When the working condition of a wet ball mill is changed, the distribution of real-time data and modeling data is inconsistent. It is difficult to accurately measure the load parameters by using the traditional soft sensor algorithm based on historical data. Therefore, a transfer learning strategy is introduced, and the robustness of the model is improved by the multi domain mechanism. The process is to preprocess and extract the characteristics of multi working conditions data, and the distribution of the edge and the conditional distribution is obtained by joint distribution fitting. Then the maximum mean discrepancy is used to measure the distribution of adaptive data, and the calculated results are applied to the regression weighted. Finally, the target domain data is used for load forecasting. The practicability and effectiveness of the model are illustrated by comparing experiments and cross experiments.