In order to solve the problem that the internal fracture state of rock mass under external stress is difficult to be judged accurately by experience， a rock fracture state recognition method based on multi-feature information fusion and least square support vector machine （LSSVM） is proposed. Firstly， the improved ensemble empirical mode decomposition（EEMD） is used to decompose the acoustic emission signal of sandstone， and a set of effective stationary intrinsic mode function（IMF）components are obtained. Moreover， the auto regressive（AR）modeling of each IMF component is carried out， and the coefficients of the AR model are extracted as the time domain eigenvector. Secondly， the frequency domain characteristics of the sandstone acoustic emission signal are analyzed by bi-spectral analysis， and the singular value decomposition of the bi-spectral matrix is carried out to calculate the singular spectrum and construct the frequency domain eigenvector. Finally， local linear embedded （LLE） are used for feature reduction， and the fusion eigenvector is normalized as the input of LSSVM， the sandstone fracture state is used as the output. Particle swarm optimization （PSO） algorithm is used to automatically optimize the parameters to realize the diagnosis and identification of rock fracture state. The results show that the fusion feature has strong robustness and the recognition rate is improved by 6% compared with the single time domain feature.