The development of method for detection of demagnetised and stator interturn short circuit faults in permanent magnet synchronous motor

Permanent Magnet Synchronous Motors (PMSMs) are widely used in automotive, aerospace and transport industries due to their high efficiency, compact design and high power density. Early detection of potential failures in PMSMs used in critical drive systems is important for the reliability and sustainability of the system. In this thesis, methods are developed for early detection of demagnetised and stator winding short circuit faults in SMSMs. Fast Fouirer Transform (FFT) was applied to the torque signal obtained from the experimental study to obtain turn short circuit fault (SCF) indicators that can be used in turn SCF detection and SCF severity classification in PMSM. The torque signal was analysed in the frequency domain in terms of harmonics for healthy, 2% SCF, 12.5% SCF and 25% SCF conditions. The amplitudes of the 2nd and 4th harmonic components of the torque in turn SCF conditions increased compared to the healthy condition. The amplitudes of the 2nd and 4th harmonic components of the torque increased with increasing turn SCF severity at all different loading and operating speeds of the motor. Therefore, the 2nd and 4th harmonic components of the torque are proposed as new turn SCF indicators that can be used for winding SCF detection in PMSM. The proposed moment-based turn SCF indicators are used in machine learning methods to detect turn SCF and classify the severity of SCF. Multilayer artificial neural networks (MLP), support vector machines (SVM), k-nearest neighbour method (KNN) and decision tree (DT) were used in machine learning methods. MLP, SVM, KNN and DT algorithms predicted the turn SCF in PMSM with 100%, 99.30%, 97.91% and 95.48% accuracy, respectively. The proposed torque-based turn SCF diagnosis method is compared with current and voltage-based SCF diagnosis methods in the literature. The proposed torque-based turn SCF diagnosis method has achieved more successful results than current and voltage-based SCF diagnosis methods. In order to detect the demagnetised fault (DMF) in the PMSM, FFT was applied to the current and moment signals obtained from the experimental study in healthy, 5% DMF and 10% DMF conditions. Motor signals were recorded at different loading and different operating speeds. The 11th harmonic amplitude of the current and the 6th and 12th harmonic component amplitudes of the torque in the demagnetised fault conditions increased compared to the healthy condition. This increase also occurred at different loading and operating speeds of the motor. Therefore, the 11th harmonic component of the current and the 6th and 12th harmonic components of the torque are proposed as new indicators for demagnetised fault detection in PMSM. The one-dimensional convolutional neural network (1D-CNN) developed from deep learning methods is used for the detection of demagnetised and winding short-circuit faults in PMSM and classification of fault type and severity. Stator raw three phase current signals are used in the proposed 1D-CNN deep network model. With the proposed 1D-CNN method, detection of DMF and winding SCF in PMSM, differentiation of faults and classification of fault severity are predicted with an accuracy of approximately 98%. Keywords: Permanent magnet synchronous motor, inter-turn short circuit fault, demagnetisation fault, convolutional neural network (CNN), multilayer artificial neural networks (MLP), support vector machines (SVM)