Wayside Defect Detector Data Mining to Predict Potential WILD Train Stops

Abstract

Advanced wayside detector technologies can be used to monitor the condition of railcar components, notify railroads of probable failures to equipment and infrastructure in advance, and predict alert of imminent mechanical-caused service failures. Using some statistical data-mining techniques, historical railcar health records from multiple Wayside Defect Detector (WDD) systems can potentially provide the essentials to recognize the patterns and develop the reliable and innovative rules to predict the failures and reduce related risks on railroads. In this paper, data from Wheel Impact Load Detector (WILD) and Wheel Profile Detector (WPD) were analyzed through comparing historical measurements for failed and non-failed wheels on the same truck to predict train stops due to high impact wheels. An exploratory data analysis was performed to identify the most critical measurements from each detector by comparing the distributions of several measurements from failed wheels to the ones from non-failed wheels. A logistic regression approach was used to predict the probability of potential high impact wheel train stops. Initial results show a 90% efficiency to predict the failure within 30 days after the most recent WILD reading.