Research On Gearbox Typical Fault Identification Method Via Deep Neural Network

Gear transmission system has been widely used in aerospace,petrochemical metallurgy,wind power generation,rail transit,weapons and equipment and other industries.With the development of modern industry,mechanical equipment is required to be faster,more accurate,and more reliable.Therefore,the research on intelligent fault identification method with high efficiency,high precision and strong generalization ability is of great significance and practical engineering value to ensure the normal operation of machinery and improve production efficiency.In recent years,artificial intelligence and deep learning theory have developed rapidly,which provides a solid theoretical basis for fault identification of deep neural network.There are still many problems in the application of deep neural network in gearbox fault identification in practical industrial environment.For example,strong noise,variable working conditions and too few fault samples will lead to high recognition false alarm rate.In order to improve the accuracy of fault identification of industrial service equipment,considering the factors such as strong noise,variable working conditions,small samples and cross domain identification,this paper studies how to enhance the sensitivity,robustness,and characterization of fault characteristics.The main research work of this paper includes:(1)In order to reduce the sensitivity of fault features in industrial environment,a referential network is proposed.The reference network includes multiple reference units,and the convolution kernel length of the reference unit is set according to the fault characteristics.The reference unit takes the health signal as the reference,eliminates or weakens the environmental noise or adverse factors in the fault signal,enhances the sensitivity of fault features,improves the accuracy of mechanical equipment fault identification in strong noise environment,and verifies the superiority and effectiveness of the reference network on the gear and bearing data set.(2)Aiming at the problem of poor robustness of identification model caused by time-varying working conditions affecting feature distribution,a depth interpolation neural network is proposed.The deep interpolation neural network is composed of multiple sub network units,a weight unit and fusion unit.The extracted features are interpolated and fused through the weight unit and fusion unit,which enhances the robustness of fault features and improves the recognition accuracy of unknown load samples.Compared with several public data signals,the recognition effectiveness and correctness of the network model are verified.(3)Aiming at the problem of over fitting of network model caused by small sample training and weak feature representation,a style generative adversarial network is proposed,including generator,discriminator and filter.The style generative adversarial network extracts fault features from the potential space of limited samples,generates new fault sample data by recombining features,and obtains a new data set with real sample features through clustering t-SNE screening.It not only effectively expands the small sample data set,but also improves the integrity of the overall characteristics of the sample.Through the comparative analysis of the complete data of the public data set and the generated sample data,the effectiveness of the style generative adversarial network is verified.(4)Aiming at the problem of low recognition rate caused by strong noise,variable working conditions,small samples and cross domain,a cross domain fault recognition method based on multi-network cooperation of reference network,deep interpolation neural network and traditional anti migration network is proposed.This method not only improves the success rate of source domain data transfer,but also has strong noise resistance and robustness.The cross-domain fault recognition experiments are carried out on different bearing and gear data sets.Compared with the traditional intelligent recognition network and transfer network,this method is effective and has high recognition accuracy.