Research On Blade Damage Identification Method Based On Convolution Neural Network

Aero-engine is the power source of aircrafts,whose safety quality directly affects the safety and performance of aircrafts.Engine blade is one of the core parts of aircraft engine components.Due to launching a pilot time working in harsh environment,engine blade is vulnerable to harsh temperature and high pressure,the influence of complex stress environment,leading to higher failure rate.In order to ensure the aircraft airworthiness flight and improve accuracy and efficiency of blade damage detection,the convolutional neural network is applied to blade damage identification so as to improve the efficiency of blade fault identification and classification,as well as to realize automatic blade damage detection and classification.The main ideas are as follows:1.The characteristics of blade damage pictures were studied which had been collected through borescope inspection,so as to make and collect pictures of the blade damage model.The optimized bilateral filtering method was used to smooth the image and preserve fine details of damage pictures.The appropriate data augmentation method was selected to expand the training set which would be used as image data set for the subsequent network test;2.After comparing some existing typical convolutional neural networks’ structure and algorithm,ResNet was selected as the basic model structure for research,and the optimization effect of several optimization algorithms was selected and tested.The optimization algorithm was combined with CNN to improve the network learning accuracy and learning efficiency,and the optimization model was determined.3.The blade damage data set was made and the optimized model was verified.When the data set was small,twice transfer learning was carried out to learn the underlying characteristics of the network as soon as possible.Aiming at the blade edge curling fault with low detection accuracy,after a specific fault was determined,the image shooting angle that could be recognized by network and the influence of the light source on the shooting angle were studied.