Study and application of a class of neural networks model with better generalization ability

This paper studies the factors that affect the generalization ability of a neural network model. Take an example of alumina concentration soft sensing in the process of aluminum electrolysis, some measures are presented to improve the model´s generalization ability. They include constructing neural networks with prior knowledge, ensuring the quantity and quality of samples through the special experiments and training neural networks both off-line and on-line. The practical application shows their effectiveness. The neural network model based on these design methods proved to be precise. It has better generalization ability and provides a reliable guarantee for advanced process control.