Identification of loose particles using deep learning convolutional network

Abstract

Loose particles inside components can be threats to their reliabilities. Automatic identification of loose particles' material has great significance for finding the source of them. Machine learning methods based on hand-craft features have been widely applied on this problem. As deep learning has made success on various domains, based on PIND (particle impact noise detection) test, a method using spectrograms and CNN (convolutional neural network) is proposed in this paper. First, signals of loose particles including different material are collected by experiments. Then signals are converted to spectrograms. Finally, spectrograms are input to CNN for training and classification. Experiments show that the method can achieve 96 % accuracy on identifying five types of loose particles and has values of practically use.