Aeronautical auto-transformer rectifier unit (ATRU) is the key power conversion device of aircraft high-voltage DC power grid. It is continuously affected by high temperature, mechanical stress, load fluctuation and other factors during operation, then its internal components may appear corresponding failure, which can lead to threaten the reliable operation and continued airworthiness of the aircraft. The spectrum of the fault signal in the rectifier part of ATRU is difficult to distinguish and the diagnostic accuracy is low, a fault diagnosis method based on genetic algorithm ( GA) combined with Bayesian regularization back propagation neural network (BRBPNN) is proposed. Firstly, an ATRU fault simulation model is implemented and then the collected signals are processed by means of time-frequency analysis so as to mine the feature information of different fault states. Subsequently, genetic algorithm is used to optimize the initial weights and thresholds of BRBPNN and the optimal GA-BRBPNN diagnosis model is established. The feature samples are introduced into the diagnosis model for fault identification and model performance testing. Finally, the experiment platform of fault simulation is built and the measured fault data is used to validate the method. The experimental results show that the diagnostic accuracy of the proposed method can reach 99. 46% for the simulated faults and the method can diagnose and identify all the samples to be tested for the actual faults. Therefore, the method based on GA-BRBPNN has good diagnostic effect and high practical value.