ARTIFICIAL INTELLIGENCE-AIDED PREDICTION OF BROKEN RAIL OCCURRENCE USING RAILROAD BIG DATA

摘要

Approximately 6,450 freight-train derailments occurred between 2000 and 2017 in the United States, causing $2.5 billion worth of infrastructure and rolling stock damage. A significant portion of freight-train derailments on mainlines in the United States are caused by broken rails. Sponsored by the Federal Railroad Administration (FRA), this study was conducted to develop an Artificial Intelligence-aided methodology for predicting the occurrence of broken rail by location and time, using the fast-growing big data in the railroad industry. A soft-tile-coding based neural network (STC-NN) is developed to predict the spatial-temporal probability of broken rail occurrence. The probabilities of broken rail occurrence are ranked to determine where and when broken rails will occur given the predicted total number of broken rails. Our broken rail prediction model accounts for network-level rail and track characteristics, historical maintenance activities, traffic and operation, track inspection record, as well as environmental factors (e.g., temperature). The implementation of the model can support railroads to identify high-risk "hot spots" on the network and, thus, prioritize their track inspection, maintenance, and resource allocation decisions.