Abstract: As a key production quality index of grinding process, particle size is of great importance to closed-loop optimization and control. This is because controlling particle within a proper range can improve the concentrate grade, enhance the recovery rate of useful minerals, and reduce the loss of metal in the sorting operation; thus, the particle size determines the overall performance of the grinding process. In fact, it is not easy to optimize or control the practical industrial process because the optimal operation largely depends on a good measurement of particle size of grinding process; however, it is difficult to realize the real-time measurement of particle size because of limitations of economy or technique. Employing soft sensor techniques is necessary to solve the problem of particle size estimation, which is particularly important for the actual grinding processes. Considering that soft sensors are applicable in many fields, the data-driven soft sensor will be a useful tool for achieving particle size estimation. However, most of the iron ores processed in China are characterized by hematite with unstable properties, and the slurry particles exhibit magnetic agglomeration, giving rise to a large number of outliers in the collected data. In this case, there are gross errors in the particle size estimation model constructed based on the data and thus unreliable measurements. Meanwhile, the traditional feedforward neural networks have the disadvantages of slow convergence speed and easily fall into local minimum during the prediction process. A single model tends to lack superiority in sound generalization, and the performance of existing ensemble learning methods will be worse under outlier interference. Therefore, in this study, based on the improved random vector functional link networks (RVFLN), the Bagging algorithm is incorporated into an adaptive weighted data fusion technique to develop an ensemble learning method for particle size estimation of grinding processes. Experimental studies were first conducted through benchmark regression issues and then validated by the samples collected from an actual grinding process, indicating the effectiveness of the proposed method.