A Novel Deep Belief Network Based on Shallow Feature Regression

To address the problem that deep belief network (DBN) does not fully utilize the feature information of hidden layer in the application process, which leads to the degradation of model prediction accuracy. A novel DBN based on shallow feature regression (SDBN) is proposed in this paper. Firstly, the initial parameters of SDBN are obtained by unsupervised pre-training. All hidden layers are added to the fully connected layer for regression, and the final prediction output is obtained by linear weighted summation of the hidden layer outputs. Then, backpropagation algorithm (BP) is applied to the supervised fine-tuning process of SDBN. Finally, experimental testing in the debutanizer column validates the validity of the SDBN model. The experimental results show that the prediction accuracy of the SDBN model is reduced by 0.053and 0.071 compared with DBN and PLSR, respectively, under the premise of guaranteeing the model speed.