| Aero engine maintenance is an important work to ensure that the engine is kept in a continuous airworthiness state,so the maintenance ability of maintenance personnel is very important.At present,the aero-engine maintenance training mainly relies on the teaching mode of on-site teaching,which is not only inefficient,but also high cost.At the same time,because the engine internal structure is relatively complex,engine parts disassembling process each are not identical,the training practices require frequent query maintenance manual,and only through the scene teaching,students on the engine overall structure is very difficult to have a comprehensive cognition,caused the traditional aircraft engine maintenance training,can’t do for students to conduct a comprehensive training effectively.In order to solve the above problems,our research group proposed an aero engine maintenance training teaching method based on augmented reality.One of the key technologies of this teaching method is to combine the 3d aero engine component model in augmented reality equipment with the virtual and real engine components.The premise of the realization of this technology is to accurately identify and locate engine parts in augmented reality equipment.By referring to the research status of augmented reality and deep learning at home and abroad,this paper adopts the identification method of aero-engine parts based on HoloLens 2 and deep learning technology.In this paper,typical target detection algorithms are studied,from R-CNN series algorithm,YOLO series algorithm to SSD series algorithm.Finally,YOLOv5 model is selected,and according to the characteristics of aero-engine parts image and the requirement of high precision in HoloLens2 when virtual and real aero-engine parts are combined.The convolutional attention module is used to optimize YOLOv5,and the feature graph in THE Neck network of YOLOv5 model is reconstructed to increase the attention to the pixel where the engine parts are in the image and improve the detection accuracy of the model.In this paper,13 kinds of common engine component data are collected,and the engine component data set is constructed by means of data enhancement,data renaming and data annotation.Iterative training of 500 epochs was carried out on the model by using the built engine component data set,and the m AP value of the final model reached 96%,and the off-line detection speed was58.1FPS.Then,the robustness of the model was tested in three aspects of target object occlusion,image brightness change and image rotation,and the optimized YOLOV5-CBAM model was compared with YOLOv5 model,YOLOv4 model,SSD model and FAST-RCNN model in many aspects by using the control variable method.The excellent performance of YOLOV5-CBAM network model in aero-engine component detection is verified.Finally,an engine component identification system based on HoloLens is built.The system adopts C/S architecture and is divided into HoloLens client and deep learning server,and transmits information through UDP protocol.HoloLens client has completed the functions of information collection and display of engine component category,position and function information,etc.The function of the deep learning service end is to train the network model and update the engine component category,position and function information required by the system according to the received video.The results show that the system can accurately identify and locate real engine components. |
