Machine learning based fault diagnosis in permanent magnet synchronous reluctance motors

Permanent Magnet Synchronous Reluctance Motors have various application areas such as industrial machines, electric vehicles, fan and pump applications due to their high efficiency and torque density, low maintenance requirements. Faults that may occur in SMSRMs prevent the motor from operating properly and cause a decrease in efficiency. This situation means that the energy resources used in the system are more loaded and costs increase. Therefore, it is very important to detect the faults that may occur in the system in advance. In this thesis, real measurement results for different operation and fault types in an SMSRM are used from shared databases for classification with machine learning based artificial intelligence algorithms. For this purpose, a short circuit fault between two stator windings was created. Current and vibration data of the motor in normal operation and as a result of two different magnitude short circuit faults were used. These data were trained on machine learning algorithms ICA and XGBoost, deep learning algorithms LSTM, TDL and CNN models. In this thesis, ICA analysis was performed with the vibration and current data used and it was found that overlapping signals can be separated by this method. Accuracy and F scores were used to evaluate XGBoost, LSTM, TDL and CNN algorithms. In the classification of vibration data, the accuracy rate was 92.02% for the LSTM model, 92.31% for XGBoost, 98.54% for TDL and 99.34% for CNN. The average F score value was above 0.90 for each algorithm. In the analysis results using current data, the accuracy rates were 89.88% for LSTM and XGBoost models, 98.24% for TDL and 99.09% for CNN. The F score was above 0.90 for the models trained using current data.