Classification of induction motor bearing faults using long-short term memory deep neural networks

Induction motors are widely preferred in the industry because they are reliable, economical and durable. The various faults often occur in the inner ring, ball and outer ring regions of the bearing components of induction motors. Therefore, it is very important to detect bearing faults at an early stage in order to increase the efficiency of operation of induction motors. In this thesis, using Case Western Reserve University (CWRU) bearing dataset and Mendeley bearing vibration dataset, bi-directional long-short-term memory type (Bi-LSTM) deep neural networks are proposed for automatic classification of faults in the inner race, outer race and ball regions of induction motor bearings on vibration data. In the study, the performance of the proposed Bi-LSTM network is evaluated as a result of feature extraction using instantaneous frequency and spectral entropy, by dividing the vibration data of normal bearing and faulty bearing into windows of different sizes such as 128, 256, 512 and 1024. In the study, it is achieved that the accuracy of the Bi-LSTM network on the test set with different window widths on the dataset prepared from normal and faulty bearing data is around 60% on average in the CWRU dataset, while it is around 75% on the Mendeley Bearing Vibration dataset. In the classification of normal and faulty bearing data, the average accuracy of the Bi-LSTM network after feature extraction with instantaneous frequency and spectral entropy is obtained to be above 95% in the CWRU dataset and over 99% in the Mendeley Bearing Vibration dataset. As a result, the proposed Bi-LSTM network is considered to be a powerful classifier for the separation of faulty and normal bearing vibration data in induction motors. In the latter step of the study, classification of the location of the faulty data and classification of the fault size experiments are carried out. In the experimental studies, it has been confirmed that high performance is obtained in the classification of induction motor bearing errors on two different datasets with Bi-LSTM, where feature extraction is applied.