Application of Spectral Kurtosis and Improved Extreme Learning Machine for Bearing Fault Classification

The condition monitoring of rotating machinery systems based on effective and intelligent fault diagnosis has been widely accepted. Traditional signal processing (SP) methods are less effective due to noises and interferences from different sources and incipient faults which remain active for a short time with a particular frequency. In recent times, SP techniques along with artificial intelligence methods are being used for fault classification. Various complex approaches in SP domain have used for feature extraction of the vibration data to design a feature set. A challenging task is to select dominant features from the available feature set for improving the accuracy of fault classification. Thus, motivated by spectral kurtosis (SK) and extreme learning machine (ELM), we propose a novel intelligent diagnosis method for fault classification of rotating machines. In this paper, SK is used as an input feature set to avoid the task of finding the dominant feature set. The extracted features are fed to ELM for fault identification. However, ELM performance primarily depends upon the hidden node parameters and the number of hidden nodes. The selection of optimum ELM parameters for good performance is an open issue. Therefore, modified bidirectional search with local search method is proposed to determine the optimum set of ELM parameters. The developed method is tested on two vibration data sets of rolling element bearings. We examined the significance of SK as a feature set and improved ELM in comparison with traditional methods. The experimental results demonstrate that the proposed method efficiently optimizes the ELM parameters to provide a compact ELM architecture and also enhances the fault classification accuracy.