An adaptive fault diagnosis approach using pipeline implementation for railway inspection

Abstract

Railway tracks must be periodically inspected. This study proposes a new approach for eliminating two major disadvantages experienced during rail inspection applications performed via computer vision. The first is the blurring effect on images, resulting from physical vibration during movement on the rail lines. This effect significantly reduces the high accuracy rate expected from anomaly inspection algorithms. The second disadvantage is the need to operate in real time. This study presents a new three-stage computer vision method approach that eliminates both disadvantages. First, a three-stage pipeline architecture is implemented and IMU-assisted blur detection is performed on images taken from the left and right rail lines. Next, a convolutional neural network is used for learning. In the third test stage, anomaly detection and classification training are conducted. By performing the implementation with parallel programming on graphic processing units, a highly accurate, cost-effective computer vision rail inspection, based on image processing and capable of operating in real time, is successfully carried out.