Data-driven soft sensors have been extensively studied in the process industry to provide an accurate online estimation of quality-related variables with easy-to-measure variables. For chemical processes with massive process variables, the performance of soft sensor models could be significantly improved by variable selection because part of these measurements is redundant or independent of quality-related variables. Generally, the variable selection is achieved by ranking process variables in order of their importance to the quality-related variables by correlation analysis. However, considering that correlation analysis methods are relative measures of variable dependence, the determination of the final variable set is quite subjective because there are several user-defined parameters. To overcome this limitation, a conditional entropy-based feature selection method is proposed. Considering that information entropy measures the degree of system chaos, the proposed method is based on the idea that the quality-related variable can be fully estimated if its information entropy is reduced to 0 by a set of optimal variables. Independent variables are first sorted by mutual information, then the conditional entropy of the quality-related variable is calculated iteratively until the result is close to 0, which indicates that the quality-related variable can be fully estimated by the currently selected variables. The final variable set is determined by further excluding the redundant variables according to information gain. The effectiveness and superiority of the method are validated through a case study on an industrial naphtha cracking furnace.