In order to accurately separate and identify the main noise sources of internal combustion engines, a method based on IVMD-RobustICA (improved variational mode decomposition, robust independent component analysis) is proposed. Firstly, a self-adaption algorithm is proposed to determine the decomposition number of variational mode decomposition. An IVMD method based on reconstructed signal energy ratio and center frequency is developed. Then the single channel signal of measured engine noise is decomposed into multiple components by IVMD. The main components are identified according to the mutual information (MI) between signal components and the source signal. Finally, the independent components are extracted by RobustICA, and the noise sources are identified by coherent analysis and time-frequency analysis. The results show that the proposed method can separate and identify the main noise sources, such as the combustion noise, the piston slap noise, and the air compressor noise.