Dynamic graph meta-learning with multi-sensor spatial dependencies for cross-category small-sample fault diagnosis in ZDJ9-RTAs

Railway Turnout Actuators (RTAs) suffer from small-sample fault diagnosis challenges because their high reliability limits fault data availability. Existing RTA small-sample methods show two major limitations—homogeneous category dependency, which prevents adaptation to newly emerging fault types, and single-sensor bias, which neglects spatial dependencies among multi-sensor signals. Focusing on the widely deployed ZDJ9-RTA, this study proposes a dynamic graph meta-learning framework to overcome these issues. The framework integrates (1) a category-shift meta-learning strategy based on dynamic task construction for rapid cross-category adaptation, and (2) a Graph-based Multi-Sensor Information Fusion Network (GMSIFN) that automatically builds sensor topology graphs and fuses multi-sensor features through GRU-gated attention to capture spatial fault propagation paths. Extensive validation on two real-world multi-sensor ZDJ9-RTA datasets — from a railway test line and in-service field deployments — demonstrates consistent superiority across 12 cross-category tasks (covering 2 to 5 novel class scenarios), achieving an average accuracy of 84.3% (3-shot) and gains of 16.1% over optimal baselines, while the field dataset exceeds 97% average accuracy in representative tasks, with peak performance up to 99.95%. These results confirm the proposed framework’s effectiveness, thereby extending small-sample studies to dynamically evolving industrial maintenance environments. Code is available at: https://github.com/huang-yu-han/GMSIFN. © 2025 Elsevier Ltd.