Research On The Remaining Useful Life Prediction Method For Aero Engines Based On Improved Convolutional Neural Network

With the rapid development of science and technology,the complexity of modern industrial systems is increasing,and there are higher requirements for reliability.Prognostics and health management technology plays an essential part in ensuring stable equipment operation and reducing maintenance costs,and remaining life prediction is the key to this technology.Data-driven remaining life prediction methods can obtain equipment degradation failure characteristics from equipment sensor data without the need to build accurate physical or mathematical models for complex components.Data-driven deep learning prediction methods are developing rapidly and have effectively improved the accuracy of remaining life prediction.However,the sensor data collected in actual industrial production is characterized by high-dimensional and complex working scenarios,which brings problems such as inadequate feature utilization and less accurate prediction.Therefore,it is a major challenge to establish an accurate remaining life prediction model according to different working scenarios.To this end,the main work of this thesis is as follows.(1)In order to focus on key features and secondary features in sensor data at the same time,a network with complementary main and secondary features for remaining life prediction is proposed.Specifically,a feature complementary network is designed to extract the primary features of equipment degradation with complementary features for predicting the remaining life of the equipment.The main network is based on principal component analysisconvolutional neural network to extract the main degradation features from sensor data.A probabilistic principal component analysis-gated recurrent unit is used in the feature complementary network to supplement the hidden degradation information contained in the non-principal components discarded during the principal component analysis,thereby improving the accuracy of the lifetime prediction.Under the experimental scenario of single fault and single operating condition,turbofan engine is selected as the research object.According to the characteristics of aero-engine with high-dimensional data and complex noise,this paper completes invalid data rejection and noise data filtering by trend analysis of sensor sequence data to obtain effective features.The key features are extracted by using principal component analysis-convolutional neural network,and the secondary features are focused on using probabilistic principal component analysis-gated recurrent unit,and the prediction results of the two networks are fused to improve the remaining engine life prediction accuracy.(2)In order to better extract deep degradation features from sensor data,a residual life prediction network with a residual parallel pooling fusion attention mechanism is proposed.Specifically,a parallel pooling fusion channel attention module is designed to extract key degradation features and retain global features in sensor data by integrating the outputs of average pooling and maximum pooling.In order to make the network focus on the critical correlation information within the sequence during training,a spatial attention module is introduced.And the residual network-based remaining life prediction network is designed according to the experimental scenarios,which can extract deep degradation features and obtain accurate life prediction models in different experimental scenarios.(3)In order to further fully utilize the sensor data,a feature fusion multi-level attentional remaining life prediction network is proposed.Specifically,a feature fusion network is first designed to extract degradation features at different scales using different sensory fields,and the extracted features are fused to obtain multi-scale degradation features;then a multi-level spatial attention module is designed and combined with channel attention to extract the intrasequence correlation information of key features;as the feature fusion network and attention module extract more effective fused features,the network further improves the accuracy of life prediction due to the more effective fusion features extracted by the feature fusion network and the attention module;meanwhile,a learning rate decay strategy is introduced to accelerate the model training speed.The three designed networks are validated under different fault and working condition scenarios,and the results show that the designed networks have good effectiveness.