Use of artificial intelligence for predictive maintenance of electric trains in South Africa

Abstract

Railway transportation plays a critical role in urban mobility and economic development, serving as a primary mode of passenger movement for millions of commuters across South Africa. Despite its strategic importance, the South African rail sector continue to face persistent operational challenges as result of ageing infrastructure, progressive rolling stock degradation, and maintenance regimes that remain largely reactive or time-based. These conventional maintenance regime approaches are not favourable when considering the demands of modern rail operations, as they fail to anticipate subsystem failures before they escalate into costly service disruptions. The inability to detect early fault conditions in critical electrical and mechanical systems has contributed to increased unscheduled downtime, high maintenance costs, and compromised passenger safety, which are challenges that have been witnessed within many railway operational environments. To address these challenges, this study investigated the application of Artificial Intelligence and Machine Learning to develop and implement a predictive maintenance model for electric trains (ET), using a fleet of passenger rolling stock classified as electric multiple units (EMU), operated by the Passenger Rail Agency South Africa as a case study. Operational data was collected from the fleet operator, and a quantitative, simulation-based methodology was followed to evaluate a machine learning predictive maintenance model. The operational data covering a period of two (2) years (2023 – 2024) was collected across three critical subsystems, the pantograph, traction motor, and auxiliary converter unit, monitoring parameters including line voltage, current, temperature, airflow, vibration, and arc frequency. Two supervised machine learning algorithms, Logistic Regression and Random Forest, were trained and evaluated using this dataset. The Random Forest model achieved superior predictive performance, with 93% accuracy, 0.91 precision, 0.88 recall, and an F1-score of 0.89. These results verify the model’s ability to effectively distinguish between normal and early fault modes with high accuracy of detection, which is crucial for safety of railway applications. Furthermore, an analysis in this study revealed that trends in the frequency of high voltage arcs, line voltage deviation and temperature rise in traction motors were discovered to occur prior to recorded fault events. This demonstrated that features derived from these data can be investigated efficiently as an information rich multi-parametric supply for proactive predictive diagnostics. This study has also demonstrated that AI-based Predictive Maintenance can materially improve rail operational efficiency through condition-based intervention, reducing unscheduled downtime. The research adds to the new field of intelligent railway asset management by simulating the integration of AI in a South African passenger rail environment. It provides a methodological framework to shift maintenance from a reactive or time-based approach to a proactive and data centric strategy. Future work is suggested to further develop this work by deploying different methodologies with more sophisticated deep learning architectures for more objects at the same time.