Research On Rolling Bearing Life Prediction Method Based On Deep Learning

The main role of rolling bearing is to bear and transfer the mechanical internal structure interaction produced by the alternating load,is one of the most important parts of mechanical equipment.In the actual working condition,the bearing in the process of operation,the complex dynamic characteristics between components,as well as assembly and cooperate with the existence of external factors,makes it very vulnerable to erosion and abrasion damage,make its reliability is greatly reduced,this to a certain extent affect the accuracy and reliability of the system.Due to the defects or faults of rolling bearings,the vibration imbalance of rotating shafts of mechanical equipment often occurs,and even the damage of rotating shafts occurs.In severe cases,catastrophic injuries and deaths may occur.If the remaining service life can be predicted accurately,it is of great practical significance to make reasonable maintenance plan and correct preventive maintenance decision.Therefore,in this thesis,the RUL prediction method of rolling bearings was studied,and an improved convolutional neural network(CNN)combined with bidirectional long and Short time memory neural network(BiLSTM)was proposed to predict the remaining life of rolling bearings.First of all,the data collected by rolling bearings are time-domain signals.The horizontal axis of the data set is time series,and the vertical axis is the amplitude of bearing vibration,which reflects the variation rule of vibration amplitude of rolling bearings with time in the process of rotation.The vibration time domain signal is converted into frequency domain amplitude by fast Fourier transform,which contains more effective information in frequency domain and can get the fault information of bearings more quickly.Therefore,the interference information can be selectively filtered to achieve fast and effective extraction of data features.In addition,considering that the FAST Fourier transform has a symmetric structure,in order to reduce the computational complexity,only the first half of the amplitude signal is taken for subsequent signal processing and model training.Secondly,the local features based on CNN are used for deep feature mining,which is used as the input of the BiLSTM network for further processing of feature data,and a trend quantitative health indicator is constructed through the network output according to the sum of the standard deviations of the data in the two directions of the bearing.Subsequently,the model data is processed by MA smoothing and curve fitting to predict the performance degradation trend,and an evaluation model is constructed to evaluate the effect of the remaining service life prediction of the rolling bearing.It is verified by experiments that the constructed health index can be effectively applied to the evaluation of bearing degradation state.Finally,a new neural network,an improved CNN-BiLSTM network,is proposed.The network uses the attention mechanism to quickly extract sparse data by ignoring or eliminating information irrelevant to the target problem,and the self-attention mechanism can reduce external information.Relying on,combined with the BiLSTM layer to extract the global features of the original data signal,and according to the sum of the standard deviations of the two directions of the bearing,independently construct a trend quantitative health indicator,and determine the failure threshold at the same time.After that,an evaluation model based on Scoring function and RMSE is constructed and applied to the evaluation of the residual service life prediction performance of rolling bearings.Finally,it is compared with the prediction results obtained by other four deep learning methods,namely LSTM neural network,CNN-LSTM neural network,BiLSTM neural network and CNN-BiLSTM neural network,and the proposed improved CNN-BiLSTM neural network is verified.