Research And Application In Particle Size Soft-Sensor Of Radial Basis Function Neural Network

The grinding-classification circuit is one of the most important parts of mineral process, and the overflow particle size distribution of the hydrocyclone is the important indicator of the grinding-classification circuit. The real-time detection of the overflow particle size distribution is very important to realize the optimizing control of the grinding circuit and to improve the grade of concentrated ore and metal recovery rate. But the existing particle size analyzer is so expensive that concentrating mill can not support. What's more, the maintenance of the analyzer is very complex.Soft-sensing technique is effective way to solve this problem. Compared with other methods, there are several advantages for soft-sensing technique. First, it is convenient to maintain equipment. Second, it is cheaper than online particle size analyzer. At last, testing speed is relatively fast, which can satisfy the requirement of real-time automatic control strategy. Therefore, this paper uses the soft-sensing technique to realize the real-time prediction of the overflow particle size distribution of hydrocyclone.This dissertation is mainly focusing on the modeling method based on RBF neural network. The detailed discussions have been expanded particularly theory of RBF neural network, selection of learning algorithm. Based on the analysis of many kinds of learning algorithm, this paper uses the nearest neighbor clustering algorithm to select the hidden layer units. There are several advantages for nearest neighbor clustering algorithm, such as small computation, a few of parameters needed to be set by experience and without predetermining number of hidden units. But there are also several disadvantages for this algorithm. For example, the algorithm is sensitive to outlier and noise spot, which is against the enhancement of generalization ability. To solve the problem, this paper proposes a kind of improved method. The number of samples is used to judge whether a cluster is formed by isolated point in this improved algorithm. Accroding to the above result, we can decide whether to delete some hidden nodes. What's more, the method of step changing is introduced in this paper, which solves the problem that the fixed step increases learning time. Based on the sunspot prediction problem, simulation experiment has done to test the effect of the improved algorithm, the results of which show that the improved algorithm improves the training speed and generalization ability.Based on the analysis of grinding-classification processes, the mathematical models and operating principle of hydrocyclone, we select some variables as auxiliary variable which the soft sensor model needs. And this paper establishs the soft-sensing model based on RBF neural network to predict overflow particle size distribution of hydrocyclone. Finally, simulation results indicate that the RBF neural network model can predict overflow particle size distribution of hydrocyclone very well.