetime Prediction of IGBTs

In critical energy infrastructures, Insulated Gate Bipolar Transistors (IGBTs) serve as essential components but are prone to unexpected failures. Precise estimation of the Remaining Useful Lifetime (RUL) of IGBTs is imperative for implementing predictive maintenance and assuring system reliability. This paper presents an innovative GPU-based approach for real-time health monitoring and lifetime prediction of IGBTs. The study explores a range of machine learning algorithms to determine the most effective one for precise lifetime prediction. Contrary to prior studies that concentrated on singular sensor data to minimize complexity and resource expenditure, this research leverages the capabilities of modern, economical, and robust GPUs to facilitate a data-driven, multi-sensor monitoring framework. The application of this approach has the potential to substantially bolster the reliability of energy infrastructure, notably in hydrogen plant. The paper conducts an exhaustive analysis of both single-variable and multivariate machine learning models, including Random Forest (RF), Long Short-Term Memory (LSTM), and Deep Neural Networks (DNNs), operating in real-time on edge GPUs. It also assesses the performance of two distinct GPU architectures - the NVIDIA Jetson Nano and Jetson Orin - in executing these machine learning algorithms.