Research On Fault Diagnosis Of Gear Under Small Sample Based On Deep Learning

The gear is the main component of the transmission system of mechanical equipment,and its operating state plays a decisive role in the smooth operation of the entire mechanical equipment.The fault diagnosis and degree analysis of gears can effectively ensure the normal operation of mechanical equipment,reduce the operation and maintenance costs of equipment and protect the safety of operators.This paper takes the vibration signal of gears as the main object,combines the principles and methods of deep learning,and addresses the main challenges in gear fault diagnosis,such as insufficient sample size,inconspicuous one-dimensional data features,fault migration diagnosis under different working conditions and deep learning model structure parameter settings,combining gear vibration signals and feature pictures in turn.The main research elements are as follows:Aiming at the problem of lack of model training caused by insufficient sample collection in complex environments,a gear fault diagnosis based on least squares generative adversarial network combined with long short-term memory neural network for data augmentation is proposed.LSGAN is used to achieve the expansion of existing small sample data to compensate for the data volume.A quantitative approach is introduced to validate the distribution of the generated and real samples,and the comparison of different generative models highlights that the LSGAN model has high generative accuracy.The diagnostic model is trained iteratively with a dataset combining generative and real samples,and the optimal sample set mixing ratio is tested by different sample ratios.The performance is compared with other diagnostic models to verify that the proposed LSGAN-LSTM model has superior fault identification.Experimental analysis is carried out on different gear signals and the results verify that the proposed method has a strong generation capability.Considering that the traditional augmented expansion methods based on one-dimensional time-series data have problems such as poor quality of the generated data and the tendency of overfitting in the training process,an expanded diagnostic model combining a two-dimensional convolutional network and a deep convolutional adversarial network(2DCNN-DCGAN)with a two-dimensional grey-scale map as the sample object is proposed.The proposed model can well reduce the amount of parameter computation required for model training,simplify the training process and reduce the iteration time compared to the data expansion model in one dimension.The proposed method is applied to different gear fault experiments and validated in comparison.The results show that the proposed method can generate more similar data under two-dimensional conditions,which is an effective improvement in diagnostic identification of fault types and training time,and has good engineering utility.Due to the defects of small samples,the diagnostic model cannot be adaptively adjusted when the working conditions change,resulting in low fault diagnosis accuracy.A fault migration learning model combining the data generation method with a deep sub-domain adaptive network is proposed.The migration from the source domain to the target domain is achieved by enhancing the fitting of data between the source domain and the target domain.Through experimental analysis of gears under different operating conditions and comparison of the methods,it is demonstrated that the proposed method has good migration diagnosis capability under insufficient data volume.