Fault detection in rotor system by discrete wavelet neural network algorithm

Abstract

This study identifies a method for detection of irregularities such as open cracks or grooves on a rotating stepped shaft with multiple discs, based on the wavelet transforms. Cracks are represented as reduction in diameter of shaft (groove) with small width. Single as well as multiple grooves are considered on stepped shaft at locations of stress concentration. Translational or rotational response curves/mode shapes are extracted from finite element analysis of rotors with and without grooves. Discrete and continuous 1D wavelet transforms are applied on resultant response curve or mode shapes. The results show that rotational response curves or mode shapes are more sensitive to shaft cracks and key contributors to identify the location of cracks than translation response curves or mode shapes. Discrete wavelet transforms are accurate enough to locate the groove of smaller size. Effectiveness of detection by wavelets transforms is analysed for single as well as multiple grooves with increase in groove depth. Increase in groove depth can be quantified by increase in wavelet coefficient, and it can be an indicator. White Gaussian noise with low signal-to-noise ratio is added to response curves and analysed for crack location identification. Intelligent techniques such as artificial neural networks are used to quantify the location and depth of crack. Discrete wavelet transforms coefficients are provided as input to the neural network. Feed forward artificial neural networks are trained with Levenberg–Marquardt back propagation algorithm. Trained networks are able to quantify the crack location and depth accurately.