Predictive maintenance (PdM) plays a crucial role in modern industrial applications by using advanced algorithms and real-time data to predict and prevent potential equipment failures. Leveraging machine learning models like XGBoost, Random Forest, and LSTM, we have focused on predictive maintenance for electric vehicles (EVs), particularly targeting components such as batteries and fuel cells. Our approach emphasizes the estimation of key parameters like the State of Health (SoH), State of Charge (SoC), and effective range prediction. This involves detailed preprocessing, including cleaning datasets, imputing missing values, and encoding categorical features, ensuring the input data is accurate and ready for robust model training.
A significant achievement has been the integration of these models into a service container for real-time EV charging pattern prediction on a Coral dev board. By refining the preprocessing pipeline, we ensured the accuracy of predictions while also focusing on energy efficiency. Using the Perun library, we benchmarked energy consumption across various CPU cores, balancing computational performance with sustainability. Additionally, we have worked on datasets containing contextual variables like vehicle type, driving style, and trip distance, standardizing effective range metrics to optimize model outputs.
Moreover, our efforts extend to predictive maintenance for EV fuel cells, where Remaining Useful Life (RUL) prediction has been a key focus. By comparing traditional methods with lightweight CNN-based solutions, we evaluated real-time energy consumption and inference times using Carla-based simulations. This work provides a strong foundation for PdM applications, combining algorithmic optimization, hardware efficiency, and data-driven insights to improve reliability and reduce operational costs in EV systems.
