An Unsupervised Framework for Mud Pumping Detection and Severity Analysis Using In-Service Train Data in Railway Track

Using machine learning techniques to analyze the monitoring data collected from in-service trains can help the infrastructure manager to detect and localize mud pumping defects automatically. However, most studies treat defect detection tasks in a supervised manner (e.g., classification), which rely heavily on manual data processing to label unhealthy conditions (such as mud pumping) for training. But a majority of measurement data actually represents a healthy condition. Supervised classification models require considerable efforts to balance the training dataset; otherwise, the models tend to perform poorly for the minority class. Unsupervised anomaly detection can handle extremely imbalanced dataset because no label information is needed in the unsupervised mode. This study proposes an unsupervised framework for mud pumping detection and severity analysis using in-service train data. The framework is based on a long short-term memory (LSTM) autoencoder and deep embedding clustering (DEC). The proposed framework is implemented on a three-year dataset collected from a section of railroads in Australia. The results show that the model can detect 5 out of 6 mud pumping events and identify their severity.

Keywords: Unsupervised learning, imbalanced dataset, autoencoder, deep embedding clustering, mud pumping.