Estimating the distribution of SOH and RUL of electric vertical takeoff and landing aircraft (eVTOL) batteries using Machine Learning algorithms

Abstract

Prognostics and Health Management (PHM) involves a thorough analysis of the health condition of a machine and its components. In the field of battery management and PHM, the State of Health (SOH) and the Remaining Useful Life (RUL) are key indicators used to assess the current condition of a battery and predict how much more service life it has. Many studies focus on point estimates for the RUL and SOH of batteries in various domains. However, for the reliability and predictive maintenance planning, it is of utmost importance to determine the uncertainty associated with the predictions of SOH and RUL so that the decision-makers can make accurate informed decisions.

In this thesis, a machine learning pipeline for estimating the distribution of SOH and RUL on 21 Electric Vertical Take-off and Landing vehicle (eVTOL) batteries that are cycled under various conditions is proposed. Utilising the segments of charge and discharge phases of these batteries, 30 features are generated and an automatic feature selection is performed using the Boruta-SHAP algorithm. The pipeline estimates the distributional mean (µ) and provides uncertainty (σ) for each capacity test cycle of eVTOL batteries by using Random Forest Regression, Convolution Neural Network with Monte Carlo Dropout, and Mixture Density Network algorithms. The accuracy and sharpness of the obtained distributions are evaluated using the Continuous Ranked Probability Score (CRPS). The results show that the Random Forest Regression performs better than the other models and predicts the distribution of SOH and RUL with an average CRPS score of 0.96% and 40.49 missions/cycles, respectively.