Abstract: A novel multi-layer neural networks is proposed, which is based on independent component analysis (ICA) for feature extraction of different modes (for example normal and bearing fault etc.), followed by a multi-layer perceptron (MLP) which implements the final classification. By the use of ICA, invariable features embedded in multi-channel vibration measurements can be captured. Thus, stable MLP classifier is constructed. The successful results achieved by the selected experiments indicated great potential of the new method in health condition monitoring of rotating machines.

Key words: Multi-layer perceptron, Independent component analysis, Mutual information, Principal component analysis