Research On Remaining Useful Life Of Turbofan Engine Based On TrellisNet

As an important part of military and civil aircraft,the health of Turbofan engine directly affects the safety of aircraft.In the past,the relevant maintenance personnel adopt the method of regular maintenance,although it can reduce the probability of engine failure to some extent,it is not the best solution because of the high maintenance cost and human resources.Condition Based Maintenance(CBM)has become a mainstream solution in recent years.Due to the big data era and the rapid development of sensor technology,technicians can obtain more real engine operation data.Compared with regular maintenance in the past,condition-based maintenance can determine the health of the Turbofan engine more quickly and accurately.The prediction of the Remaining Useful Life(RUL)of the Turbofan engine can be given by studying the past time series data of the engine.Therefore,the research on Turbofan engine RUL prediction has become the focus and hot spot of researchers in the near future.In view of the fact that new deep learning models are rarely used for RUL prediction in the current field,the prediction model in this field has become a bottleneck,and it also limits the further improvement of prediction accuracy.Since the engine operating data used is time series data,we looked forward to finding excellent deep learning models for processing time series data in the field of Natural Language Processing(NLP)to solve the above problems.This paper focused on the task of predicting the RUL of Turbofan engine based on TrellisNet,including data preprocessing,data feature extraction,modeling of predicting the RUL,etc.The main research contents include data feature extraction of autoencoder based on LSTM,the RUL prediction model based on Temporal convolution network(TCN)and the RUL prediction model based on TrellisNet.First of all,TCN,which has excellent performance in the field of NLP,is introduced into the field of RUL prediction.The dilation convolution in TCN makes it have a larger receptive field.Under the same network depth,compared to Long Short-Term Memory network(LSTM),it can Process longer sequence data.And its residual structure facilitates the training of the model.Based on the advantage of extracting data features from LSTM self-encoder,a RUL prediction model based on LSTM self-encoder and TCN was proposed.Experiments were carried out on the NASA dataset C-MAPSS.Compared with the prediction models such as CNN and LSTM,the proposed method is significantly improved on the complex dataset.Next,we focused on TrellisNet,which is a special structure of TCN.Compared with TCN,Firstly,the TrellisNet network adopts a weight sharing method,that is,the Filter used on each layer of the network is the same,which reduces the number of network parameters.Finally,the TrellisNet network combines the original input with the result of the hidden layer in each layer,so that the model can make full use of the information of the original data.TrellisNet performed well on datasets such as MNIST,PTB,PMNIST,and CIFAR-10.Therefore,this paper proposed a RUL prediction model based on TrellisNet.In the experimental part,compared with DBN,CNN,LSTM and other models,TrellisNet has a certain improvement in complex datasets.