Prediction Of Bearing Remaining Life Based On Neural Network Ensemble Learning

Bearing is one of the most commonly used rotating parts in mechanical equipment.When the bearing fails,the whole mechanical equipment will not work properly,which will cause irreparable losses to the enterprise.If the predicted Remaining Useful Life(RUL)can be used as the maintenance basis to timely take maintenance measures for the equipment to avoid the loss of production and construction of enterprises,the bearings RUL can be predicted accurately and timely during the operation of bearings.Deep groove ball bearings represent the largest market in all bearing categories,while the datadriven RUL forecasting method has the advantage of flexible modeling based on the size of data samples.Therefore,this thesis chooses a deep groove ball bearing as the research object and applies the Data-driven RUL prediction method to study the RUL prediction of bearings.Because the state data of bearings have time series attributes and there is a strong correlation between the different state data,this thesis takes the convolution neural network,recurrent neural network,and Attention mechanism as the starting point to study the RUL prediction of deep groove ball bearings.Three deep learning models,BiGRU-Attention,TCN-Decoder,and improved Transformer are built to predict the bearing RUL in the PHM2012 dataset.The prediction results show that the three models can effectively predict the bearing RUL,but their prediction effect is still not satisfactory compared with the latest research methods in this field.Based on this,this thesis applies the method of Ensemble Learning(EL)to integrate three deep learning models.Finally,the integrated model is validated with the bearing dataset of PHM2012.The results show that the performance of the integrated model is more stable and the forecast error is smaller than that of the single model.In addition,aiming at the problem that RUL prediction can not intuitively reflect the running state of equipment,this thesis proposes an improved TCN model to forecast the running state of bearings in a limited range,and then validates the model on the re-divided PHM2012 dataset.The results show that the method can accurately predict the running state of bearings for a certain period of time in the future.To sum up,there are three main parts of this thesis.Firstly,three in-depth learning models are built to predict the RUL of bearings respectively,and the experiments are carried out on the PHM2012 dataset.Secondly,aiming at the problem that the forecasting effect of the three sub-models is not perfect,the weighted average integration method is selected to integrate the three models.Finally,the data set of PHM2012 is validated.The result shows that the integrated model has a better forecasting effect than the single model.Finally,in view of the problem that RUL prediction can not intuitively reflect the running state of bearings,a method for predicting the running state of bearings in a limited range is put forward and verified on the re-divided dataset of PHM2012.The results show that this method can accurately and intuitively reflect the running state of bearings in the future.