Application Of Deep Neural Network Based On YOLOv3 In Aircraft Surface Defect Recognition

Aircraft structure is subject to fatigue load and corrosive environment during its service life,so it will produce damage defects such as fatigue and corrosion.Generally manifested as the corrosion and aging of the aircraft skin surface.The severely corroded joints of the beam and skin joints caused structural deformation,desoldering and perforation.Corrosion active spots appeared on the aluminum alloy connection site and began to corrode under the film.A series of problems such as severe corrosion of screw connection parts.It will have a serious impact on flight safety,from aggravating aircraft component damage to flight quality,and from serious accidents such as the emergence of fuel tanks,landing gears,and torn skins during flight,which is safe for flight and ground personnel.Causes serious security threats.In order to solve the problem of aircraft surface defect recognition,a deep neural network prediction model based on YOLOv3 algorithm is developed for aircraft surface defect recognition.Collect the surface defect pictures of the aircraft through various channels.Take representative samples as the data set,and divide them into 5 kinds of defects: skin peeling,skin crack,thread corrosion,skin deformation,and skin tear.The built-in aircraft surface defect prediction model was trained and tested.The model is a convolutional neural network with a depth of 129 layers.After 4000 iterations,the loss function keeps converging at 0.2691.The recognition speed of the single image in the test environment is between 30-50 ms,and the recognition accuracy on the training set and test set is 90.8% and 76%,and the recall rate is 80% and 72%,respectively.And draw the following conclusions based on the experimental results.First,sample images with different resolutions have an impact on the recognition accuracy of aircraft surface defect prediction models,but the impact is mainly concentrated on the recognition accuracy on low-resolution images and the recognition accuracy on medium-high-resolution images.Smaller.Second,the aircraft surface defect prediction model has a good recognition effect for single or clustered defects with a relatively large distance from each other;the model has a general recognition effect for aircraft skin defects where the defects appear densely overlapping or have similar external characteristics;For aircraft skin defect recognition under night and strong light conditions,if the aircraft skin defect features are not significantly occluded,it has a better recognition effect.If the aircraft skin defect features are significantly occluded,the model recognition fails.Finally,through a comparison experiment with the model based on Faster-RCNN algorithm on the same data set,it is found that the model based on Faster-RCNN has higher confidence in the same target than the model based on YOLOv3;the rivet defect based on small targets The recognition accuracy rate in recognition is 86% higher than the YOLOv3 model 83%;the recognition accuracy rate of the remaining defect types is between 68%-74%;the recognition speed of a single picture in the test environment is 50ms-2s between.Therefore,in the prediction of target confidence and the recognition of small targets,the model based on Faster-RCNN is better than the model based on YOLOv3;but it is inferior to the model based on YOLOv3 in recognition accuracy and speed.In particular,in terms of recognition speed,the Faster-RCNN model has poor real-time performance.Under the same conditions,the prediction time of a single sample based on the Faster-RCNN model is more than twice that of the model based on YOLOv3.