Soft Sensor Modeling Of Aluminum Strip Grain Size Based On Ant Colony Optimization Algorithm And RBF Neural Network

The aluminum electromagnetic casting control system is a nonlinear complex system relating to many fields. The core problem of production process steady-state optimization is the construction of production process model and the determination of process control system parameters. The grain size is an important index for measuring the aluminum strip's quality. Until now, there is no instrument can be used to measure the value of grain size online. In order to achieve optimal control of the production process, the soft-sensor measurement technique was introduced to achieve online measurement of grain size. In this paper, the soft sensor model of aluminum strip's grain size has been established based on RBF neural network, and model parameters has been optimized by ant colony optimization algorithm (ACO).RBF neural network has been successfully used in many fields because of its capability of simple structure, fast training speed and good generalization ability. Based on analysis of electromagnetic casting technology of aluminum, some easily measurable variables were chosed as the secondary variables associating with the grain size. After preprocessing and PCA of collected sample data, the variables were used as inputs of the soft senor model, then the soft sensor model of aluminum strip's grain size was established based on the RBF neural network.To solve the problems of unstable model structure and poor generalization ability caused by uncertainty of the center of hidden layer in the RBF neural network model, based on basic ant colony algorithm, ant colony clustering algorithm was used to optimize the center of hidden layer, to determine the center and the broadband of network.The result shows that, the soft sensor model based on RBF neural network has good approximation and fitting ability. After optimizing the center of the RBF neural network by using ant colony clustering algorithm, convergence speed, stability and generalization ability of the model have been significantly improved.