Abstract: The shortcomings of the existing feature extraction and machine fault detection methods are analyzed. Combining HHT-DDKICA with support vector data description (SVDD) method, a new fault monitoring algorithm for hoist machine is proposed. Vibration signals of hoist machine are filtered into multiple interesting frequency bands and intrinsic mode functions (IMFs) are obtained through empirical mode decomposition (EMD). Then HHT denoising method and a signal energy criterion are adopted to select effective IMFs. Since single IMF may consist of some nonlinear vibration sources, an alternative data dependent kernel independent component analysis (DDKICA) method is presented to separate source signals. The method can determine a proper kernel function according to training samples, and the optimal model parameter can be also achieved by solving a DDKICA model selection criterion in the empirical feature space. Considering distributions of features extracted by DDKICA, SVDD is adopted to extablish new statistics and confidence limits. Hoist machinery application shows the efficiency of the proposed method.

Key words: Data-dependent kernel independent component analysis, Fault monitoring, Feature extraction, Hilbert-Huang transform, Support vector data description