Research On Fault Diagnosis And Remaining Life Prediction Of Rolling Bearings Based On Convolutional Neural Network

As one of the most important key components in rotary machine,the rolling bearing has been widely applied to people’s life,industrial manufacturing,national defense construction and other fields.Its operational state often directly affects the accuracy,reliability and other important parameters of the entire machine.Breakdown of the rolling bearing is very likely to cause major accidents,directly resulting in economic loss and even personal injury.However,the rolling bearing has a longer life dispersivity than other mechanical parts.As its service lifespan varies from different working conditions,it is unadvisable to determine the service cycle of the bearing absolutely according to the design life.Therefore,carrying out fault diagnosis and predicting the remaining life of the rolling bearing are of great significance to reinforce the service management and maximize the use value of bearingsDuring studies on fault diagnosis of the bearings,the vibration signal collected during operation is employed to evaluate the bearings for fault diagnosis.In order to effectively use the signal,features of the signal under different fault states are firstly analyzed in this paper,and the periodic impulse response characteristics,characteristic frequencies,and frequency doubling characteristics of the fault signals are studied.Then,the features contained in the original signal are identified and extracted.From the perspectives of time domain,frequency domain,and time-frequency domain,the time-domain parameter extraction,frequency-domain parameter extraction,wavelet packet decomposition,VMD decomposition and other feature extraction methods are selected and combined with theories related to sample entropy and multi-scale permutation entropy to acquire the key information that characterizes the bearing’s operating state,and establish the feature set of the experiment for fault diagnosis.Finally,a CNN bearing fault diagnosis model based on multi-scale features is established,and verified by experiments on the dataset of Case Western Reserve University,in which comparative experiments are carried out against SVM,ELM,and KELM.During studies on predicting the remaining life of the bearing,in order to further explore the gradual process of rolling bearings from startup to failure,the law of evolution of the periodic life cycle degradation is analyzed firstly to clarify the relation between the degradation state and the remaining life.Later,an STFT-CNN-based model for predicting the remaining life of the bearings is built.Before being input to the CNN model,the original signal is STFT transformed to convert the unidimensional timing signal to the time-frequency domain.At last,the model is experimentally verified by the IEEE PHM 2012 dataset,and comparative experiment is carried out against the BP neural network model.According to the experimental results,with a diagnostic rate of more than 99.7%on the CWRU data set,the multi-scale features-based CNN bearing fault diagnosis model proposed in this paper outperforms other prediction models including SVM,ELM and KELM in diagnosing the same dataset,which shows the advantages of the model.The prediction experiment of the remaining life also produces better results.With IEEE PHM 2012 dataset,the STFT-CNN-based remaining life prediction model can detect the starting point of degradation.Compared with results of experiments on BP neural network,the proposed model provides a remaining life curve that better fits the actual curve and is proven to have higher accuracy.