DC serial arc fault recognition in aircraft using machine learning techniques

• The SVM – RFE technique is used in the context of DC serial arc fault recognition in aircraft applications. • A feature engineering and analysis method is presented. • Optimisation of the signals sampling frequency regarding recognition time vs recognition performance (robustness and reliability). • Recognition rates up to 98% are achieved with the smallest subset of relevant variables. Arc fault detection represents one of the major challenges for protection systems used in aeronautical industry due to the high demand in terms of reliability, robustness and detection time. Current aircraft has no system capable of recognize arc faults respecting those requirements. The problem becomes harder with the voltage levels increase expected in More Electric Aircraft (MEA) and for all electric or hybrid propulsion, pushing studies to consider more carefully the phenomenon of arc faults. This study proposes a machine learning approach to help detect DC serial arcs, which is a challenge in terms of recognition. The analysed database contains both current signal measurements of arc faults and nominal behaviours. A classifier is implemented based on the extraction of relevant features from the conventional current signals. The selection and design of the model is based on the Internal SVM – RFE technique. The obtained results demonstrate the best trade-off between all the performance requirements. The methodology is able to achieve a recognition rate of 98%.