Research On Fault Diagnosis Algorithm Of Planetary Gearbox Based On Deep Learning And Transfer Learning

The planetary gearbox is widely used in metallurgical machinery,aviation and navigation,lifting transportation,textile and chemical industry and other modern industrial equipment.The core function is to change the speed and torque to match the power source and load.However,due to the influence of heavy load and changeable environment during the actual operation,the key parts of the planetary gearbox,such as the sun wheel and the planetary wheel,are prone to malfunction of the transmission system.What’s worse,these faults will lead to the shutdown of the whole transmission system,and even cause disastrous consequences if they are not detected and eliminated in time.To solve these problems,this thesis studies the fault diagnosis algorithm of planetary gearbox based on deep learning and transfer learning.The main research objects,content and innovations are as follows:(1)Planetary gear box usually works in harsh natural environment,thus the vibration signal is inevitably affected by the background noise.To weaken the influence of noise on the diagnosis effect,this thesis firstly investigates and analyzes the commonly used signal denoising methods at home and abroad.Based on the principle and process of VMD algorithm,then a novel denoising method called VMDSR is proposed.It could effectively select the number of modal decomposition,and eliminate the noise signal through selective reconstruction.Experimental results show that the proposed VMDSR is superior to other popular denoising methods in terms of signal-to-noise ratio and root mean square error.(2)Deep learning presents a powerful feature extraction capability,which could adaptively learn useful features from original data.Compared with conventional signal processing methods,it avoids the tedious and inefficient manual feature extraction.Therefore,this thesis first systematically studies the basic structure as well as relevant algorithms,and reviews some popular deep learning model CNN.Then the principles of residual learning unit and SE module are analyzed emphatically which jointly deduce a novel MSCNet model.Combined with the advantages of the above two components,MSCNet can not only enhance the useful information extracted from the network,but also avoid the frequent gradient disappearance in the deep networks.Finally,by integrating VMDSR and MSCNet,a planetary gearbox fault diagnosis method was proposed,and the performance is verified by several groups of comparative experiments.(3)Deep learning methods usually require training samples with rich types,adequate quantity and labeled information of fault.Otherwise,it is hard for models to learn features and results in unsatisfactory diagnosis accuracy.However,in practical engineering applications,there is a lack of labeled classification sample data,which makes it difficult to train a model with high diagnostic precision.To solve this problem,this thesis proposes an improved VMDSR-MSCNet method based on transfer learning.Specifically,a MSCNet model is firstly pre-trained by other datasets in related fields.Then a small number of planetary gearbox samples are introduced to fine-tune the parameters,so as to effectively overcome the problem of insufficient labeled samples.Finally,the experimental results show that the proposed method is superior to the conventional VMDSR-MSCNet method in terms of convergence speed and diagnostic accuracy.