Abstract:  In practical application, in order to solve the problem of low fault diagnosis accuracy due to the non-stationary nonlinear vibration signal of planetary gearbox in the early stage of fault development, a gear fault diagnosis method based on MEEMD-SDP image feature and deep residual network (DRN) was proposed. Firstly, modified ensemble empirical mode decomposition (MEEMD) was used to decompose gear vibration signals to obtain intrinsic modal function(IMF) components that could reflect gear vibration signals. Secondly, the IMF component extracted by symmetrized dot pattern (SDP) method was transformed into the feature vector of snowflake image features in polar coordinates. Finally, the deep residual network (DRN) model was introduced to realize the recognition and classification of different gear faults, and the comparison with the convolutional neural network (CNN) model was made. On the gearbox data set published by Southeast University, a comparative experiment was made on the identification accuracy of different models for gear state faults.The experiment results show that the SDP image features can fully represent the gear state information, and the average accuracy of DRN model for gear diagnosis is significantly higher than that of CNN model, reaching 98.1%. The research results have certain theoretical and practical value for improving the accuracy of gear fault identification of existing planetary gearboxes.

Key words: gear transmission; intrinsic modal function(IMF);ensemble empirical mode decomposition(MEEMD); symmetrized dot pattern (SDP);image feature; deep residual network (DRN)