부산대학교 도서관

Faults in electrical machines are mostly detected by analyzing direct measurements. However, an early detection can be difficult with this measurement evaluation. With the knowledge of the most important tolerances, that influence the output performance of the machine and the significant frequency orders, a method to classify the machine's quality at the End of Line (EoL) test bench and for condition monitoring is studied. With a transient drive simulation stochastically distributed tolerances are simulated, the associated frequency orders analyzed and compared to test bench measurement results. Machine Learning (ML) is applied on the simulation data and evaluated for an EoL test procedure as well as eccentricity classification where a large portion of the test cases is correctly classified. If the ML models are applied on the measurement data the results are classified slightly too high, but the gradation of the machines remains equal compared to the measurement analysis. If no fault cases are available for training, an autoencoder can be utilized to identify the fault cases. As a result, the autoencoder trained on the basis of the simulations can detect the fault cases well. If the autoencoder is applied on the measurement results, the gradation of the machines remains.