Data-Driven and Model-Based Prognostics and Health Management for Embedded Systems

Electronics are fundamental to most engineering systems. Despite increasing demand driven by rapid digitalization, the Internet of Things, and autonomous vehicles, Prognostics and Health Management (PHM) for electronic systems, particularly those with safety-critical functions, remains limited. While interest in applying machine learning to PHM is growing, its adoption in electronic systems is still in its early stages. This thesis presents a data-driven real-time PHM framework for condition monitoring, fault diagnosis, and multi-step ahead forecasting, all integrated into an unified data pipeline and deployed directly on the edge device.
An optimal model is first developed by reducing system measurement noise through exponentially weighted moving average (EWMA) and exponentially weighted moving standard deviation (EWMS). This is followed by feature selection using SHAP (SHapley Additive Explanations), an additive feature attribution method. Finally, hyperparameter optimization is performed using a well-defined search space, and the model’s performance is validated through cross-validation. The optimized model, an adapted multivariate bidirectional LSTM (Bidi-LSTM), is then deployed on the target embedded system for real-time inference and multi-step ahead forecasting.
The proposed PHM methodology was evaluated in two real-world case studies: (1) an electronic control unit (ECU) used in industrial applications and (2) battery state-of-charge (SOC) estimation for lithium-ion batteries. The data-driven PHM framework outperformed state-of-the-art model-based methods, including Kalman filter-based estimators. It achieved an overall classification accuracy exceeding 99.98% in the ECU experiment and demonstrated superior performance in SOC estimation, with a lower mean absolute error (MAE) and improved forecasting accuracy. The optimal models were deployed on two hardware testbeds to evaluate execution time and resource consumption, yielding positive results for both case studies. The results confirm the generalization capability of the proposed PHM framework, demonstrating its adaptability to different embedded systems with minimal adjustments. By integrating real-time inference and multi-step-ahead forecasting, this approach provides actionable insights for predictive maintenance, enhancing system reliability and operational efficiency.