Research On Highly Robust Transfer Failure Diagnoses For Planetary Gearbox

Planetary gearboxes are widely used in the key part of transmission system of vehicles such as speed changer,speed reducer and differential,because of their small size,large transmission ratio and high efficiency.Besides,planetary gearboxes are also applied in the driving system of helicopters and wind turbines.However,planetary gearboxes are prone to failures thanks to the bad working conditions,load fluctuation,fatigue stress and other problems.The failures of planetary gearboxes will lead to serious safety hazards and huge economic losses.Unfortunately,it is difficult to achieve real-time and accurate health monitoring and fault diagnoses by traditional manual maintenance methods.Therefore,some deep learning and transfer learning techniques are applied in this study to achieve real-time,intelligent and accurate fault diagnoses for planetary gearboxes.In this study,rectified linear Tanh(ReLTanh)is proposed to solve vanishing gradient problem of Tanh,and multi-scale transfer mechanism(MSTM)and multi-scale transfer voting mechanism(MSTVM)are proposed to improve the classical domain adaption models.The main contributions are shown as follows:(1)Tanh is a classical activation function with good noise robustness and nonlinear activation characteristics.However,in the deep model,Tanh will cause the vanishing gradient problem,which limits the fitting ability of the models,and affect the final diagnostic accuracies.In order to keep all the advantages and solve gradient vanishing problem,ReLTanh is proposed.ReLTanh is composed of two lines in positive and negative intervals and a Tanh curve in the middle.The slopes of the two lines can be adjusted by adaptive thresholds,so as to seek for a balance among gradient stability,nonlinearity and noise robustness.ReLTanh can help deep models to achieve fault diagnosis with high robustness and high accuracy.The feasibility and effectiveness of ReLTanh were verified theoretically and experimentally.(2)Plenty of labeled samples are needed during the training process of deep learning model,but they are unavailable in practice.Thus,domain adaption transfer models are applied to failure diagnoses,and they use the experience learned from source domain with sufficient labeled samples to solve the diagnosis tasks in target domain without labeled samples.However classical models can output only one transfer feature with high dimension and fixed scales,it is likely to lose important information while performing domain confusion operation.In order to solve these problems,MSTM is proposed in this paper to remold and improve the classical deep domain adaption models.The MSTM block has four parallel pipelines composed with multi-scale convolution and pooling operations,and multi-scale transfer feature can be outputted.MSTM is helpful to reduce the distance between domains,accelerate the convergence of the model,and improve the diagnostic accuracy in the target domain.(3)MSTM strengthens the transfer loss only,and it is likely to destroy the balance between classification loss and transfer loss.In order to solve this problem and further improve domain adaption models,MSTVM is proposed.MSTVM can remold the MSTM block based on Inception block,and multiple branch classifiers are constructed based on bagging strategy.So MSTM is kept,and multi-scale voting mechanism is introduced to enhance both transfer loss and classification loss at the same time.MSTVM greatly improves the robustness of the transfer models to random initialization parameters and sample sets.