Independent component analysis (ICA) is a new effective technique for separation of statistically independent sources. Generally, ICA requires that the number of sensors must be no less than the number of independent sources to ensure enough information for separation of all sources. In some practical applications, this requirement of ICA is not met and we are interested in separation of only one source. A new method called wavelet-ICA filter is proposed in this paper that attempts to extract the independent feature by only using one transducer. The method employs ICA to regularize the wavelet decompositions of a signal to find the independent feature. Morlet wavelet is employed in this application for its nonorthogonality. The analysis of the feasibility of the method is shown in this paper. The effectiveness of this method is demonstrated by applying it to both simulated signals and vibration signals collected from a gearbox for periodic impulse detection. PCA is also used for the same purpose as comparisons. The results show that the proposed method using ICA is much better than PCA for separation of an independent feature.
[1]
Erkki Oja,et al.
Independent component analysis: algorithms and applications
,
2000,
Neural Networks.
[2]
Guillaume Gelle,et al.
BLIND SOURCE SEPARATION: A TOOL FOR ROTATING MACHINE MONITORING BY VIBRATIONS ANALYSIS?
,
2001
.
[3]
M. Brenner.
Non-stationary dynamics data analysis with wavelet-SVD filtering
,
2003
.
[4]
Andrew Ball,et al.
A STUDY OF THE NOISE FROM DIESEL ENGINES USING THE INDEPENDENT COMPONENT ANALYSIS
,
2001
.
[5]
Aapo Hyvärinen,et al.
A Fast Fixed-Point Algorithm for Independent Component Analysis
,
1997,
Neural Computation.
[6]
G. Delaunay,et al.
BLIND SOURCES SEPARATION APPLIED TO ROTATING MACHINES MONITORING BY ACOUSTICAL AND VIBRATIONS ANALYSIS
,
2000
.