Abstract:The stellar spectroscopy in the Department of Astrophysics has extremely important research prospects. Since the LAMOST (Large Sky Area Multi-Object Fiber Spectroscopy Telescope), a large-scale scientific astronomy survey project independently developed by China which has been the world's largest scientific device for obtaining space spectroscopy data. The sixth data (DR6) is currently available to astronomers around the world. The classification of stellar spectral data is extremely important in studying the content of astronomical observation data analysis. In order to take into account both fast operating speed and accurate classification accuracy, this paper analyzes F, G, K, M-type star spectrum in DR5 based on the bias estimation convolutional neural network method (BECNN). The core idea of BECNN is to replace the bias parameter of the flexible maximum transfer function with the bias parameter of the Taylor expansion of the deviation function, thereby reducing the error and improving the accuracy. Comparing this method with the existing two algorithms, the Neural Network (NN) and the Convolutional Neural Network (CNN). The results show that the automatic classification accuracy of the BECNN algorithm in F, G, K, and M star spectra is 93.177%, 88.349%, 93.807%, 89.255%; the CNN algorithm is 91.646%, 87.671%, 92.701%, 89.054%; the NN algorithm is 90.819%, 87.417%, 91.325%, 88.092%, and the fusion of pairwise star spectral data is used as a test Sample set for further verification. It is concluded that the automatic classification accuracy of BECNN spectrum is higher than CNN and NN. This method has good reference value in future research on special celestial body search and fine classification of star spectrum.