Research On Deep Learning And Its Application On The Defects Detection For Aero Engine

Aero engine is the core of aircraft,its key components usually run in harsh environments with high temperature,high pressure and high speed.Long-term running in this environment will result in various defects on components of aero engine,such as cracks,nicks,indentations,dents,ablation,corrosions,blade curling and other defects.As the power heart of the aircraft,the important task of aviation maintenance is based on the internal conditions of the aero engine,in time to determine whether it meets the flight standards.At present,industrial endoscopy is a commonly used detection method for aero engine.Through the endoscope,inspectors can easily observe the depths of the engine,but this method is limited by human factors.Due to the complexity of the engine internal structure and the variety of defect category,shape,location and area,the traditional image processing method is difficult to solve the above problems.Convolution Neural Network(CNN)has the ability to automatically learn the feature of samples,which opens a door for the automatic detection of engine defects.From Regions with CNN(R-CNN)model to the improved Fast R-CNN model,and finally to the faster Faster R-CNN model,the detection accuracy and detection speed of above modes have been constantly improved.The Single Shot MultiBox Detector(SSD)for real-time detection is a state-of-art object detection algorithm.In this paper,we will focus on the above-mentioned object detection model based on convolution neural networks and its application on the defects detection for engine.Several network models are deeply studied in this paper.Faster R-CNN with small network,Faster R-CNN with large network and SSD model using large network are choose for training engine defect samples.The test data set is sent to the model for testing after training.We found that the detection accuracy of the three models is better than the data given by the original author.Especially,the detection accuracy of SSD network model using large network is 89.36%,and the average detection speed is about 29.8 frames per second.It means that the detection speed has reached the real-time video image processing requirements.This model can meet the industrial production real-time detection standards.This paper implement an engine defect recognition system by using the above three training models.This system provides a platform for rapid diagnosis of engine defects,doing tests and applications.Although the final detection accuracy of the three models can’t reach 100% accuracy,but the research in this paper provides a feasible solution for the automatic detection of engine defects.It laid the foundation for the final application.