Track Fastener Image Processing Based On FPGA And Deep Learning

Track detection has always been a key research topic in the field of rail transit,and the state detection of track fasteners is an important part of the track detection task.As a key component connecting the rails and sleepers,the track fasteners are in good condition to ensure the safe and stable operation of the train.In this process,once the fasteners have various abnormal states,such as reverse installation,missing,broken,etc.,it will lay a great hidden danger for driving safety.Therefore,ensuring the health of track fasteners has received more and more attention from the railway department.At present,the ways to realize track fastener detection are mainly divided into two categories:(1)manual inspection;(2)track inspection vehicle detection.However,both methods have inherent disadvantages.For example,manual inspection requires long maintenance mileage,high maintenance costs,and high workload,which is prone to missed inspections,making it difficult to meet the needs of the rapid development of the railway system.For track inspection vehicle detection,the detection period is long,and the normal operating track line is occupied during the detection process,so the detection efficiency is low.In order to improve the existing track fastener detection methods,this dissertation proposes a track fastener detection method based on Field Programmable Gate Array(FPGA)and deep learning.This system is designed to use the embedded platform to achieve high-precision and real-time detection of track fastener images.In order to achieve the above goals,this dissertation takes the high-definition track fastener images as the research object to carry out the research on the state detection of track fasteners.In the work of fastener state detection,this dissertation proposes a track fastener detection method based on LYOLOv4-Tiny deep convolutional neural network,which ensures the accuracy of detection and limits the computational complexity of the model.In the FPGA hardware deployment work,this dissertation proposes a deep convolutional neural network model accelerator based on multiple acceleration strategies such as serial-to-parallel conversion,cyclic pipeline,array partition,ping-pong buffer,and multi-channel transmission.The main work of this thesis can be summarized as follows:(1)In the work of track fastener detection,this dissertation proposes the Promosaic data argumentation technology to make up for the shortage of defective fastener samples.In addition,this dissertation further improves the average detection accuracy through the Swish activation function,and then achieves the precise location of the fastener position through the improved loss function.Finally,considering that the detection model should be deployed in embedded devices,this dissertation uses the model compression method to compress the original network,and the effectiveness and good performance of the model in this dissertation are verified by comparing with different prediction models on multiple indicators.(2)In the work of FPGA hardware deployment,aiming at the high model complexity,large amount of calculation,slow detection speed and other problems,a FPGA-based convolutional neural network accelerator was designed and implemented,and the performance and required resources of the accelerator were analyzed.Using deep convolutional neural network to detect fasteners on the Zynq hardware platform,the detection accuracy exceeds 98%.Finally,a comparative experiment with various algorithms is carried out,and it is concluded that the method in this dissertation has certain advantages.(3)The track fastener detection software is designed,and each function is tested by using field data.The results show that the software meets the requirements of field application.Figure 44,Table 15,References 90...