Data-Model Interactive Remaining Useful Life Prediction Of Bearings

Bearing is a kind of precise component to ensure the normal operation of mechanical equipment,if its failure may lead to equipment shutdown,or even cause more serious safe accidents.With the increasingly high requirements for high-end,precision and automation of mechanical equipment,accurate prediction of bearing failure time has become an urgent demand in the industry.Timely evaluation of the current health status of bearings and prediction of the remaining effective working time is of great significance to improve the safety and reliability of mechanical equipment.Mechanical equipment based on random degradation characteristics of monitoring big data,considering the multiple performance degradation and degradation model is difficult to synergy and match each other,and the uncertainty of remaining useful life prediction is difficult to quantify under the coupling of multi-dimensional features.In this thesis,bearing is taken as the research object,focusing on its performance degradation state and based on vibration data in degradation stage.The remaining useful life prediction based on the interaction between degradation feature extraction and random degradation modeling(Data-Model Interactive)under big data is carried out.Firstly,an adaptive recognition method of degradation stage based on outlier cleaning is proposed for the interference of abnormal impulse on degradation stage recognition.The outlier detection method combining global abnormal segments detection and accurate locating of abnormal impulses is constructed,realizing accurate locating of impulse-types outliers.The screening criteria and iterative removal strategy for abnormal segments are proposed to realize the fast recognition of segments which contains abnormal impulses.After the cleaning of outliers,the 3σ method is introduced to detect the degradation points to distinguish the health stage and degradation stage of bearings.Finally,the adaptive cleaning of degradation data in the context of big data is realized,which can significantly improve the stability of bearing health indicators and degradation point identification,and provide high-quality data basis for remaining useful life prediction.On this basis,aiming at the problem of no interaction between degradation feature extraction and stochastic process degradation model,a data-model interactive remaining useful life prediction method based on linear stochastic process is proposed.On the basis of vibration data of degradation stage based on outlier cleaning,a closed-loop feedback mechanism for constructing multi-dimensional feature fusion of health indicators and degradation model is established by minimizing the mean square error between predicted life and real life.The inverse optimization of the fusion coefficient for multidimensional statistical features makes the health indicator construction and degradation process modeling interactive.Then,the prediction of remaining useful life and its distribution is obtained based on stochastic process calculation.Finally,the interactive modeling method of remaining useful life prediction modeling and degradation feature fusion is used to enhance the representation accuracy of fusion health indicator on bearing degradation state and improve the prediction accuracy of remaining useful life by simple linear stochastic process.Furthermore,considering the dynamic characteristics of bearing degradation performance and the variable sources of degradation process uncertainty,a data-model interactive remaining useful life prediction method based on nonlinear stochastic process is proposed for bearings with nonlinear degradation characteristics.Based on the idea of data-model interaction,a nonlinear wiener process model considering adaptive drift rate and measurement error is constructed.The degradation state of the model is composed of drift coefficient and potential degradation value,and the kalman filter algorithm is used to update the state adaptively.In order to consider the adaptive drift rate in the future long-term degradation process,the exponential weighted average method is used to aggregate the estimated drift coefficient sequence from the predicted time to the failure time.To achieve accurate prediction and uncertainty measurement of the remaining useful life.Finally,the predicted value of remaining useful life of nonlinear degenerate bearings converges rapidly to the real value,which improves the stability and accuracy of prediction results.Finally,the adaptive recognition method of degradation stage based on outlier cleaning and data-model interactive remaining useful life prediction method proposed in this thesis are verified experimentally on the real bearing experimental data set.The results show that the proposed method solves the problems of difficult identification of degradation stage and no interaction between degradation feature extraction and degradation model,and verifies its advantages in improving signal quality and prediction accuracy.