Multi-feature Extraction of Metal Cracks using Based on Wavelet Neural network

Abstract

Abstract The traditional ultrasonic testing mostly uses a single feature, which induce the inaccuracy of the crack detection. This paper proposes a metal crack detection method based on multi-feature extraction. First, the characteristics of time-frequency information of the signal obtained by using wavelet packet transform is briefly introduced. Secondly, the designed wavelet neural network model consists of first layer with three neurons (wavelet packet transform of the ultrasonic echo signal, the peak and energy function), and the output layer with one neuron presenting recognition features to identify and analyze metal cracks. The experimental results show ultrasonic echo measurement with the improved method proposed model can effectively reduce the influence of white gaussian noise while significantly improving the recognition rate of metal cracks.