Selection of secondary variables is an effective way to reduce redundant information and to improve efficiency in nonlinear soft sensor. A novel method based on kernel independent component analysis (KICA) and false nearest neighbors method (FNN) is proposed on selecting the most suitable secondary process variables. The first step is to convert the non-linear operating variables into the linear space with kernel method. One the basis, they are projected into the independent ones with KICA transformations. In order to compare the different impacts on the operating variables, each original variable is eliminated orderly from original datasets with FNN in KICA subspace. In this way, it is possible to trace the important cause for the prediction. The result shows its validity with the verification in hydrocyanic acid (HCN) process industry.