Research On Wing Icing Shape Recognition Technology Based On Enhanced U-Net

Icing is one of the important factors affecting the flight safety of aircraft.It must be paid attention to from design and manufacture to operation,and the icing detection of key parts of the aircraft is also particularly important.At present,the icing signal is obtained by using the single point icing sensor.This method is difficult to give the overall icing state of the aircraft area.Visual detection often has some empirical judgment,which can only be used as an auxiliary detection device for aircraft icing.In this thesis,the wing icing shape recognition method based on U-Net is proposed to realize the accurate and specific recognition of wing icing,and complete the pixel level recognition of wing icing shape.Firstly,the simple icing tunnel experimental platform is designed and built,which solves the problems of difficult and high cost of wing icing image sample acquisition.The simulated wing icing environmental conditions include wind speed of 0～18.5m/s;The diameter of supercooled water drop is 26.3～32.5 m;Water droplet content:0.8～0.9g/m~3 and temperature-15～0℃.The wing icing image sample is obtained through the icing experiment.Sample enhancement is realized by image processing methods such as Gaussian noise and HSV transform,and the wing icing image sample set is constructed for the neural network training of wing icing shape recognition.CBAM is added to the skip connection channel of U-Net network model to solve the recognition problems of high transparency and fuzzy boundary of icing.The Att U-Net improves the recognition ability of icing boundary by processing the characteristic map of decoder process with channel attention and spatial attention.Through comparative experiments,it is concluded that the recognition recall of Att U-Net network model with CBAM in the image samples of wing icing test set is 93.48%,which is 8.69 percentage points higher than that of U-Net network model.Complex background and interference factors are common problems in wing icing image samples,which leads to the low recognition precision of U-Net network model.The Res-block is introduced into the RU-Net network model to prevent the gradient from disappearing in the training process,further excavate the characteristics of icing area and distinguish background interference,so as to improve the recognition precision of U-Net network model.The comparative experimental results show that the recognition precision of and RU-Net network model for image samples of wing icing test set is 91.49%,which is much higher than the recognition effect of U-Net of 78.32%.The anti-interference ability of complex background and the accurate recognition ability of icing fuzzy boundary are improved by integrating CBAM and Res-block into the U-Net network model.At the same time,the morphological analysis method is used to process the output of model,and the enhanced U-Net network model for wing icing shape recognition is built.The performance comparison experiments of U-Net,Att U-Net,RU-Net and enhanced U-Net are carried out on the same software and hardware platform.The results compared with the U-Net network model,the recall and precision are increased by 11.98 and 17.00 percentage points respectively,which fully verifies the improvement of the network model performance by the improved module.The wing icing image samples with blue sky,gray sky background and different wing surface logos are used to the enhanced U-Net network model for recognition test,and the precision is more than 90%,which verifies the generalization ability of the network model.Finally,the enhanced U-Net network model was used to identify 5min,10min,and 20min image samples during wing icing.The results show that at the beginning of wing icing,the wing leading edge area will quickly collide with supercooled water droplets and form thin ice throughout the area.Over time,the icing area of wing surface changes less.The number of pixels occupied by the icing area in the identification results of the enhanced U-Net network model is counted,and the results reflect the change of the area of the icing area during the icy process of the wing,which fully verifies the perception ability of the enhanced U-Net network model to the change of the icing area of wing.