Study In The Detection Of Defects Of Fasteners On High-speed Railway Catenary Support Devices

In the electrified railway industry,the pantograph-catenary system plays an important role in transmitting power from the traction network to vehicles.However,due to the vibration and excitation in long-term operation,sophisticated mechanical and electrical interactions exist between the pantograph and catenary,which inevitably causes a high defect rate of the pantograph-catenary system and strongly influences operation safety.For example,fasteners serving as the connection of the cantilevers on the catenary support devices may loosen,break or are even missing.Non-contact detection is widely adopted with the great advances in imaging technology.However,nowadays,railway personnel manually detect the defects by reading a large volume of data from captured HD images offline.With the massive construction of high-speed railways,personnel can easily get vision fatigue and correspondingly miss some defects,which is also not efficient.It is necessary to develop an automatic defect recognition method based on the catenary support device images.As deep convolutional neural network(DCNN)prevails in image processing,a DCNN-based detection system is proposed in thesis to analyze and process the catenary support device images and predict the defects of the fasteners on the cantilever joints.Firstly,since various components are installed on support devices and are captured by cameras in global views,cantilever joints should be extracted from the HD images.Thesis analyzed the advantages of several object detection frameworks,such as Faster R-CNN,SSD,YOLO and chose SSD framework to localize the joints on the catenary support device.Considering that the joints usually cover small area in global images,we adjusted SSD framework to contain more shallow layers as detection layers.A dataset comprising four cantilever joints’ labels and location is established.Based on the training data,a model is trained to precisely locate and extract the joints by adjusting the related parameters.To analyze the working states of multiple and tiny fasteners,thesis refers to face recognition to design a three-stage cascaded DCNN.After extracting the cantilever joints in Stage 1,an improved YOLO framework is used to extracted six types of fasteners from the four joints.In Stage 3,a DCNN classifer is built to classify the three states of fasteners.Since Stage 2 is much easier,a semi-supervised training is designed to decrease the cost of human labeling.Due to the limitation of defect samples,the numbers of three-state training data are balanced.Thesis uses the catenary support device images of Ju-Yue section as training data and that of Heng-Zhu section as testing data.Extensive experiments and comparisons of the defect detection of catenary support devices along the Wuhan-Guangzhou high-speed railway line indicate that the system can achieve a high detection rate with good adaptation and robustness in complex environments.