Gap Monitoring Of Switch Machines Based On Lightweight Object Detection Algorithm

The position of the switch rail and the degree of track tightness on the turnout can be reflected by the gap size inside the switch machine.Gap monitoring of the switch machine plays a significant role in the safe and stable operation of the train.Gap monitoring mainly consists of video surveillance and gap detection.Embedded video surveillance has the advantages of stable operation and low cost,which is widely used in this field.At present,gap detection methods mostly utilize traditional image processing algorithms.Still,their detection accuracy is easily affected by complex environmental factors such as vibration,and different image processing parameters need to be set according to the type of switch machine and the installation position of the camera.The above factors lead to its poor robustness and applicability.Deep learning algorithm can learn from a large number of data to achieve end-to-end detection.The adaptability of the model is strong,and its application to gap detection can make up for the shortcomings of traditional image processing methods.As the deep learning algorithm requires high hardware computing power,the lightweight processing of switch machine gap detection model based on deep learning has important engineering application significance to improve the detection accuracy,speed and applicability of the model.The research of the gap monitoring system in this paper is mainly divided into two parts.The first part studies the realization of real-time gap monitoring on the embedded development board.The second part explores the gap detection model based on lightweight object detection.The main work done in this paper is as follows:(1)This thesis constructs a real-time gap monitoring system on an embedded development board based on the i.MX6 q processor.According to the gap monitoring system requirements and hardware conditions,the functions of video capture,video coding and network transmission in the gap monitoring area are designed and implemented.Finally,the gap of the switch machine can be monitored in real time through the upper computer.(2)This thesis proposes a gap detection model based on deep learning,constructs a switch machine gap image data set,and trains and tests various object detection models on gap image data set.Through the evaluation and comparison of indicators such as gap detection accuracy and model calculation complexity,YOLOV4-tiny is selected as the basic network structure of the gap detection model.(3)A lightweight processing strategy for the gap detection network model is proposed.The computational complexity of the network model is reduced by sparse training,channel pruning and fine-tuning.The knowledge distillation is applied to the fine-tuning process,and the loss function in the knowledge distillation is optimized and improved according to the network structure of the model,which improves the detection accuracy of the lightweight gap detection model.In this thesis,the final lightweight gap detection model was tested and evaluated.The average accuracy of gap detection is 98.4%.The memory consumption and computing complexity of the model are reduced by 65.6% and 33.6% respectively,and detection speed is improved by 38.3%.The results indicate that the gap monitoring system based on the lightweight object detection model can meet the requirements of railway monitoring.The lightweight processing strategy proposed in this paper promotes the practical application of the object detection algorithm in the gap detection.