Rock mass quality classification is an important foundation for engineering design and construction， and it is also an important research topic at present. Taking into account the complexity and uncertainty of rock masses and the differences of rock masses in different regions， machine learning methods are widely used in rock mass quality evaluation. A case database was established by collecting 55 sets of measured samples and 17 sets of interpolated samples from different regions. RQD， uniaxial saturated compressive strength （Rw）， rock mass integrity coefficient （Kv）， structural plane strength coefficient （Kf）， groundwater seepage volume （ω） are determined as the input conditions of the model， and the rock mass quality grade is the output condition. Based on the case library， an improved two-stage regression migration learning （Two-stage TrAdaBoost.R2）-Isolated Forest multi-factor rock mass quality grade prediction model is proposed. The advantages of this model are of follows： （1） The idea of migration learning is introduced into the rock mass quality classification. Taking into account the differences of rock masses in different regions， using the idea of weight adjustment， a sample similar to the target rock mass is selected from the known samples to assist in the training of the model. Solved the problem of insufficient training samples， and achieve high-precision prediction of the model when there are fewer learning samples in the target field. （2） When using the migration algorithm to classify the quality of the rock mass， the classification problem is transformed into a regression problem. The regression algorithm is used to predict the quality of the rock mass. Only one model can be used to judge the multiple levels of the sample， which overcomes the limitation of the classification algorithm in solving the multi-classification problem. （3） The sample weight is adjusted in two stages， which solves the problem of the source domain weight falling too fast in the TrAdaBoost algorithm. （4） Combined the Two-stage TrAdaBoost.R2 algorithm with the Isolated Forest anomaly detection algorithm，the influence of abnormal data on the model is eliminated， and the stability of the model is increased. The trained model was used to make multiple judgments on 12 samples of the first phase underground project of Guangzhou Pumped Storage Power Station， and the prediction accuracy of the model was evaluated by the mean square error. The average mean square error of the test sample is 0.067， and the prediction accuracy is high. It proves that the model has good performance in the application of rock mass quality grade prediction.

http://www.goldsci.ac.cn/article/2021/1005-2518/1005-2518-2021-29-6-826.shtml