Impact of different data processing methods on model diagnostic performance

Abstract

During fault diagnosis of rolling bearings of rotating machinery in nuclear power plant, there are different data processing methods for original vibration signals, and different data processing methods have different diagnostic accuracy under the same diagnosis model. [Purpose] When the performance of the model is limited, in order to improve the diagnostic accuracy of the model, it is necessary to study the data processing method. [Methods] Therefore, the original vibration signal is processed into five ways: time-domain diagram(TD), time grayscale diagram(TGS), fast Fourier transform frequency-domain diagram(FFT), short-time Fourier transform time-frequency diagram(STFT) and continuous wavelet transform time-frequency diagram(CWT) for the initial feature extraction of the original vibration signal, and the five methods are compared and analyzed in the original data, data with added noise and data with less sample size. [Result] The experimental results show that the FFT data processing methods has a more obvious test accuracy under the noisy data and small sample size, and its accuracy is 1.9% and 6.7%(data under 50 sample size) higher than the suboptimal methods in both cases. [Conclusions] Therefore, in the case of noisy data and small sample size, the diagnostic performance of the model can be further improved by adopting the FFT data processing.