Abstract:

In view of the difficulties in obtaining instability data of open-pit mine dump and the small amount of sample data, a discrimination model of slope stability state of open-pit mine dump based on migration learning algorithm was proposed. According to the actual geological conditions and rainfall conditions of the dump slope of F open-pit mine in Shaanxi Province, a similar simulation test scheme of slope with different soil-rock mixing ratio was designed under the condition of rainfall. The data of water content, earth pressure and pore water pressure of the slope model were collected and processed. Considering the influence of small sample data set on the classification accuracy of GBDT model, using the idea of transfer learning, the sample weight of source domain data set and target domain data set was iteratively updated by TrAdaBoost algorithm, and the GBDT model was used as the weak learner for data sample training. Finally, according to the classification result of the weak learner, the weighted majority voting method was used to generate a TrAdaBoost-GBDT dump slope stability discrimination model based on transfer learning to improve the discrimination accuracy of small sample data label categories. The results show that the proposed dump slope stability state discrimination model has a better performance in judging the stable state than other algorithm models, and the values of accuracy, precision, recall and area under curve(AUC) are 93.3%, 87.5%, 100% and 93.8%, respectively. Compared with other algorithm models, this model can improve the accuracy of slope stability discrimination of small sample data sets, and make up for the low accuracy of machine learning classification results of small sample data sets.

Key words: dump slope, stability state discrimination, transfer adaptive boosting-gradient boosting decision tree (TrAdaBoost-GBDT), transfer learning, small samples