A Data Based Diagnostic Method for Current Sensor Fault in Permanent Magnet Synchronous Motors (PMSM)

Abstract

In highly automated electric vehicles, the reliability of the electrical powertrain system is very important. A fault of a current sensor should be detected in an early stage to avoid a critical failure, which would lead to breakdown of the vehicle. This paper focuses on the gain fault of a current sensor on a permanent magnet synchronous machine (PMSM) and proposes a data based diagnostic concept, which is able to detect the fault and its severity in short time. After analyzing simulation data in healthy and faulty conditions, several basic features of current signals in time domain are generated. Subsequently, the three most effective features for the fault detection are chosen with a proposed feature selection tool. Then, three different machine learning algorithms (linear regression, decision tree and neural network) are used to train models with the selected features. The performance and characteristics of each model are compared. The neural network model has the lowest prediction error on the severity of fault. For standstill condition, another well performed diagnostic concept is developed with the same approach. The advantage of the data based approach is to reduce the effort on searching appropriate features by using machine learning algorithm. Therefore, diagnostic concepts for new faults or machines can be quickly developed because the most work of a data based diagnostic concept can be done automatically.