False alarm reduction in railway track quality inspections using machine learning

Track quality geometry measurements are crucial for the railways’ timely maintenance. Regular measurements prevent train delays, passenger discomfort and incidents. However, current fault diagnosis or parameter deviation relies on simple threshold comparison of multiple laser scanners, linear variable differential transformer (LVDT) and camera measurements. Data threshold exceedances enact maintenance actions automatically. However, issues such as measurement error, and sensor failure can result in false positives. Broad localisation resolution prevents trending/ inferencing by comparison with healthy data baseline at the same position over periodic inspections. False alarms can result in costly ineffective interventions, are hazardous and impact the network availability. This paper proposes a novel methodology based on convolutional neural network (CNN) technique for detecting and classifying track geometry fault severity automatically. The proposed methodology comprises an automatic flow of data for quality assessment whereby outliers, missing values and misalignment are detected, restored and where appropriate curated. Improved, “clean” datasets were then analysed using a pretrained CNN model. The method was compared with a suite of machine learning algorithms for diagnosis including k-nearest neighbour, support vector machines (SVM), and random forest (RF). The analysis results of a real track geometry dataset showed that track quality parameters including twist, cant, gauge, and alignment could be effectively diagnosed with an accuracy rate of 97.80% (CNN model). This result represents a remarkable improvement of 38% in comparison with the traditional threshold-based diagnosis. The benefits of this research are not only associated with maintenance intervention cost savings. It also helps prevent unnecessary train speed restrictions arising from misdiagnosis.