Past literature shows that artificial neural networks (ANN) can be successfully applied for fault diagnosis of rotating machinery but the results reported in the past literature are not satisfying. It is discovered that the main reasons for this poor accuracy are, noise present in the signals and usage of feature set that do not describe the signals accurately. Data obtained in field contains good degree of noise and this noise hurts the performance of the network. Although neural network, as a function estimator removes noise from time series to a certain extent, denoising prior to the modeling can greatly improve its ability to capture valuable information. The features used to describe the vibration signals implicitly define a pattern language. If the language is not expressive enough, it would fail to capture the information that is necessary for classification and hence regardless of the learning algorithm used, the accuracy of the classification function learned would be limited by this lack of information. Signal de-noising and feature selection are therefore highly desirable to improve the classification performance of the network. In this paper de-noising based on wavelet transforms and feature selection process based on genetic algorithms is presented. The implicit assumption made in all the past literature is, multi defects do not occur. But in reality we can find lot of cases with multi defects. So, cases with multi defects are also considered in this paper. GA is also used to select optimum network parameters. A multi layer feed forward neural network (MLFNN) is trained with error back propagation (EBP) learning algorithm. To have a complete understanding of the concepts behind the classification accuracy first the study is started with ideal signals derived by synthesizing different sinusoids composed of possible frequencies and amplitudes of sinusoids normally observed in machine vibrations and faults such as unbalance, misalignment and defects in anti friction bearings. The test results show that the proposed method improves the performance of diagnosis to 99.2% even with 15% random noise in the input signals.

This content is only available via PDF.
You do not currently have access to this content.