Yield Prediction using Artificial Neural Networks

Abstract

This paper presents a yield prediction method for a hard disk drive manufacturing processusing artificial neural networks. The networks presented are multilayer perceptrons trained by thebackpropagation algorithm. Eight networks are used to predict the yield based on the concept ofMixture of Local Expert. Seven of them act as a local expert and predict yields of each work stationin the whole process. The eighth network serves as a gating network which combines and executesinformation from each local network and then makes a prediction. The experimental results suggestthat the networks built have a fair forecast capability. This paper also uses the Noise Injectionapproach in conjunction with the cross-validation stop training method to build a neural network. Theexperiment shows that such approaches help to improve yield prediction capability.