Prediction Of Remaining Useful Life Of Products Considering Sample Difference

Data-driven product remaining useful life prediction research is of great significance for promoting the transformation and upgrading of the manufacturing industry and building a strong intelligent manufacturing country.As the starting point of prediction,the number and distribution of data samples determine the design of research methods and the accuracy of prediction results.In order to construct a product remaining life prediction method that can fit the characteristics of the data samples,on the basis of fully analyzing the characteristics of different sample data,this paper constructs the remaining useful life prediction methods under the same distribution situation of small sample,different distribution situation of small sample and the same distribution situation of large sample,respectively,as described below.(1)Aiming at the problem of remaining useful life prediction in the case of small samples with the same distribution,this paper proposes a method based on feature engineering and model integration.First,perform feature engineering on the acquired product data,and construct new features from different perspectives;secondly,screen the features obtained by different feature construction methods to obtain the core features that can characterize the degradation information of aero-engines;finally,select the appropriate base learner,and propose a weighting method for model integration to obtain the final product remaining life prediction model.Experiments on the C-MAPSS dataset show that the method proposed can effectively extract product degradation features and further improve the accuracy of product life prediction for the product life prediction problem under the same distribution scenario with a small sample.(2)Aiming at the problem of predicting the remaining life of products under different distributions of small samples,this paper proposes a method for predicting the remaining useful life of products based on the DWMTr A algorithm.Firstly,the shortcomings of the multisource weighted Tr Adaboost algorithm applied to the remaining useful life prediction problem are analyzed.Secondly,in view of the above problems,a source domain sample weighting method based on Mahalanobis distance is proposed,and the limit tree regression algorithm is used as the base learner,and then the specific process of the DWMTr A algorithm is proposed.Finally,by clustering the data in C-MAPSS to classify different working conditions,six transfer tasks are established and the prediction results are discussed.The results show that for the prediction of product remaining useful life under different distribution scenarios of small samples,the DWMTr A algorithm proposed in this paper can effectively expand the small sample data set,realize the transfer of knowledge under different data distributions,and improve the accuracy of product remaining life prediction.(3)Aiming at the problem of predicting the remaining life of products under the same distribution of large samples,this paper proposes a method for predicting the remaining useful life of products based on the combination of Hilbert-Huang transform and gated recurrent unit neural network.First,the bearing data is preprocessed using the Hilbert-Huang transform that can reflect the local characteristics of non-stationary signals;then,a root-mean-square-based health indicator is established,and the degradation stage of the bearing is divided by the top down time series segmentation method;Finally,a gated recurrent unit neural network is trained using the degradation period data after the HilbertHuang transform,and the remaining useful life of the bearing is predicted.The PHM2012 bearing data set is used to verify the method proposed in this paper,and the experimental results demonstrate the effectiveness of the method proposed in this paper.On the basis of considering sample differences,this paper provides a systematic and efficient solution for the prediction of product remaining life under different sample sizes and distribution conditions from the perspectives of machine learning,transfer learning and deep learning,which has significant theoretical and practical engineering value.