Caustic Ratio Of Alumina Production Soft Measuring Neural Network Model Correction Method

In the process of High-Pressure Digestion (HPD) of alumina, the ratio of soda to aluminate (RSA) is a very important production index, and its magnitude will directly reflect the quality of alumina. However, ANN is based on the historical data, even if it is precise initially, it can't adapt to the working conditions' alteration actually, and also, the prediction data of the RSA will deviate largely. Therefore, the new runtime parameter must be adopted. When we want to solve this problem we often use artificialities to collect the specimens and make the off-line training, and then interrupt system and replace primary ANN. But there is a great deal of workloads and time-consuming work in artificial selecting specimens, the training efficiency is influenced greatly. In this paper, through analyzing ANN model of RSA of alumina soft-sensing, the author refers to the more proper method of on-line rectification.This article summarizes three primary questions about the on-line rectification method of ANN of RSA: (1) when ANN needs to be re-training; (2) how to setup new on-line training patterns; (3) how to carry out this training. By analyzing the characteristics of the object which needs rectify, this paper lays on-line rectification model pertinently. The author, with a view to offering a perspective on the study of new patterns with firm verify and data amending, successfully resolve the above-mentioned questions.The practical application proves this method of on-line rectification not only improves predicted level of soft-sensing model, but also resolves the problems that ANN hardly adapts to the new working conditions as time pass by. Moreover, as the execution proceedings are all included the computer progressing, and work efficiency and credibility can be improved greatly.