Deep Learning Informed Methods for Automating Analytics of Railway Infrastructure Health Conditions

鐵路基礎設施健康狀況自動分析的深度學習方法

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
  • Zijun ZHANG (Supervisor)
  • Kwok Leung TSUI (Co-supervisor)
Award date11 Aug 2022

Abstract

The increasing railway transportation loads and higher train speeds lead to growing interest from researchers to develop advanced railway inspection technologies. Analyzing the health conditions of railway infrastructure is especially critical to ensure the safety of rail transport, since severe rail surface defects or any missing components would induce extra vibrations in train operations, contributing to a boosted risk of derailment. However, the traditional practice relies on experienced maintenance crews walking along railway tracks to periodically check their health condition, which is time-consuming and can involve human errors. Thus, automated analytics of railway infrastructure health conditions is attached great significance to the massive construction of railway network.

The recent advancement of image processing techniques and deep learning principles have benefited many railway-related analytic tasks, directly enhancing the inspection accuracy and speed. Based on different image datasets, three issues have been studied and addressed efficiently in our research.

First of all, deficient or lapsed components can cause train derailments and other severe accidents. It is indispensable to perform regular detection on major components of railway track, which requires an accurate and fast solution to detect categories, locations, and shapes of various components. In this problem, two methods are presented with different focuses. The first one is a one-stage method based on feature concatenation and multi-scale output layers. With the pre-defined anchors, the major component detection is realized by predicting offsets of the anchor coordinates. The proposed method is capable of improving the detection performance but also requires a significant amount of computations. To achieve a better tradeoff between detection accuracy and computational complexity, we propose the second method as an attention-powered deep convolutional neural net-work. Compared with the first method, it is anchor-free and employs two types of positional embedding to strengthen the detection of major components, especially edge components, which offers a new speed-quality solution to enable faster and more accurate image-based rail component detection.

Furthermore, foreign objects around the railway tracks are often the causes of rail transport accidents, such as train damages and unscheduled stops. Thus, it is of great importance to inspect the presence of any foreign objects on the rail track site. In the issue of rail foreign object detection, a three-step deep generative framework is introduced to learn the feature representations of normal rails and identify abnormal rails based on well-designed anomaly scores. Unlike the existing methods relying on a large number of abnormal rails, our proposed framework overcomes the scarcity of abnormal data in practice, allowing the detection of foreign objects at both image and pixel levels without pre-defining their scopes.

The effectiveness of the proposed methods is validated thoroughly with both collected and synthesized rail images. Results of the computational studies and comparative experiments with benchmarking methods further validate their suitability and superiority on railway infrastructure health condition analytics.