Enhanced YOLOv11 for Image-Based Anomaly Detection in Freight Train Gate Chains

To address the low accuracy and poor generalization of existing brake chain anomaly detection algorithms for freight trains in complex scenarios, this paper proposes an improved You Only Look Once version 11-based real-time image detection framework. The simple attention module is integrated into the backbone network to enhance feature extraction, while the content-aware reassembly of features up-sampling operator improves texture reconstruction of brake chains, particularly for small targets like lock chains. A dynamic wise intersection over union loss function with adaptive non-monotonic focusing weights is designed to mitigate localization errors caused by complex backgrounds. Experiments on a self-built dataset (2,000 linear array images) show that the improved model achieved a 99.5% mean average precision on the test set, a 5.4% increase over baseline You Only Look Once version 11, with single-frame inference time of 12.9 ms. This method enables real-time, high-precision detection of brake chain tightness, offering a lightweight solution for intelligent safety monitoring in freight train operations.