Research On Engine Life Prediction Method Based On Feature Fusion

With the rapid development of industrial internet,the complexity of modern industrial systems is continuously increasing,and the security and reliability of the system are increasingly valued.Prognostics and Health Management(PHM),as a very important technology,face increasing challenges in its important component remaining useful life(RUL)prediction.RUL prediction can provide early risk prediction by predicting the time length from the current time to the object failure time,which is helpful for early fault detection and avoiding accidents.Aircraft engine is a high-tech mechanical device regarded as the "heart" of an aircraft.Timely and accurate estimation of the RUL of aircraft engines can help formulate scientific maintenance plans and implement predictive maintenance.In this study,a sample analysis of the C-MAPSS dataset of aircraft engines was conducted as a preliminary step.Subsequently,various data preprocessing techniques were employed,including sensor feature selection,data standardization,sliding window reconstruction of data,and segmented linear degradation,in order to improve data quality and reduce noise.Then,the reconstructed data was processed by a multi-scale convolutional neural network for feature extraction to obtain hidden features of the original data.Afterwards,feature fusion methods such as hierarchical attention mechanism and concatenation fusion were used to fuse the differences in features extracted from different scales of convolutional layers,both within and between convolutions.Finally,the fused features were inputted into long short-term memory networks and gated recurrent unit networks according to the time-series characteristics,enabling multi-dimensional feature fusion and prediction of remaining useful life.During the training process,techniques such as batch normalization,dropout,and dynamic learning rate were used to improve the model’s prediction performance and reduce the risk of overfitting.This study proposed a multi-dimensional feature fusion-based residual life prediction model(MDFF),which was evaluated and validated using the C-MAPSS dataset.Through experiments comparing the effectiveness of single-scale convolution and multi-scale convolution feature extraction methods,the multi-scale convolution was used as a spatial domain feature extraction method.The residual life prediction model based on multi-dimensional feature fusion was compared with long short-term memory network(LSTM),Gated recurrent unit(GRU),Convolutional neural network(CNN)and multi-scale convolution models,and it was found that the multi-dimensional feature fusion method can effectively improve the prediction performance of the residual life prediction model.Finally,the evaluation indicators of the multi-dimensional feature fusion-based RUL prediction model were compared with those in high-quality published works,showing that the proposed model outperformed most of them.This study validated the effectiveness and advanced nature of the proposed RUL prediction method based on multi-dimensional feature fusion.