Prediction Of Remaining Useful Life Of Aero-engine Based On Multidimensional Long Sequence

The level of industrial development is the embodiment of a country’s comprehensive strength,and industry plays a leading and driving role in the development of the national economy.As the pearl of the industrial field,the aeroengine is the product and representative of cutting-edge technology in the industrial field.The aero-engine industry not only undertakes the important task of ensuring national security,but also is the backbone of promoting industrial development and improving the national economy.In recent years,with the rapid development of industrial modernization and intelligence,aero-engine systems also have become more complex.Due to the complex structure and the harsh and changeable working environment of the aero-engine system,how to grasp its health status in a timely and effective manner is very important to ensure the safety,reliability and economy of the engine operation.Remaining Useful Life(RUL)prediction is one of the core tasks of Condition Based Maintenance(CBM)or predictive maintenance.How accurately and effectively to get the remaining useful life from the aero-engine monitoring data has important research value for prognostics and health management(PHM)of aero-engine system.In this paper,deep learning methods and artificial intelligence technology are used to conduct research on the remaining useful life prediction task of aero-engines based on multidimensional long-sequence monitoring data of aero-engine.As the benchmark data,the Commercial Modular Aero-Propulsion System Simulation(CMAPSS)data sets released by NASA are used to verify the validity of the prediction methods proposed in this paper.The main research contents of this paper are as follows:(1)Aiming at the limitations of the single prediction model in the task of aeroengine RUL prediction,a prediction method based on improved single model is firstly proposed to improve the adaptivity of single model,and then a prediction method based on combined model is proposed to improve the poor accuracy and generalization of single prediction model.First,the adaptive improvement of the single prediction model is carried out,and the characteristics and performance of each single model in the prediction task are studied and analyzed through experiments one by one.Second,the combined prediction model is constructed based on the performance characteristics of each single model.Finally,the prediction performance of the combined models are analyzed through experiments,and the effects of each single model in the combined models on each other are analyzed.The experimental procedures of single and combined models based on Convolutional Neural Network(CNN)and Recurrent Neural Network(RNN)are carried out with C-MAPSS dataset as the base data for experimental study and validation analysis.The final experimental results show that the proposed improved single model-based prediction method is more adaptive in the aero-engine RUL prediction task;the combined model-based prediction method can achieve the effect of mutual enhancement among the single models,which finally makes the combined model based prediction method have better generalization and prediction accuracy.(2)In view of the high dimensionality and long time sequence of aero-engine monitoring data,it is difficult to comprehensively consider the global features,local features and deep abstract correlation information implied between the degraded features,a method is proposed for predicting the remaining useful life of aero-engine based on deep dilated convolutional model.The method can significantly increase the model’s receptive field by the deep learning model builds by deepening the network structure,which improves the model’s feature mining ability for multi-dimensional long sequence data and the performance of comprehensive analysis of complex feature information.The experimental results show that the method can significantly improve the prediction accuracy in the task of aero-engine remaining useful life prediction(3)The setting of hyperparameters in deep learning models has an important impact on the comprehensive performance of the models.As the deep learning model structure deepens,the number of hyperparameters and the calculations in the training will increase dramatically,resulting in reduced model accuracy,a surge in training time,and higher hardware requirements.Meanwhile,the existing prediction models still suffer from weak generalization ability when dealing with different prediction tasks.Aiming at the above problems,this paper improves the prediction models by incorporating the Hyperband algorithm in the proposed prediction methods.By dynamically setting hyperparameters in the model and autonomously allocating computer arithmetic power,the Hyperband algorithm improves the generalization ability of the model when handling different tasks.While it reduces the subjectivity of hyperparameter setting and the upper limit of hardware demand in the training process,shortening the model training time.Finally,it improves the adaptability and prediction accuracy of the models.