Train motion prognostics and classification from multi-source decentralised sensors using unsupervised data-driven technology

• A novel automated train motion prognostics and classifications is established. • Hunting phenomena and associated rolling radius differentials from multi-source sensors are identified. • Over hundreds of real-life data have been collected from multi source decentralized sensors. • New unsupervised deep learning has been robustly developed, customized, and verified by field measurements. • Railway risk and safety can be improved from monitoring rolling radius differentials. This study proposes a novel scenario to recognize railway train motions using K-means clustering and data collected by smartphone sensors. Traditional methods often depend on high-cost and maintenance intensive sensors. These methods face financial and logistical challenges which limit their widespread application. This study collects train acceleration using smartphones onboard and uses K-means to classify different train motions from the extracted features in both time- and frequency- domains. The result demonstrates that this approach not only addresses the latency with the traditional methods but also enhances the accuracy of train motion classification. This successful endeavor underscores the potential of integrating machine learning with smartphones to efficiently address railway motion classification challenges which enhances real-time monitoring and predictive maintenance.