Research On Gearbox Fault Recognition Based On Improved Deep Belief Network

As an important transmission component in mechanical equipment,the gearbox’s working condition will directly affect the performance of the entire equipment.Once the fault is not solved in time,it may lead to huge economic losses and even casualties.Therefore,it is of great significance to study the gearbox fault identification technology and ensure its normal operation to prevent accidental and malignant accidents.At the same time,due to the complex and variable working conditions of the gearbox,and the lack of long-term,systematic collection of fault samples under various working conditions in the engineering practice,label samples are insufficient.Therefore,gearbox fault identification becomes more difficult under variable working conditions.Based on this,this paper takes the gearbox as the research object and the fault identification under variable operating conditions as the research target,focusing on the two main aspects of the gearbox’s mixed-domain feature extraction and fault recognition technology.The specific research content is as follows:(1)On the basis of common fault modes and vibration characteristics analysis of gearboxes,according to the single domain features which are difficult to fully reflect the vibration state information and the fault recognition has one-sidedness problem.The time domain,frequency domain,and time-frequency domain features are combined to form a mixed domain feature set.Aiming at time-frequency domain feature extraction,a twodirectional two-dimensional principal component analysis(TD2DPCA)adaptive feature extraction method enhanced by Synchroextracting Transform(SET)is proposed to avoid introducing human factors.The results of gearbox failure experiments show that: Mixeddomain feature set formed by combining multi-dimensional features contains more comprehensive fault information which is conducive to gearbox fault identification.At the same time,the enhanced TD2 DPCA features by SET have higher fault recognition accuracy than Short-time Fourier transform and wavelet transform.(2)For the current phenomenon that time-frequency transformation is mostly used for qualitative analysis of equipment state and lack of quantitative evaluation,a characteristic index of mechanical equipment performance degradation combined with SET and complex wavelet structural similarity(CWSS)evaluation algorithm is proposed.In this method,SET time-frequency analysis is firstly performed for the vibration signal.Then comparing this time-frequency map with the time-frequency map of the vibration signal during normal operation according to the CWSS algorithm,to obtain the performance degradation index.The bearing life data is used for verification,and it is compared with the pulse index,root mean square value,STFT-CWSS,etc.The results show that the SET-CWSS feature can quantify the degree of bearing performance degradation and is more sensitive to early faults.(3)Aiming at the current problem that serial connections are used in combination with multi-network outputs and fusion of multi-domain features,the different contributions of different types of features to fault recognition are ignored.Based on the property of Attention mechanism,which can achieve dynamic weighting of features and improve information acquisition capability,a DBN gearbox fault recognition method that combines attention mechanism is proposed.In the method,restricted Boltzmann Machines(RBM)are first used to learn features,and secondly,multi-dimensional features are fused by dynamic weighting of attention.Finally,through gearbox fault experiments,it is proved that attention dynamic weighting improves the utilization of effective features of fault recognition,and has higher fault recognition accuracy than other fault recognition methods.At the same time,the proposed method can be used for multi-core parallel computing,reducing the training time of complex fault identification models.(4)Aiming at the difficulties of fault identification induced by the large difference of the same fault of the gearbox under variable load conditions and lack of fault samples in engineering,a cosine loss function that reduce the sensitivity of the fault identification model to the variation of load condition and semi-supervised A-DBN model are proposed.The gearbox fault experiment proves that the proposed method can improve the accuracy of fault recognition under variable load conditions,speed up the model learning speed,and have better generalization ability than other methods.In the case of insufficient label samples,the proposed method can avoid the problem that the supervised learning method cannot use the unlabeled samples.By learning the fault information contained in the unlabeled samples,it can maintain high fault recognition accuracy.In order to compare with the existing research results,the open bearing fault data sets were tested.The results show that the solutions proposed for the two problems of variable working conditions and lack of label samples have certain advantages.