Soft sensor modeling based on modified PSO

Standard particle swarm optimization (PSO) has the drawback of trapping into local minima easily when used for the optimization of high-dimension complex functions with a lot of local minima. In order to deal with the problem an improved PSO algorithm with crossover operator is developed. Better particles are selected in this algorithm, thus can avoid premature convergence to local optimum as well as accelerate the convergence speed. Four high-dimension complex benchmark functions are introduced to test this method. Simulation analysis shows that improved PSO algorithm has better capabilities in convergence accuracy and speed as well as its global search performance by comparison with normal PSO algorithms. Finally the improved PSO based neural network (NN) soft sensor model for ethylene yield is developed, results of the application in industrial process control show that this model has high prediction precision and good generalization ability, it can satisfy the need of spot measurement.