| Rotary machines are widely used in various industrial fields,such as wind turbines,automobiles,aero engines,and so on.Remaining useful life(RUL)prediction is a key component of Prognostic and Health Management(PHM).Accurate RUL prediction of rotating machines can help predictive operations,avoid casualties and improve the economy.Therefore,accurate RUL prediction is of significance.Since industry 4.0,with the rapid development of sensing technology,signal processing,and artificial intelligence,data-driven methods,especially RUL prediction methods based on deep learning(DL),have developed rapidly in academia and industry.For accurately predicting RUL of rotating machinery,facing the gap that the majority of DL approaches only can learn the input based on a single learning mode,the differentiated deep learning methodology is proposed firstly,then a series of differentiated deep learning models are developed,which are combined with PHM related theories.The following are the innovations and main contents of the article:(1)To comprehensively monitor rotating machinery and accurately describe different dynamic degradation processes,a deep differentiated learning model based on the concept of interlayer ordered information was proposed for multi-source inputs(multi-source original data and multi-dimension features).Aim at the limitation that the sequence information of the mechanical equipment input lacks a clear physical meaning,a new type of long short term memory is presented called LSTM-AON networks based on attention ordered neurons,which divides the hierarchy of multi-source input via ordered neurons guided by attention,so that it can use the sequence information of neurons to realize differentiated learning based on the input hierarchy,and then improve the predictive performance.Meantime,aiming at the coupling effect between different hierarchy information,a novel LSTM variant(Cocktail LSTM)with multi-hierarchy updating rules is proposed,which can make relatively full usage of the interlayer sequence information and further improve the prediction ability of RUL.The effectiveness of the above methods was verified by IEEE 2008 simulated aero-engine data and actual wind turbine gearbox bearing life data.(2)Passive differentiated learning methods based on ordered information hierarchy have the extension limitation of multi-hierarchy,thus active differentiated deep learning methods are proposed to break through the hierarchical limitation.a)Facing the gap that most neural networks(NNs)including LSTM cannot process input data in different updating modes based on the trend degree of input,an active differentiated learning model called multi-cellular LSTM(MCLSTM)is firstly proposed.Firstly,a trend division unit is proposed to determine the input data trend,and then several cell units are designed to update the cell state in different modes according to the data trend level.Therefore,MCLSTM can extract the different degradation trends.Based on MCLSTM,a deep learning model for multi-source input has been successfully applied in RUL prediction tasks of IEEE 2008 simulation of aero-engine data and actual wind turbine gearbox bearing life data.b)Multi-source input can not only reflect the degradation trend of rotating machinery in the temporal dimension,but also reflect the health state of rotating machinery at this moment in the spatial dimension.Therefore,a spatiotemporally active differentiated learning method was proposed to fully monitor the degradation process of rotating machinery from a spatiotemporal perspective.For spatial dimension,spatial multi-differentiated convolution structure(MCNN)and spatial multi-differentiated involution structure(MINN)are proposed respectively,then multi-differentiated feature extraction is executed according to the health state level of multi-source input.For temporal dimension,multi-cellular bidirectional LSTM(MBLSTM)and multi-head recurrent gated unit(MGRU)are proposed to learn multi-differentiated features according to the(bidirectional)unidirectional trend level of multi-source input.Based on spatiotemporal multi-differentiated networks,sequence and parallel deep structure are proposed,and a concise structure is further proposed to accelerate.The effectiveness of the above methods was verified by IEEE 2008 simulated aero-engine data and actual wind turbine gearbox bearing life data.(3)The RUL prediction accuracy of deep learning methods based on pattern recognition is always reduced when the full-life data exist distribution discrepancy.At present,the transfer learning prediction methods usually reduce the single distribution distance of the high dimension features of source domain and target domain to achieve the domain invariant feature learning and improve the prediction ability of across distribution.The high-level features mean that the prediction model is deep,which also means a huge computational burden in the field of RUL prediction.However,a single distribution distance reduction is usually not as close to the real domain invariant distribution compared to multiple differentiated distribution distance reductions.Therefore,based on the above defects,this paper proposed the micro-transfer of multiple differentiated distributions,which carried out the multiple cellular differentiated distribution transfer study within the level based on the recurrent network(MCLSTM).By the recurrent characteristics,the proposed model has no need for the depth,so it is also a prospect of shallow level feature transfer.The effectiveness of the above methods was verified by IEEE 2008 simulated aero-engine data and actual wind turbine gearbox bearing life data.(4)The method based on pattern recognition requires high quality and large quantity of life data,but in practical engineering,life data are often limited and incomplete.Thus it is of significance to develop an intelligent RUL prediction method under limited samples.Firstly,the health indicators(HIs)with stable threshold are constructed based on variational encoder(VAE),then two LSTM variants based on attention mechanism named partiality LSTM(LATMPA)and Macroscopic-microscopic attention LSTM(MMALSTM)are developed orderly to select HIs and then learn differently in different time scales.By the proposed step-by-step prediction method,the RUL prediction of rotation machines under limited samples is executed.The validity of the above methods was verified by the experimental life data of gear contact fatigue in mechanical transmission laboratory of Chongqing university and the life data of actual wind turbine gearbox bearing. |
