Research On TEDS Image Defect Visual Inspection System Driven By Data And Knowledge Symbiosis

Railway construction is an important category in the field of transportation in our country.With the large number of railway lines being opened and the gradual increase of railway operating mileage,the monitoring of the safety status of railway operation becomes more and more important.At present,our country has already carried out the laying of railway safety image intelligent analysis equipment.These systems mainly use artificial image discrimination methods,which are heavy in tasks,difficult to meet real-time detection requirements,and are prone to missing and false alarms.Therefore,the demand for the development of railway image monitoring and detection system based on artificial intelligence is becoming more and more urgent.This thesis is based on the Trouble of Moving EMU Detection System(TEDS)organized and constructed by the China National Railway Group Limited.The thesis aims to use a new type of algorithm driven by data and knowledge symbiosis to design a set of TEDS image intelligent auxiliary recognition systems.The designed system can detect and analyze the parts such as axle end bolts and brake pads of CR400 AF and CR400 BF models in real-time.The main innovative work and achievements of this thesis are as follows:(1)An ORB(Oriented FAST and Rotated BRIEF)feature extraction positioning model based on Gaussian probability field is proposed.This thesis uses ORB and k-means clustering to construct a feature dictionary and locates the target based on the character probability field.The method can solve the problem of poor system robustness caused by excessive reliance on artificial semantic features in traditional algorithms.Experiments are carried out on the parts data set of Electric Multiple Units(EMU).The results show that the designed algorithm has a high recognition rate of 99.6% and88.2% for the location of axle end bolts and brake pads,respectively.This innovative algorithm improves the accuracy of traditional algorithms for target locations.Besides,the recognition rate is faster.(2)A deep learning probabilistic positioning method of Attention-YOLO(You Only Look Once with Attention Mechanism)is proposed,which solves the problem of assigning attention to salient areas of the image.The channel attention and spatial attention mechanisms are added to improve the network structure of YOLOv3(You Only Look Once Version 3)and improve the positioning performance of the network.The positioning experiment analysis on the labeled data set shows that the improved model has an average detection accuracy of 98.7% and 91.7% for shaft end bolts and brake pads,respectively,which is 2.9% and 4.5% higher than the positioning accuracy of the YOLOv3 algorithm.(3)A target detection algorithm driven by data and knowledge symbiosis is proposed.The two positioning models designed are weighted and fused through the idea of probability to obtain more accurate target positioning;Deep Convolutional Neural Networks(DCNN)is designed for target defect detection.Experimental analysis on the defect data set shows that the designed algorithm model has a defect-recognition accuracy of 96.4% and 72.3% for shaft end bolts and brake pads,respectively,which is3.2% and 3.6% higher than that of the YOLOv3 algorithm.(4)The TEDS image intelligent auxiliary recognition system is designed for engineering realization.The system is based on the Ubuntu operating system,Py Torch framework,and Anaconda3 software.It uses GPU to improve image processing capabilities.The interface platform is used to test the system,which verifies the effectiveness of the system’s safety status monitoring and detection functions of EMU components.