This paper presents a novel Adaboost.R training algorithm by weight trimming, which increases the training speed when dealing with large datasets and retain the forecast precision. At each iteration, the algorithm discards most of the samples with small weight and keeps only the samples whit large weight to train the weak learner. During training, only a small portion of the samples are used to train the weak learner, so the speed is increased. The method has been applied to mining safety monitoring, the experimental results show that the method has good effects for large-scale data.