Potentials of machine learning in electric drives production using the example of contacting processes and selective magnet assembly

Abstract

Machine learning (ML) is a key technology in data driven industries. In general, ML algorithms offer insight in complex processes by analyzing measured data without acquiring in-depth domain knowledge. In contrast to common physical simulations they do not require excessive computational time and are well suited for real time analysis. This study focuses on transferring the potential of ML to the production of electric drives. Three major issues are identified: the preprocessing of the data, dealing with small data sets and the selection of an appropriate machine learning algorithm. Depending on the specific application in production, different algorithms, for example support vector machines, neural networks, random forests or boosted algorithms come into consideration. The potential of ML in electric drives production is demonstrated using two concrete applications: In the case of contacting technologies, such as thermo and ultrasonic crimping, ML algorithms for predictive maintenance, quality management and process control are considered. The second use case covers the selective magnet assembly. Here, a ML-based concept is proposed that predicts the cogging torque by analyzing magnet properties, as well as process parameters.