Research On Temporal Deep Transfer Learning Prediction And Remaining Useful Life Evaluation Of Bearing Under Multi-conditional Degradation Sequences

Rolling bearings,as a key component of mechanical equipment,are prone to various failures under complex working condition such as high load and strong impacts.Effective Remaining Useful Life(RUL)can contribute to the prevention of major accidents.The RUL prediction of bearings consists of three main steps: determining the start point of early fault,constructing a Health Indicator(HI)that reflects the degradation state of the bearing and building a RUL regression prediction model.Due to various factors,there are challenges in actual engineering applications: 1)The collected monitoring data comes in the form of streaming data,which requires real-time detection of early faults and determination of the start point of RUL prediction;2)Bearings are in normal operation for a long time,with less fault data,so it is necessary to use the degradation data from different working condition as a way to explore the temporal evolution mechanism of bearings and built HI with strong temporal characterization;3)The distribution of temporal degradation features is inconsistent across working conditions,so it is necessary to transfer domain knowledge with the help of transfer learning to improve the generalization ability of the RUL prediction model.To solve the above problems,this paper takes vibration signals of bearings as data objects,focus on the temporal characteristics of degradation data,research temporal deep transfer learning algorithms theoretically,concentrate on how to transfer fault knowledge between different working conditions,and then construct multiple deep transfer learning models to solve the RUL prediction under different working conditions.The main research content include:(1)Aiming at the unstable problem of detection model for online dara,an online detection approach via self-adaptive deep feature matching is proposed.This approach includes offline and online stages.At the offline stage,a new health state assessment algorithm is first proposed,and then construct a Support Vector Data Description(SVDD)model based on deep features.At the online stage,a self-adaptive matching strategy named one-dimensional anchor is proposed.First,reconstruct online samples adaptively by one-dimensional anchor mapping.Second,the deep feature are extracted,and then input to the trained SVDD model for online early fault detection.Experimental verification is carried out on the IEEE PHM Challenge 2012 dataset,and the results show that the approach proposed in this paper can effectively online detect early fault,determine the start point of early fault and has good robustness.Aiming at the HI construction problem of insufficient amount of data under different working conditions but strong temporal characteristics,a health indicator construction method based on multi-scale deep feature transfer is proposed to dig the degradation knowledge of bearing.The method is built based on domain adversarial neural network,by introducing maximum mean discrepancy and Laplace regularizes,this model can enhance features’ discriminant ability from global scale and local scale.Beside,a new HI assessment metric is proposed,the metric can accurately evaluate the tendency similarity and geometric similarity in different working conditions scenario.Comparison experiments on IEEE PHM Challenge2012 dataset and XJTU-SY dataset show that,compared with existing HI construction methods,the proposed method outperforms several typical transfer learning methods and non-transferable deep learning methods in terms of monotonicity and correlation.Focused on the problem of the inconsistent distribution of bearing under different working conditions,a prediction method of bearing remaining useful life is proposed based on deep temporal feature transfer.The model uses the temporal convolutional network to exploit inherent temporal features from the degradation trend of multiple bearings.Then,a domain adaptation algorithm is proposed for minimizing sequence similarity,and a support vector machine(SVM)model is constructed to predict the remaining useful life.Experiments are performed on the IEEE PHM Challenge 2012 dataset.The result shows that the proposed method has good prediction performance on insufficient fault data and low dependence on the working conditions.To address the problem of how to online estimate the remaining life of bearings under the extreme condition of unknown working conditions,from the perspective of time series degradation information transfer,a new time series transfer recursive prediction model integrating prior degradation information is first constructed.Secondly,the transfer component analysis is adopted to adapt the common feature space of the predicted online degradation sequence and existing offline sequences.By extracting domain-invariant features and constructing SVM model,the RUL of online bearing can be evaluated.The experimental results on IEEE PHM Challenge 2012 dataset show that the proposed method can accurately estimate the degradation trend only using early fault data,and provide an effective solution for the online RUL estimation under an unknown working condition.In summary,this paper is oriented to the characteristics of actual engineering.By studying the deep transfer learning theory,a series of new and effective solutions are provided for the RUL prediction under different working conditions,which improves the accuracy and robustness of the prediction model and has significant theoretical and engineering application value.