Abstract: Rock mass instability prediction plays an important role in ensuring the safe production of mines. Due to the features of rock mass like nonlinearity,non-uniformity,discontinuity and anisotropy,it is difficult to describe the inter- nal state of rock mass. The method of rock mass stability evaluation based on single physical quantity monitoring is hard to meet the requirements. The real-time state of internal rock structure could be obtained in the acoustic emission signal from the rock mass instability based on the irreversibility of acoustic emission phenomenon. Therefore,originated from the D-S evidence theory,a rock mass instability prediction method based on the weighted D-S evidence theory and fusion multi-domain features is proposed. Under laboratory conditions,red sandstone,as the experimental object,is tested in uniaxial compression with type RMT-150C rock mechanics test system and the rock mass instability acoustic emission signal is collected. Subsequently,the multi-domain features in the acoustic emission signal of unstable rock mass are firstly extracted. The sensitive features of unstable rock mass are selected in each stage by the sensitive fea- ture evaluation method. Evidenced from the multi-domain sen-sitive features,an optimized BP neural network predic- tion model is used for preliminary prediction. Secondly,the basic probability distribution function of each evidence with the model outputs is calculated by the posterior probability modeling method. Then,the similarity measurement optimi- zation is applied on the basic probability distribution function. Lastly,a prediction model based on weighed D-S evi- dence theory and multi-domain features fusion is built. The experimental results show that the accuracy of prediction model of rock mass instability has been improved manifestly through the multi-domain features fusion based on D-S evidence theory in the decision level. The replacement of single feature parameter with multi-domain feature fusion in the data extraction provides more credible inputs for the prediction model. The more features fused in the model,the more accurate outputs will be obtained. Modifying the basic probability distribution function by weighed thinking and similarity measurement optimization could effectively eliminate the high-conflict evidences,which might lead to tradi- tional D-S evidence fusion algorithm failure.