Wooden Sleeper Decayed Detection for Rural Railway Prognostics Using Unsupervised Deeper FCDDs
share
It is critical for railway managers to maintain a high standard to ensure user safety during daily operations. Top-view or side-view cameras and GPS positioning system have enabled progress toward automating the periodic inspection of defective features and assessing the deteriorated status of the railway components. Frequently, collecting deteriorated status data constraints time consuming and repeated data acquisition, because the temporal occurrence is extremely imbalanced. Supervised learning approach requires thousands of paired dataset of defective raw images and annotated labels. However, one-class classification approach has a merit that fewer images enables us to optimize the parameters for training normal and anomalous feature. Simultaneously, the visual heat map explanation enables us to discriminate the localized damage feature. In this paper, we propose a prognostic discriminator pipeline to automate one-class damage classification towards defective railway components. We also sensitivity analyze toward the backbone and the receptive field based on convolutional neural networks (CNNs) using pretrained networks: baseline CNN27, VGG16, ResNet101, and Inception Networks. We also visualize the explanation of the defective railway feature using a transposed Gaussian upsampling. We demonstrate our application for railway inspection in an open-accessed dataset of defective railway components, and wooden sleeper deterioration in rural railway. The heatmap is so important that the hazard-marks could cause an operational delay, an urgent inspection, and unexpected accident to passenger impact in railway inspection. Furthermore, we mention its usability for prognostic monitoring and future works for railway components inspection in the predictive maintenance of railway systems.
READ FULL TEXT