Abstract: In order to improve the recognition accuracy of rock burst disaster, a semi-supervised over-sampling framework is proposed to recognize the microseismic signal. Principal component analysis(PCA), Wavelet transform and Fisher linear discriminant(FLD) are used to extract features of the multi-channel data set. Then, over-sampling method and semi-supervised learning method are used to construct a balanced data set based on the features. Furthermore, six classifiers are trained based on the balanced microseismic data set, and the original data set is from Yancon Group Company. Experimental results show that the semi-supervised over-sampling framework is actually able to improve the classification accuracy. Compared with the method which only uses over-sampling algorithm, the semi-supervised over-sampling framework can obtain higher values of recall and F1 score on 5 classifiers, with CPLE or SELF as the semi-supervised method. The six classifiers are KNN, LR, FLD, RF, SVM and Adaboost. Overall, this framework not only extracts the compressed representation of the high-dimensional data set, but also deals with the microseismic signal recognition problem based on a balanced data set.