In recent years, the spectral data of celestial bodies observed have achieved a dramatic increase thanks to the successful implementation of various projects of spectral sky survey. Therefore, higher requirements for the automatic classification and analysis of spectrum are proposed for large-scale projects of spectral sky survey. The classification problem is transformed into a regression one in this paper, and a method of spectral category regression based on the residual depth network is put forward to conduct a prediction of MK spectral subtype on stellar spectrum. The network is mainly composed of 25 convolution layers, 1 maximum pooling layer, 1 average pooling layer, full connection layer and 12 residual structures. The maximum pooling layer is used to filter features, the convolution layer to extract features, and the average pooling layer to reduce parameters and improve efficiency. The residual structure can prevent the degradation of the network, extract high-dimensional abstract features by deepening the network and improve training speed. Considering the non-zero probability of data with false labels and corrupted data, Log-Cosh is adopted as a loss function in this paper to reduce the negative impact of bad samples. 80 000 spectra that are randomly selected from LAMOST DR5 are used as the experimental data. The spectra are divided into the training set, verification set and test set according to the proportion of 7∶1∶2 after eliminating the bad value and normalizing the flow. The experiment includes two parts. In the first part, the network is adopted to carry out a prediction on the spectral subtype, and the maximum absolute error, the average absolute error and the standard deviation are used to compare the performance of convolution kernels with different shapes. The predicted value is taken as the abscissa and the label as the ordinate, and the second-order nonlinear fitting is used for all sample points in the test set, a straight line that is coincident with y=x is obtained, proving that the model can predict the spectral subtype well. The second part is concerning the internal analysis of the model. The main characteristics of the model in predicting four types of spectra, A, F, G, K, are mainly explored with the method of category activation mapping, thus endowing the model with interpretability. In the text data set, 91.4% of the spectral prediction errors of this method are within 0.5 spectral subtypes, and the average absolute error of the prediction is 0.3 spectral subtypes. It is shown that the method proposed in this paper can better predict spectral subtypes with faster speed and higher accuracy according to the comparison of the same data set with nonparametric regression, Adaboost regression tree and K-means.