In order to solve the problem that neutral point clamped(NPC) inverters has many kinds of open faults and similar characteristics, which cause low diagnosis accuracy. CEEMDAN and improved deep belief network was used to investigate open circuit fault of three level NPC inverter. Firstly, the topology and working principle of NPC inverter are introduced and different IGBT open fault types are defined. Then CEEMDAN decomposition method and Pearson correlation-energy value criteria are combined to select sensitive IMFs, their spectrum envelopes are obtained by Hilbert transform, multiple energy and entropy of IMF components and envelope spectrums are constructed as high dimension fault characteristic vectors. Finally, multi-feature vectors are transmitted to DBN optimized by improved Harris hawk optimization algorithm. Different simulation compared experiments show that IHHO-DBN model has better diagnosis result than other traditional ones. It is concluded that there are some certain values from the proposed method in this paper.