Steady data judgment algorithm and neutral network robust training in boiler combustion optimization

Abstract

This paper proposes a steady data judgment algorithm, which includes three major parts, i.e. gross error detection, steady data judgment and steady data re-sampling. In order to minimise the impact caused by noise when training neutral network, the paper introduces the RANSAC algorithm and put forward a RANSAC-BP algorithm. In this algorithm a t-distribution based confidence interval is constructed to replace the distance threshold, which is selected by experience; the algorithm culls noisy data during neutral network training and then re-train the neutral network with noise free data. The algorithm has been validated by a simulation experiment. © 2012 Springer-Verlag GmbH.