Since Huang proposed the Empirical Mode Decomposition (EMD) in 1998, mode decomposition has been widely studied, but EMD and relative developed algorithms are still generally lack of adaptability and mathematical theory. This paper propose a new mode decomposition algorithm called Neural Mode Decomposition (NMD) based on Fourier neural network (FNN) and frequency clustering. Firstly, a FNN is constructed to decompose and learn the information of each amplitude modulation frequency component and non-periodic component in the raw data. Secondly, the frequency components obtained by the FNN are clustered into multiple Intrinsic Mode Functions (IMF) with separated spectrum based on the energy of each frequency component learned by FNN. Practical decomposition results on a series of artificial and real data show that NMD algorithm can effectively implement mode decomposition, better reflect the characteristics of raw data than EMD, and has higher adaptability than Variational Mode Decomposition (VMD).