Low Frequency Oscillation (LFO) seriously threatens the stability and security of a power grid. Most traditional LFO identification methods regard the measured signals as stationary signals or process them as stationary signals. This ignores their non-stationary characteristic. However, in a power system with high proportion of renewables and power electronics, LFO parameters have a wide range of time-varying characteristics. Traditional methods find it difficult to identify them accurately. This paper proposes a low frequency oscillation identification method based on a Convolutional Neural Network (CNN). First, the time domain features of the original signals are extracted as a preprocessing operation. Then, 16 signals of different time-domain characteristics are sent to the trained CNN networks. Finally, the full connection layer synthesizes the outputs of each network to obtain the final identification results. Simulation results show that this method can identify the frequency and damping factor of LFO signals quickly and accurately, and has good anti-noise performance. In contrast to the Prony method, it can identify whether a new oscillation mode is introduced in the process of oscillation. This work is supported by the Key Research and Development Program of Sichuan Science and Technology Department (No. 2019YFG0142) and Science and Technology Project of State Grid Sichuan Electric Power Economic and Technological Research Institute (No. SGSCJY00GHJS2000015).