Wear Particles Classification Based On Wavelet Transform and Back-Propagation Neural Network

In the present paper, the application of wavelet transform to wear particles classification is studied. A wavelet method for wear particle classification is putted forwarded, according to the inherent properties of the wavelet being able to characterize the image information at each individual scale. First, the noise, distort and background of the original wear particle images are filtered, and the size of the wear particle images is processed to be the same. Then the two-dimensional discrete Daubechies wavelet is chosen to transform the wear particle images processed in three different scales. As a result, smoothed images and three detail images, in vertical, horizontal and diagonal directions, are obtained in three different scales. Four types of features, variance, mean, wavelet modulus extremum and energy of sub-images, are extracted from each sub-image as a wear feature vector to distinguish different wear particles. Second, based on back-propagation (BP) neural network, which possesses a powerful nonlinear mapping ability, a classifier is built. The wear particle feature vector is chosen as the input of the classifier, and six types of wear particles, rubbing particles, spherical particles, fatigue chunk particles, severe sliding particles, laminar particles and cutting particles, are chosen as the output of the classifier. Many different typical wear particles in six types as training samples are used to train the radial basis function neural network. After the goal of the neural network given is met, the training process is finished and the wear particle classifier based on radial basis function neural network is built. Finally, some practical examples show that the method of wear particles classification using wavelet transform is effective, and the comparison among the wavelet transform, the Fourier transform and the statistical functions shows that the method of the wear particle classification based on wavelet transform is more accurate.