Application of particle swarm optimization algorithm in soft sensor modeling

In order to improve the pressure sensor´s current stability and temperature drift performance, a soft sensor regression model was modeled based on least square support vector machine (LS-SVM). According to the difficulty in selecting penalty factor and kernel parameter which are called hyper-parameters in LS-SVM when modeling, particle swarm optimization (PSO) algorithm and ergodicity search algorithm (ESA) are proposed to optimize it. Then a soft sensor model would be reconstructed according to the optimal hyper-parameters. The experiment results show that the time of optimizing the hyper-parameters by PSO (about 1676 second) is decreased to that of ESA (about 211884 second), and mean squared error (MSE) of the prediction model with optimal parameters got by PSO (about 1.25times10^-6) is reduced to one sixth of that by ESA (about 8.35times10^-6). So PSO algorithm has more superior performance on global optimization and convergence speed. The optimal soft pressure sensor model got by PSO has more superior current stability and temperature drift performance, and the disadvantage influence of the change of working current and circumstance temperature on the sensor is decreased greatly. This method is feasible that combining PSO algorithm with LS-SVM in soft sensor modeling. It has definite development space and practical application value.