Brake pads on high-speed trains running through mountainous regions of central and western China face a demanding task. The railway lines there are shaped by substantial terrain variations, producing numerous long ramps where trains depend heavily on friction between brake pads and discs to hold a stable speed and stop safely. Because the pads wear continuously under these conditions, monitoring their damage is fundamental to safe operation, and researchers have now introduced a new artificial intelligence approach designed to make that monitoring work reliably across a range of train speeds.
The challenge, as described in a study published in Applied Intelligence, is that variable-speed braking on long ramps causes significant variations in the characteristics of signals recorded at the brake interface. A monitoring system trained on data from one rotational speed may fail when the train brakes at another. The authors therefore frame cross-speed signal monitoring as a recognition problem involving multiple target domains, meaning the model must adapt simultaneously to data from several speeds rather than just one. Their solution is a reinforced multi-target domain adaptation network, abbreviated RMTDAN, which was developed specifically for cross-speed brake pad condition monitoring.
The network combines three components that the authors describe as working synergistically. The first, random mix-up augmentation, generates diverse samples with altered distributions, which is intended to enhance the model’s extraction of categorical features when data distributions change and to ease adaptation to multiple target domains. The second, centroid similarity pseudo-labeling, produces reliable labels for unlabeled data across the multiple target domains, improving model precision during the alignment process. The third, prototype contrastive learning, constructs category prototypes for the source domain and for each target domain. This contrastive mechanism reduces the distributional distances between representative features of the same category in different domains, while maximizing similarity between inter-domain prototypes to reinforce domain alignment.
Together, these components progressively align data distributions recorded at different rotational speeds. To evaluate the approach, the team carried out multi-target domain adaptation experiments using vibration acceleration data collected in the laboratory at different rotational speeds. According to the reported results, the proposed method enhances domain adaptation capability by learning data distributions at multiple speeds, thereby enabling effective cross-speed brake pad condition monitoring for high-speed trains operating on long ramps. The authors also report that the method achieves higher recognition accuracy than other domain adaptation algorithms they compared it against.
The study, authored by Min Zhang, Lingzhi Qiao, Daqian Ji, Caoyuan Hu and colleagues, appears in Volume 56 of Applied Intelligence as article number 528, with a publication date of 29 October 2026. Its findings are based on laboratory-collected vibration data rather than measurements from trains in commercial service, so the reported accuracy gains reflect performance under controlled experimental conditions at the tested rotational speeds. Within that scope, the work addresses a practical safety concern: on long ramps, where braking is sustained and speed regulation is critical, a monitoring system that can recognize brake pad damage across changing speeds could help ensure that wear is detected regardless of the operating conditions a train encounters on mountainous routes.
Subject of Research: Cross-speed condition monitoring of high-speed train brake pads using a multi-target domain adaptation network
Article Title: Cross-speed condition monitoring of brake pads for high-speed trains on long ramps based on a reinforced multi-target domain adaptation network
Article References: Zhang, M., Qiao, L., Ji, D., & Hu, C. (2026). Cross-speed condition monitoring of brake pads for high-speed trains on long ramps based on a reinforced multi-target domain adaptation network. Applied Intelligence, 56(16), Article 528. https://doi.org/10.1007/s10489-026-07483-9
Image Credits: AI Generated
DOI: 10.1007/s10489-026-07483-9
Keywords: brake pads, high-speed trains, domain adaptation, condition monitoring, machine learning, vibration signals, railway safety, long ramps, pseudo-labeling, contrastive learning, Cross-speed, condition
News Source: Denise Maddox. (October 29, 2026). New AI Network Monitors High-Speed Train Brake Pads Across Different Speeds. Scienmag.
