Abstract: As a critical equipment in nuclear power plants, valve leakage would have an impact on system safety and stable operation. Using acoustic emission technology to identify valve leakage and distinguish between different leakage modes of internal and external leakage is helpful for targeted maintenance in the future, which is of great significance for improving maintenance economy. Valves are key equipments in nuclear power, and there are two main leakage modes. One mode is external leakage, which means that the fluid in the pipeline flows from the inside of the valve to the outside through the flange joint surface or the gap between the valve stem and the packing. Another mode is internal leakage, which refers to the leakage at the sealing surface of the valve disc inside the valve body due to poor closure or structural defects in the valve disc itself. In the case of internal leakage, the fluid in the pipeline cannot be completely cut off. The application of acoustic emission technology can achieve the identification of valve leakage, but it can only identify whether the valve has leaked and cannot identify whether the valve has leaked internally or externally. Distinguishing different leakage modes is of great significance for the subsequent development of maintenance plans. A valve leakage mode recognition method based on empirical mode decomposition (EMD) and random forest was proposed in this paper. Firstly, the leakage acoustic emission signal was decomposed by EMD to obtain component signals in different frequency bands. Secondly, Fourier transform was performed on each component signal to obtain its spectrum, and spectrum energy ratio features were extracted from the spectrum. Due to the wide frequency distribution range of the leakage acoustic emission signal, the number of features is relatively large. Therefore, it is difficult for general logic algorithms to accurately identify the leakage mode. Machine learning algorithms are very suitable for processing multi- dimensional data, so it can be used to classify valve internal and external leakage. Random forest is one of machine learning algorithms that uses decision trees as classifiers, which has significant advantages in processing data with high feature dimensions. Finally, an intelligent recognition model was established based on the random forest algorithm to achieve automatic identification of internal and external leakage. The method was validated using experimental data, and the results show that multiple component signals are obtained by decomposing the valve leakage acoustic emission signal using EMD, which refines the spectrum differences of the acoustic emission signal under different leakage modes. The spectrum energy ratio of each component signal was used as a feature for valve leakage mode recognition, achieving accurate characterization of valve internal and external leakage. A leakage mode recognition model for electric valves based on random forest was established, achieving accurate identification of different leakage modes.