Signal imaging with gramian angular field based on the current signal and envelopes for anomaly detection of rotating machine

Abstract

In the context of rotating machinery systems, it is inevitable to encounter variable load conditions. These conditions result in alterations to the rotation speed of the motor rotor, consequently inducing fluctuations in the magnetic flux within the motor. Such variations manifest in the motor current signal in a manner resembling a fault frequency akin to the supply frequency. Due to the minimal difference between the supply frequency and the fault frequency, extracting the latter proves challenging. Therefore, this study introduces a methodology for discerning fault characteristics by employing time series signal imaging. The diagnostic process for abnormalities comprises the following sequential steps: (1) transforming 1-dimensional time series data into 2-dimensional images, (2) training Convolutional Neural Networks (CNN) based on normal data utilizing self-labeling techniques, and (3) computing indicators of integrity. The study advocates the utilization of Gramian Angular Field (GAF) for imaging purposes. GAF images possess the attribute of encapsulating temporal information from the signal. The envelope of the current signal is derived through Hilbert transform. Both the original signal and the envelope signal are then translated into GAF images. The image derived from the original signal captures the characteristics of the supply frequency, whereas the image from the envelope signal delineates the features of the defect frequency. CNN, renowned for its efficacy in image recognition, serves as a feature extractor. Notably, a self-labeling technique is employed to construct a learning dataset exclusively from normal data. Through this approach, CNN is trained as a feature extractor for normal data. The Mahalanobis distance is subsequently adopted as an indicator of integrity.