Aiming at the problem of lack of nuclear magnetic resonance spectroscopy data in the prediction of total hydrogen activity of crude oil, combined with the theory of deep learning, a pre-processing method of spectral data amplification is proposed. According to the correlation coefficient of the sample input and the label, the random noise is added to the original sample to generate the virtual sample. The sample data structure is processed to facilitate the feature extraction of the convolutional neural network, and the data redundancy is added to improve the structure to further improve the integrity of the data feature extraction. A regression forecasting convolutional neural network(RF-CNN) is designed to realize the regression prediction of the total hydrogen content of crude oil. Experiments show that, for the regression prediction of total hydrogen properties, the amplified data not only solves the over-fitting phenomenon in the original data training, but also has more stability and accuracy than the traditional partial least squares(PLS) dimensionality reduction method.