A Comprehensive Methodology for Intelligent On-board Condition Monitoring of Electric Drives

Abstract

Electric drives are often used to power electrified vehicles such as electric cars or pedelecs. As any other mechanical system, electric drives wear out with time, which, consequently, increases the possibility of sudden failures that might compromise the safety of people in the surroundings of the vehicle powered by the damaged drive. A method to effectively increase the reliability of electric drives is the on-board condition monitoring of the device. By applying condition monitoring, failures can be detected and classified in advance, which allows to take preventive measures before a failure occurs.

This thesis introduces a comprehensive methodology which comprises all relevant stages for the development of intelligent algorithms for condition monitoring: from the data acquisition and pre-processing to the systematic generation of accurate and compact machine learning models for timely preventive measures. The main objective of the introduced approach is to predict the remaining useful life of an electric drive. Furthermore, the presented approach also enables the on-board fault diagnosis if required. The proposed methodology is validated with a proprietary database consisting of data collected with electric drives for pedelecs. The findings of this case study show that the introduced methodology enables a reliable identification of the end of useful life of electric drives, which is required for data labeling. Moreover, the generated models are able to accurately predict the end of useful life of an electric drive with sufficient time in advance, and to correctly identify the damaged element as well. Finally, this study shows how the resulting models are suitable for an embedded implementation for on-board condition monitoring due to their reduced size and computational complexity. In a second case study, the failure prognosis algorithms are compared with other related techniques using a benchmark database. This study demonstrates the high efficiency of the proposed method to estimate the time to failure with very compact models, when compared to other related approaches.