Discovering New Prognostic Features for the Harmonic Reducer in Remaining Useful Life Prediction

Table 1

Summary of feature extraction.

Extract signal methods	Feature parameters	Definition [25]	Use situation	Advantages	Limitations
Signal analysis	Time-domain method	Calculates statistics from the signal in the time domain, which can reflect the degradation tendency of machinery	RMS [26–28]	Small computational burden; simple structure	Little fault information
			Kurtosis, mean, skewness, RMS, etc. [29–31]	Reflect the degradation evolution of the run-to-failure machinery [32]	Limited fault information
	Frequency-domain method	It is based upon the transformed signal over the frequency domain, which can distinguish and detach specific frequency components of interest	Cepstral, envelope analysis, Fast fourier Transform (FFT), higher-order spectra [30, 33, 34]	Small computational burden; simple structure	Limited fault information
	Time-frequency domain method	It can identify time-dependent variations of frequency components in the signal	Short-time fourier transform (STFT), wavelet transform (WT), and Hilbert–Huang transform (HHT) [30, 31, 35, 36]	Fully mine the data information; higher accuracy of RUL prediction; analyze nonstationary signals	Data redundancy; large computational burden
Deep learning methods	—	It can learn high-level abstract representations from the signal automatically and accurately	SparseAutoencoder, deep belief network, RNN, long short-term memory [37–40]	Extract feature automatically and accurately	Large demand for training data; rely on the temporal feature in signals [25]; limited accuracy [41]

Illustrative example 1: Atamuradov et al. [29] proposed a new health indicator construction method for point machine prognostics composed of a hybrid feature selection, which extracted 8 time-domain features from the point machine sliding-chair CM data. Illustrative example 2: in study [36], the FFT method is applied to extract 12 time-frequency domain feature data, including root mean square, peak, and mean, from the vibration signal to fully mine the data information. Then, the PCA is used to calculate the eigenvalues of the multidimensional data index covariance matrix and feature vector as the new comprehensive index data, which be used as input data in the prediction model. Illustrative example 3: Deutsch and He achieved accurate RUL prediction based on deep belief network (DBN), which made use of the self-taught feature learning ability of the DBN to extract deep representation features [38].