A feature fusion classification model of lung sound recognition based on cnn was proposed to solve the problems of high memory occupancy, low classification accuracy, inadequate feature extraction and unbalanced training data in the current lung sound classification model. Using data enhancement technology to expand the data set, the problem of insufficient parameters in CNN training is alleviated, and the classification accuracy is improved. In particular, a feature fusion strategy is introduced, in which the feature spectra obtained by the short-time Fourier transform (STFT), the Meir frequency reciprocal coefficient (MFCC), and the wavelet transform (WT) are fused in series to enhance the feature representation and enable the model to learn higher-level and more abstract feature representations. Experiments conducted on the ICBHI 2017 dataset validated the validity of the proposed model, and significant improvements were observed on various performance indicators. Specifically, the specificity was significantly improved by 45.15%, the accuracy by 19.16%, and the F1 score by 13.57%.