Research On The Failure Mechanism And Residual Life Prediction Method Of Rolling Bearing Based On Vibration Excitation Response

Rolling bearings have a complex internal structure,work in relatively harsh environments and often operate under changing conditions,making them one of the most prone to failure and easily damaged components in rotating machinery.Bearing failures can directly lead to machinery downtime,economic losses and even major accidents.Health monitoring and Residual Useful Life(RUL)predictions for rolling bearings are therefore of great importance.In the context of industrial big data,it is easy to collect the spatial and temporal state data of rolling bearings through various sensors,which provides good data support for the implementation of bearing health management,and the powerful data feature extraction capability and non-linear function mapping capability of Deep Learning(DL)also provides strong technical support for the research of rolling bearing residual life prediction.This paper takes rolling bearings as the research object,and investigates the construction of bearing health factors and remaining life prediction methods based on deep learning on the basis of the failure mechanism of rolling bearings under vibration excitation response.The main research contents are as follows:(1)Research on rolling bearing failure mechanism based on vibration excitation response.Taking the rolling bearing of planetary gearbox as the research object,a simulation model of specific fault vibration excitation response under different operating conditions is established.Firstly,the vibration transmission path of the rolling bearing of the planetary gearbox is analysed,and the meshing frequency and fault characteristic frequency are calculated.On this basis,a rolling bearing fault simulation model is established,and the influence of each parameter on the performance of the model is studied and analysed.The amplitude-frequency modulation model,the meshing force and the meshing phase difference model are established to simulate the numerical signal of the bearing under normal operating conditions.Considering factors such as bearing load variation,the vibration signal under bearing inner and outer ring fault is simulated numerically,and finally the correctness of the model is verified by experiment.(2)From the above analysis of rolling bearing failure mechanism,it can be seen that the vibration signal of the bearing will be more complex when it fails,and the extraction of the deep degradation characteristics of the bearing needs to consider the denoising ability and feature extraction efficiency of the model.In view of this,a Manifold Regularization Stack Denoising Autoencoder(MRSDAE)combined with Kernel Principal Component Analysis(KPCA)is proposed.Rolling bearing health factor construction method.The unsupervised deep learning method Stack Denoising Autoencoder(SDAE)is used for deep feature adaptive extraction of rolling bearing full-life vibration signals;a streaming regularisation term is added to the SDAE network to maximise the retention of the internal structure of the encoder hidden layer data and improve the model’s validity.The constructed model makes full use of the powerful feature extraction capability of SDAE to extract the deep features of the bearing vibration signal,and the extracted deep features are then input into the KPCA algorithm network for non-linear dimensionality reduction to obtain rolling bearing health factor curves with certain trend,monotonicity and time correlation.The proposed method is validated by XJTU-SY and PHM2012 public dataset,which shows that the proposed method has good generalization ability,is generalized and can simplify the feature extraction steps.(3)A rolling bearing RUL prediction method based on MRSDAE-KPCA combined with Bidirectional Long and Short Term Memory(Bi-LSTM)network is proposed.Using as input the bearing HI curves extracted from the MRSDAE-KPCA based model in the previous section,the Bi-LSTM network,which can fully consider the past,present and future information of the time series data,is used to construct the prediction model in the bearing life prediction stage.By comparing the effects of different network layers and different optimization algorithms on the prediction model,the best network layer and optimization algorithm are selected and the prediction model is constructed.The proposed method is compared with other deep learningbased methods,and the results show that the proposed method is significantly better than other methods in the bearing RUL prediction task.(4)The rolling bearing RUL method based on Bi-LSTM networks proposed in the previous chapter completed the prediction of the remaining service life of the bearings using the bearing health factors extracted by MRSDAE-KPCA.However,there is some overfitting phenomenon in the operation of the model and the ability to capture the timing features needs to be improved,which limits the application of the method.In view of this,an unsupervised Multi-scale Densely Gated Recurrent Unit(UMDGRU)based rolling bearing RUL prediction method is proposed.The method consists of a pre-trained MRSDAE network initialisation feature layer,a multiscale layer,a jump gated recurrent unit layer and a dense layer.By incorporating the multiscale and dense layers,the network has the ability to capture time series data features and integrate attentional information at different time scales.The network is also an end-to-end network,combining feature extraction methods with RUL prediction models,and pre-training using an unsupervised MRSDAE network,which is efficient,fast and easy to apply.Validation on the dataset shows that the proposed method has better generalisation capabilities and achieves higher prediction accuracy than other data-driven methods.