Modeling Difficult-to-Measure Process Parameters Based on Intrinsic Mode Functions Frequency Spectral Features of Mechanical Vibration and Acoustical Signals

Abstract

Heavy mechanical devices of complex industrial process produce soundly mechanical vibration and acoustical signals. Some difficult-to-measure key process parameters have direct relationship with these signals. A newly ensemble empirical mode decomposition (EEMD), Fast Fourier Transform (FFT), Mutual information (MI), and Kernel partial least squares (KPLS) based modeling approach is proposed to measure these process parameters. At first, different scale intrinsic mode functions (IMFs) of mechanical vibration and acoustical signals are obtained using EEMD technology. Then, FFT transforms these multi-scale IMFs into frequency domain, and MI based feature selection method selects interesting frequency spectral features. Finally, KPLS constructs the final soft sensor models using the selected features. Experimental results based on vibration and acoustical signals of ball mill demonstrate this approach is more effective than other exist multi-scale decomposition based methods.