In response to the problem of limited labeled samples in bearing fault diagnosis under the background of big data,a deep semi-supervised small sample classifier(DSSC)was designed and used for bearing fault diagnosis.Firstly,a deep learning classifier is constructed using a constrained Boltzmann mechanism,parameter pre-training is completed using unlabeled samples,and parameter tuning is combined with labeled samples to obtain an accurate classification model.Then,the activation function is improved to solve the gradient vanishing problem,improve convergence speed,and enhance classification performance.Finally,the signal was decomposed into a serial of intrinsic mode functions by variational mode decomposition,and sample entropies were calculated as the features,which were input to the classifier to recognize the fault types.The results of bearing fault diagnosis experiment and high-speed bearing fault diagnosis of wind turbines show that DSSC can fully utilize unlabeled and labeled samples to achieve bearing fault pattern recognition,and has advantages such as high fault diagnosis accuracy,good stability,less time consumption,and good real-time performance.It provides a new method for bearing fault diagnosis and can also be used for data cleaning,providing effective samples for intelligent fault diagnosis.
bearing fault diagnosisbig datadeep learningsmall-sample classificationactivation fu nction
NETL
NSTL
万方数据
