Empirical Research in Anomaly Detection for Rotating Machines Diagnosis using Deep Learning

This paper analyzes a deep learning model technique that diagnoses the normal/abnormal status of rotating machinery by transforming vibration data and extracting features. In the case of vibration data, it was collected from a rotating shaft alignment machine, and its normal/abnormal status was defined by the degree of axis alignment twist. Due to the characteristics of the vibration data, it is difficult to distinguish between two instances (normal and abnormal). It focuses on various data transformation techniques and feature extraction methods that explain the distinction between instances.In the phase of data preprocessing, the data is transformed to the frequency domain using Fast Fourier Transform and Wavelet Transform. The data then utilized Auto-encoding for feature extraction and MLP and CNN as models for predicting whether a machine is normal or abnormal. Consequently, this paper compares and evaluates the results of each process utilizing two data transformation techniques, one feature extraction method, and two models. Finally, it evaluates the performance of the model by using accuracy for the classification of rotational machine failure.