Aero-engine Damage Detection Based On Object Detection And Semantic Segmentation

Aircraft engines are critical to the aircraft being able to be safe and stable.In the course of progressively increasing work,the internal components of the engine are subject to damage from different sources and therefore the engine need to be regularly inspected for inside damage.Most of the current inspection methods of major airlines are based on manual determination of the internal pictures obtained by borescope camera.However,this relies heavily on the personal ability and experience of the technician,which may explore potential safety hazards(such as mis-inspection or missed inspection).As artificial intelligence is widely adopted in various fields,deep learningbased target detection and semantic segmentation techniques are becoming more advanced and popular.Compared to traditional methods,they can achieve more superior performance in many real-world scenarios.In this paper,the target detection based on deep learning and semantic segmentation algorithm is proposed to monitor the damage of each component in real time on the aero-engine borescope images.In this paper,the borescope images acquired under real scenarios were summarized and the damage on different components were analyzed in a targeted manner.Three aero-engine damage datasets,vortexer,large bend and blades,were subsequently obtained by manual annotation.Then semantic segmentation branches were added to the single-layer target detection network YOLOX model based on the encoder-decoder network structure.Two model input methods including single-layer feature input and multi-layer feature input are designed according to the feature requirements of different datasets,making the damage detection model to perform the task of damage detection and damage segmentation simultaneously.Also,this paper introduces a self-supervised learning based on unsupervised learning,and proposes a multilayer comparison learning method in comparison with the learning method Mo Co,with the combination of the characteristics of the damage detection model.The model constructs positive and negative samples based on the three layers of features output from the FPN,expanding the variability between the features of different layers so that each layer can better characterize damage of different complexity.Furthermore,for the deployment work of damage detection models in practical scenarios,this paper designs and develops a real-time aero-engine damage diagnosis system.The system performs real-time acquisition of borescope images through an image acquisition card and accompanying SDK,and uses a classification network to obtain the component categories in the images,and finally visualizes and persistently stores the damage data obtained through component-specific damage detection modules.Finally,a full experimental analysis was conducted to select the most suitable segmentation head addition scheme for each dataset.It is verified that the proposed multi-level comparative learning method can improve the detection accuracy of each damage category and thus can be more effective than the Mo Co method.The pretrained damage detection model of the backbone network achieves higher detection accuracy on all three aero-engine damage datasets.In the comparison with the classical model,the proposed method achieves better detection accuracy in all evaluation metrics,and the advantages of the proposed method in model deployment are demonstrated by the comparison of GLOPs and FPS.