Abstract: In modal parameter identification from ambient vibration responses through power spectral analysis, manual intervention is often required to identify frequency peaks. To overcome this limitation, a convolutional neural network (CNN)-based approach is proposed. The convolutional neural network is trained based on the graphical features of natural frequency peaks and false peaks to obtain a network model for identifying real peaks, so that modal parameters can be identified automatically. Due to its graphical features, this network model has good generalization ability and can perform transfer learning. This method combines the advantages of power spectrum method and convolutional neural network to solve the problem of a large number of false modes generating from traditional power spectrum method under noise interference and relying on manual identification of natural frequencies, significantly improving the accuracy and efficiency of modal parameter identification. To verify the feasibility and robustness of the method, ASCE Benchmark structural experimental data and vibration responses of the in-service bridge are analyzed. After the acceleration response are analyzed by power spectrum analysis, training and validation datasets are generated for training convolutional neural network models. The results show that this method can identify the peaks of natural frequency with high accuracy and has good robustness of noise, providing efficient and reliable technical support for modal analysis for the structures under ambient vibration.