Quality Prediction Model Based on Variable-Learning-Rate Neural Networks in Tobacco Redrying Process

As tobacco redrying process has characteristics of multi-interference, strong coupling, great hysteresis, nonlinear and uncertainty, it is very difficult to establish the physical model. This paper presented an innovative method with the variable-learning-rate-based back propagation neural network (BPNN) for establishing the quality prediction model of tobacco redrying process. First, characteristics of the process and correlation of the process variables are analyzed, and determine eight input parameters and two output quality indicators of model. Then, a quality prediction model of tobacco redrying process is established by using the BPNN structure. In the process of network training, BP algorithm is improved by using the method of variable learning rate, and satisfactory prediction results are obtained. Finally, in order to verify the effectiveness of this method, the improved BPNN model is applied for simulation experiment, and is compared with ordinary BPNN. The prediction results show that the improved model possesses strong self-learning function and higher prediction accuracy.