Stellar spectrum classification is one of the important tasks of stellar spectrum analysis. Chinese large-scale survey project LAMOST can obtain massive stellar spectral data. In order to efficiently classify massive stellar spectral data, especially stellar spectral subtype data, we need to study fast and effective stellar spectral automatic classification algorithms. This paper proposes a hybrid deep learning algorithm based on Transformer feature extraction, Bert+svm (abbreviated as Besvm), to classify the spectral subtypes of type A stars automatically. The algorithm takes 26 line indices of the spectrum of A-type stars as input features and uses the Bert model to perform a deeper learning of the 26 line indices. By learning the internal correlation of the 26 line indices, it extracts the spectrum more conducive to the A-type stars classification characteristics. The extracted new features are input into the classifier algorithm Support Vector Machine (SVM for short), and then the three subtypes A1, A2, and A3 of the A-type star spectrum are automatically classified. Previously, the SVM algorithm has been applied in the stellar spectrum classification task, and some derivative SVM algorithms also have a higher classification accuracy rate in the stellar spectrum classification task. Compared with the SVM algorithm previously applied to the stellar spectral classification task, our hybrid deep learning algorithm is less affected by the signal-to-noise ratio of the data, and the low-signal-to-noise ratio data can also have a higher classification accuracy. The amount of data used is relatively small. This paper verifies the effectiveness and superiority of the algorithm through five sets of experiments: Experiment 1 is used to compare and select excellent kernel functions, and through the matching experiment of spectral data, the radial basis kernel function RBF is finally selected; Experiment 2 compares the performance indicators of the Besvm algorithm with the other four traditional excellent algorithms verify the superiority of the Besvm algorithm; Experiment 3 is used to test the stability of the Besvm algorithm; Experiment 4 analyzes the influence of the amount of data on the Besvm algorithm; Experiment 5 analyzes the influence of different signal-to-noise ratios data on the classification accuracy of Besvm algorithm. The analysis of comprehensive experimental results shows that the hybrid deep learning algorithm Besvm proposed in this paper can still maintain a high classification accuracy rate on a small-scale data set with a low signal-to-noise ratio. The overall classification error rate of Besvm is below 0.01, which is much lower than the error rate of the classic traditional machine learning algorithm LDA algorithm, Bp neural network algorithm, SVM algorithm and Xgboost algorithm. The classification accuracy is too limited by the number of hidden neurons.