Modelling and Detecting Faults of Permanent Magnet Synchronous Motors in Dynamic Operations

Abstract

Permanent magnet synchronous motors (PMSMs) have played a key role in commercial and industrial applications, i.e. electric vehicles and wind turbines. They are popular due to their high efficiency, control simplification and large torque-to-size ratio although they are expensive. A fault will eventually occur in an operating PMSM, either by improper maintenance or wear from thermal and mechanical stresses. The most frequent PMSM faults are bearing faults, short-circuit and eccentricity. PMSM may also suffer from demagnetisation, which is unique in permanent magnet machines. Condition monitoring or fault diagnosis schemes are necessary for detecting and identifying these faults early in their incipient state, e.g. partial demagnetisation and inter-turn short circuit. Successful fault classification will ensure safe operations, speed up the maintenance process and decrease unexpected downtime and cost. The research in recent years is drawn towards fault analysis under dynamic operating conditions, i.e. variable load and speed. Most of these techniques have focused on the use of voltage, current and torque, while magnetic flux density in the air-gap or the proximity of the motor has not yet been fully capitalised. This dissertation focuses on two main research topics in modelling and diagnosis of faulty PMSM in dynamic operations. The first problem is to decrease the computational burden of modelling and analysis techniques. The first contributions are new and faster methods for computing the permeance network model and quadratic time-frequency distributions. Reducing their computational burden makes them more attractive in analysis or fault diagnosis. The second contribution is to expand the model description of a simpler model. This can be achieved through a field reconstruction model with a magnet library and a description of both magnet defects and inter-turn short circuits. The second research topic is to simplify the installation and complexity of fault diagnosis schemes in PMSM. The aim is to reduce required sensors of fault diagnosis schemes, regardless of operation profiles. Conventional methods often rely on either steady-state or predefined operation profiles, e.g. start-up. A fault diagnosis scheme robust to any speed changes is desirable since a fault can be detected regardless of operations. The final contribution is the implementation of reinforcement learning in an active learning scheme to address the imbalance dataset problem. Samples from a faulty PMSM are often initially unavailable and expensive to acquire. Reinforcement learning with a weighted reward function might balance the dataset to enhance the trained fault classifier’s performance.