A sampling method on the basis of imitation orthogonalization is proposed to ensure the typicality and ergodicity of samples. Considering the characteristics of back titration (BT) during cobalt removal with arsenic salt, such as many influencing factors and strong coupling, kernel principal component analysis (KPCA) is applied at first. Through KPCA, the effective characteristics of data can be extracted to reduce the dimensions of variables and to eliminate the coupling between variables. Then the extracted characteristic components are utilized as the input of radial basis function (RBF) neural network. Owing to there are many parameters in RBF neural network, which means that it is difficult to obtain the global optimal parameters, rival penalized competitive learning (RPCL) algorithm is adopted first to determine the original values of hidden nodes. On this basis, the improved particle swarm optimization (IPSO) is employed to select the parameters of RBF neural network. It is proved by the simulation results of industrial data that the BT prediction model is effective. KeywordsSampling method; KPCA ; RBF; IPSO; BT
[1]
Wu Tie-bi.
Improved operational pattern-PID based control of ORP in cobalt removal process using arsenic salt
,
2014
.
[2]
Wen Yu,et al.
Neural networks for the optimization of crude oil blending
,
2005,
Int. J. Neural Syst..
[3]
Erkki Oja,et al.
Rival penalized competitive learning for clustering analysis, RBF net, and curve detection
,
1993,
IEEE Trans. Neural Networks.
[4]
Tiebin Wu,et al.
Particle Swarm Algorithm Based on Boundary Buffering-Natural Evolution and its Application in Constrained Optimization
,
2014
.
[5]
G. Houlachi,et al.
The Removal of Cobalt from Zinc Electrolyte by Cementation: A Critical Review
,
2000
.
[6]
Bernhard Schölkopf,et al.
Nonlinear Component Analysis as a Kernel Eigenvalue Problem
,
1998,
Neural Computation.