Signal-Based Feature Extraction and SOM based Dimension Reduction in a Vibration Monitoring Microsystem

Summary

We describe different signal based feature extraction methods which we are using in a intelligent microsystem for machine vibration monitoring. Each of these methods enables a special view to the machine vibration signal. The statistical methods like the quadratic mean value and the percentiles are useful for a fast global view. Signal processing methods as envelope detection, fourier transforms, wavelet transforms and time frequency methods are special views to the spectral composition of the vibration signal. The dimension reduction is realized by two self-organizing maps, one for the statistical features and one for the time frequency distribution. With a “Fuzzy Neural-Gas” network, which is a integration of the “Neural-Gas” Network and the Fuzzy c-Means algorithm, we are classifying the output of the self-organizing maps. The described hierarchical neural network integrates the dimension reduction and the classification for the machine vibration monitoring.