There are some disadvantages in the traditional model algorithm for the soft sensor, such as low predictive accuracy and poor fusion ability. Therefore, a multi-model soft sensor algorithm is proposed based on the D-S rule and difference autoregressive moving average(ARIMA) model. Firstly, the adaptive fuzzy kernel clustering method(AFKCM) and least squares support vector machine(LS-SVM) are used to establish multiple sub-models. Then the output of the soft sensor is obtained through the fusion of the sub-models based on the weight factor calculated by D-S rules. The ARIMA model is used to realize the dynamic correction to the static multi-model output. Simulation results and industry application indicate that, comparing with the traditional soft sensor, the proposed method has better predictive performance and fusion ability.