Research On Deep Attention Network For Aero-engine Bearing Health State Recognition

The structure of aero-engine is complex and the working conditions are changeable,which makes the bearing of aero-engine prone to failure,thus causing the aero-engine to fail to work normally,and even causing an air accident.Therefore,it is of great theoretical significance and engineering value to carry out the research on the health status recognition of aviation bearings.Aiming at the problem that the traditional bearing health state recognition method is limited in the application of aviation bearing engineering background,this thesis constructs a deep attention framework,and proposes three evaluation models based on this framework to achieve high-precision recognition of aviation bearing health state under three engineering backgrounds of non-stationary working conditions,strong noise interference and sample imbalance.Aiming at the problem of low recognition accuracy of traditional methods in dealing with aviation bearing signals under non-stationary conditions,this thesis proposes an evaluation method based on multi-head attention.This method adopts the’ feature enhancement ’ strategy.Firstly,the feature extraction module of CNN and GRU is constructed to obtain the deep features of the signal.Then,a multi-head attention module based on signal features is established to make the network pay attention to and fuse the information of different representation subspaces to improve the saliency level of fault features.Finally,the effectiveness of the method is verified by the vibration experiment of aviation bearing.Aiming at the problem of low recognition accuracy of traditional methods when dealing with aviation bearing signals under strong noise interference,this thesis proposes a multi-scale attention recognition method.This method adopts the ’ threshold denoising ’ strategy.Firstly,an adaptive threshold denoising sub-module is constructed and introduced into the multi-scale feature extraction module to reduce the noise interference in the multi-scale features.Subsequently,an attention fusion module based on multi-scale features is constructed to organically fuse the multi-scale features after noise reduction,thereby strengthening the bearing fault features in a noisy environment.Finally,the effectiveness of the method is verified by the vibration experiment of aviation bearing.Aiming at the problem of low evaluation accuracy of traditional methods when dealing with aviation bearing signals with unbalanced distribution of samples,this thesis proposes a deep metric attention recognition method based on the ’ double loss synergy ’ strategy.This method first constructs a channel attention mechanism and introduces it into the feature extraction module to enhance the feature extraction ability of the network.Subsequently,the measurement and classification loss coordination technology is established and applied to network training,which makes the distance between various types of samples increase,the distance within the class decreases,and the separability is stronger.Finally,the effectiveness of the method is verified by the vibration experiment of aviation bearing.Finally,an aero-engine bearing health status recognition system based on Python language is developed,which supports users to perform offline training,online recognition and visual analysis of aero-engine bearing signals.