Time-frequency analysis for non-stationary signal from mechanical measurement of bearing vibration

Abstract

The aeroengine has been an important power source in aircraft and usually its operating condition is under mal-condition, so keeping it operating in safety status is essential. To improve performance of fault diagnosis method applied in aeroengine condition monitoring, a novel approach combining wavelet transform and neural network is proposed. By locating the converging local modulus maxima of the wavelet transform at fine scales, the wavelet transform has a good effect on detecting signal singularity or sharp transients. When the signal changes abruptly, it results in the wavelet maxima. The effective feature vectors are acquired by discrete orthonormal wavelet transform based on multi-resolution analysis and inputted as variables of neural network. The wavelet neural network can be extended by adding appropriate units to the network which are trained to represent signal time-frequency characteristics. The proposed neural network has three advantages over a typical neural network: data driven learning, local interconnections and entropy based self-organisation. The synthesised method of recursive orthogonal least squares algorithm is used to calculate the network parameters. The simulation results and actual applications show that the method can effectively diagnose and analyse aeroengine vibration fault.