Research On Equipment Health Assessment And Remaining Useful Life Prediction Method Based On LSTM

With the rapid development of science and technology,equipments in the fields of aerospace,manufacturing,energy,and metallurgy are becoming more and more intelligent and complex.Traditional maintenance strategies have problems such as"under-maintenance" or "over-maintenance",which is difficult to meet actual maintenance demand.In order to ensure the safe and reliable operation of the equipment,in the big data era,it is of great significance to study the data-driven prognosis and health management(PHM)method for the complicated scenarios of equipment systems.The core connotation of PHM is based on state monitoring data,using intelligent algorithm to assessment system health and predict remaining useful life(RUL)in real time,and based on this,develop maintenance plan.As a typical deep learning model,long short-term memory(LSTM)network has been widely used in machine translation and time series prediction,and has achieved good results due to its advantages in extracting time-dependent features.Therefore,this thesis focuses on PHM core of complex equipment systems,combined with LSTM network,and deeply study the data-driven health assessment modeling method and RUL prediction method.The details are as follows:1.A new health assessment method based on LSTM network and variational autoencoder(VAE)hybrid model is proposed.The method is trained in an unsupervised manner.On the one hand,by combining with the LSTM network,the long-term and short-term time dependence in the state data can be effectively extracted;On the other hand,combined with the variational autoencoder,the state data and the correlation between the dimensions can be mapped into the continuous hidden space to achieve deep feature extraction,and it also has good robustness to noise.In this thesis,the performance of the algorithm is verified based on the PHM08 public dataset,and a comparative experiment is carried out with the traditional modeling method.The results show that the proposed method has better monotonicity and robustness,and its effectiveness and superiority are verified.2.A RUL direct prediction method based on multi-layer LSTM network is studied.The state monitoring data is essentially time series data.Therefore,combined with the advantages of LSTM network processing time series data,this thesis constructs a sample based on time window sliding and designs a RUL direct prediction framework based on multi-layer LSTM network.Finally,the experiments were carried out on two public datasets,PHM08 and C-MAPSS.The results show that the multi-layer LSTM network can achieve higher prediction accuracy than the traditional shallow machine learning model.3.Designed and developed a health management system based on the Azure cloud platform.Based on the main research algorithm of this thesis,the turbine engine equipment is used as the research object,and the turbine engine health management system based on Azure cloud is built.The core function module is verified and tested.The above research results show that,based on the PHM application scenario of the equipment under complicated conditions,combined with the deep learning theory,the thesis proposes a new method based on LSTM-VAE and a direct RUL prediction method of multi-layer LSTM network.To some extent,it solves the problem that traditional methods have too much domain-dependent expert experience,deep features are not available and non-universal.Comparative tests were conducted on public datasets such as PHM08,and better results were obtained.