A Study on Deep Learning Application of Vibration Data and Visualization of Defects for Predictive Maintenance of Gravity Acceleration Equipment

Hypergravity accelerators are a type of large machinery used for gravity training or medical research. A failure such equipment can be serious problem in terms safety costs. This paper proposes prediction model that proactively prevent failures may occur hypergravity accelerator. An experiment was conducted to evaluate the performance method proposed this paper. 4-channel accelerometer attached bearing housing, which is rotor, and time-amplitude data were obtained from measured values by sampling. The trained with transfer learning, deep learning replaced VGG19 Fully Connected Layer (FCL) Global Average Pooling (GAP) converting vibration signal into short-time Fourier transform (STFT) Mel-Frequency Cepstral Coefficients (MFCC) spectrogram input 2D image. As result, has seven times decreased trainable parameters VGG19, it possible quantify severity while looking at defect areas cannot seen 1D.