In order to detect stationary and transient harmonics under noise interference, an empirical wavelet transform based harmonic detection method for power system is proposed for the first time. Firstly, the empirical wavelet transform is used to extract a set of AM-FM components from power harmonic signals, and the separation of each harmonic and fundamental wave signals is realized. Then, the separated harmonic components are transformed by Hilbert transform to obtain the amplitude and frequency parameters of each harmonic and the start-end time of transient disturbance. The simulation results of multi-class harmonic signals show that the proposed method avoids the mode aliasing problem existing in traditional Hilbert-Huang transform effectively and realizes the adaptive decomposition of multi-frequency harmonic signals even at low signal-to-noise ratio. It ensures the accuracy of detection results of various parameters and has good noise robustness and real-time detection. This work is supported by Science and Technology Project of State Grid Corporation of China (No. 52199918000C).