Modified Deep Belief Network Model and Its Application

Abstract

Aiming at the problem that the traditional deep belief network (DBN) model has long training time and it is difficult to find the global optimum because of the fixed step. A modified DBN model was proposed. In the DBN pre-training phase, the adaptive step (AS) was introduced into the Contrastive Divergence (CD) algorithm to solve the problem of finding the global optimum difficulty because of the fixed step, improving the accuracy and oscillation resistance. In the fine-tuning phase, the BFGS quasi-newton method (QNM) was used to speed up the convergence and reduced the training time. The modified DBN model was applied to the fuel consumption prediction of the descending section of the aircraft. The experimental outcomes prove that compared with the traditional DBN and its variant model, the modified DBN model improves the convergence speed and the prediction accuracy.