Research On Prediction Of Remaining Useful Life Of Aero-engine Based On Data Driven

As one of the most important core components of an aircraft,the safety and reliability of an aircraft engine is of great significance to ensure the safety of the aircraft and personal safety.In the process of actual operation and use,the performance and health status of aircraft engines will inevitably degrade gradually under the combined effect of various internal and external factors,and when the degradation reaches a certain level,it will not be able to complete its normal tasks and functions,which will cause irreversible economic and personnel losses.Remaining useful life(RUL)prediction is one of the most critical parts of the failure prediction and health management(PHM)system,and it is also the most difficult and challenging comprehensive technology.If the RUL of an aero-engine can be accurately predicted,appropriate maintenance plans can be made in advance to reduce maintenance costs and effectively protect aircraft and personnel safety.With the rapid development of modern industrial sensor technology and artificial intelligence technology,industrial Internet of things system can obtain massive equipment sensor monitoring data,and these big data have the characteristics of massive,high and non-linear,which are difficult to be processed by traditional methods,so the data-driven RUL technology based on data has been widely researched and applied.Deep learning methods with powerful data processing capability,without the need to know the exact failure mechanism of the device and professional signal processing knowledge,can deeply mine the hidden information in sensor monitoring data to achieve more accurate RUL prediction,and have already achieved remarkable results in the field of RUL prediction.Based on the monitoring signals of aero-engine sensors,the data driven RUL prediction method is studied deeply in this paper.The main research contents of this paper include:(1)To solve the problem that most deep learning methods do not carry out adaptive weighting for different sensor input features,a RUL prediction model(FARes Net)based on feature attention mechanism and Residual Network(Res Net)is proposed.Firstly,a feature attention mechanism module with compression excitation is used for adaptive weighting processing of multi-sensor input features.Then,two parallel Res Net branches are used to extract time series features along the time dimension and feature fusion along the sensor feature dimension respectively to deeply mine the hidden features of the data.Experimental results verify the effectiveness of the proposed RUL prediction model,which has lower prediction error compared with other deep learning models.(2)In view of the problem that current RUL prediction methods based on deep learning(1D-CNN,LSTM,etc.)can effectively mine the time information of sensor sequence,but it is difficult to effectively model the spatial relationship between sensors.,a RUL prediction model(MSCNN-GAT)based on multi-scale convolution and Graph Attention Network(GAT)is proposed.Firstly,multiple different 1-D CNN kernels are used to mine the data features of sensor time series signals at multiple scales.Then,Pearson correlation coefficient was used to construct the adjacency matrix of the sensor network,and the graph structure of the sensor network was constructed,and then GAT was used to aggregate the multi-sensor data.Experimental results on CMAPSS data set show that the proposed RUL prediction model can effectively mine the time and space dependence of multi-sensor data,improve the performance of RUL prediction,and is suitable for RUL prediction tasks under various fault conditions.