Data Preprocessing Method in Motor Fault Diagnosis Using Unsupervised Learning

Motor failure diagnosis using the existing deep learning is a lot of research using the CNN algorithm. Supervised learning is an algorithm in which a user specifies a correct answer of data and learns based on the correct answer. Therefore, CNN is very clear to classification of characteristics data. However, data that is not learned cannot be classified. The existing fault diagnosis using CNN failed to verify the data even if only the sensor measuring vibration in the same environment was changed. In this paper, a research was conducted using the K-means algorithm, which is unsupervised learning. In order to apply K-means to fault diagnosis, FFT data is divided by frequency. The center point is calculated by applying K-means algorithm for each frequency. The correct data set was determined based on the input data. Using the input data, the answer data set is calculated similarity and distance to the centroid of the cluster classified by each frequency, and the data with high probability are selected to classify the correct answer data set classified as K-means as normal or fault. By using the above method, even if the sensor is changed, it is possible to verify the data of similar trend. In order to verify the proposed algorithm, data were collected using different sensors and compared with the conventional CNN algorithm. The CNN failed to classify the data when the sensor was changed, but the proposed algorithm classified the data and showed high accuracy.