Introduction: Curve squeal is a loud tonal noise from railway traffic that contributes to environmental noise pollution. Squealing arises from friction-induced vibrations in the wheel-rail contact during vehicle curving under specific friction conditions. Environmental factors are known to influence these friction conditions and, consequently, the tendency of squeal occurrence. This study investigates whether environmental variables measured from the wayside of the track can be used to predict curve squeal using machine learning models.

Methods: Separate prediction models were developed for low-rail and high-rail squeal using a labeled dataset collected during a 19-month measurement campaign at a commuter railway curve in Sweden. Multiple machine learning algorithms were evaluated, and feature importance analysis was used to investigate the environmental predictors associated with low-rail and high-rail squeal. The practical usefulness of the prediction models was further assessed for adaptive top-of-rail friction modifier (TOR-FM) application.

Results: Environmental variables contained useful predictive information regarding curve squeal occurrence, although the overall predictive performance remained limited. The tree-based ensemble models generally achieved the highest performance, while the feature importance analysis indicated that low-rail and high-rail squeal relied on different feature patterns. The operational assessment demonstrated that the prediction models could reduce unnecessary TOR-FM triggering while maintaining high squeal detection rates.

Discussion: The results demonstrate the potential of combining environmental monitoring with machine learning to support site-specific predictive railway noise mitigation and maintenance decision-making. The differences between low-rail and high-rail prediction further indicate the value of treating the two squeal types as separate prediction tasks.