A method of distributed mutual information is proposed for selecting secondary variables in a soft sensor. The mutual information between the predicted primary variable and the secondary variables is obtained by estimating the probability distribution of every secondary variable and the predicted variable. This information indirectly reflects the linear or nonlinear correlations between the predicted variable and the secondary variables. A threshold value is obtained by t-test approach as a criterion to judge the correlation of variables. Subsequently, the variables whose mutual information is greater than the threshold value are further screened to be selected as the relevant variables or to be discarded as weakly relevant variables. Finally, a soft sensor model is built based on the support vector machine algorithm with the selected secondary variables.