An Unsupervised Learning Framework for Track Quality Index and Safety

Abstract

Rail track geometry defect remains a primary cause of train accidents in the USA. With over $1 billion lost annually to geometry-related accidents, there is an urgent need to re-assess the analysis and treatment of geometry defects. The development of track quality index (TQI) takes a contracted view of track assessment by focusing only on quality without any safety consideration. Track geometry safety limits are set by Federal Railroad Administration based on raw track geometry data. Since different variations of track geometry parameters exist, there is a skepticism about the effectiveness of the current bipartite analytical approach of track quality and safety. These results into two maintenance regimes: regular and spot maintenance. This study aims to create a framework through which a hybrid index that combines both safety and quality can effectively eliminate costly spot maintenance practices. This index would be used to create a data-driven maintenance scheme that maximizes the time between two maintenance cycles and minimizes disruptions. This technical note describes the proposed framework for creating such an index using unsupervised machine learning.