Abstract: Abstract: Rolling bearing is an important part of agricultural machinery, its health directly affects the normal operation of the equipment. However, the vibration signal of the bearing is weak, has complex non-stationarity and nonlinearity. In actual working conditions, bearing fault usually exists on the form of compound faults. Therefore, it is of great significance to extract feature information and identify fault intelligently for bearing compound faults. This paper proposed a diagnosis method combining improved hilbert-huang transform (IHHT) and convolution neural network (CNN). Hilbert-huang transform (HHT) includes empirical mode decomposition (EMD) and hilbert transform, which can analyze signal according to its local time scaled characteristic, but there are also problems in HHT, such as mode mixing and false intrinsic mode function (IMF) components. Ensemble empirical mode decomposition (EEMD), which utilizes statistical characteristics of uniform frequency distribution of Gaussian white noise, can be used to control mode mixing in HHT, however, it is difficult to determine the amplitude of white noise in EEMD. In this paper, multiple population differential evolution (MPDE) algorithm was used to optimize the added white noise amplitude for the adaptive selection of parameters. For the problem of the false IMF component, the criterion of selecting sensitive IMF based on the minimum distance between IMF components and original signal was discussed, MPDE-EEMD and screening theory of sensitive IMF components was used to improve HHT in order to adaptively extract time-frequency characteristics of fault signal. In order to accurately to identify the health status of equipment, deep learning model based on the convolution neural network was used to realize the intelligent fault identification. Based on the AlexNet model, the CNN network model for rolling bearing fault diagnosis was constructed by traversing all possible combinations of models. For training network, the IHHT time-frequency diagram generated by training set was input into CNN to continuously update network parameters. Then the model was applied to testing set and output the fault identification result. Two different bearing faults experiments which contain single faults and compound faults were used to test the feasibility of the IHHT+CNN method. The experiment results showed that the diagnosis accuracy of IHHT+CNN method for single fault and compound fault was 100% and 99.74% respectively. In comparison with the BP neural network, discrete wavelet transform (DWT) +CNN and short-time fourier transform (STFT) +CNN, the results of the presented method in this paper showed more accurate than other diagnosis results, under the different experimental loads and conditions, the robustness of bearing fault diagnosis method was verified. Compared with