<p>In recent years, significant progress has been made in the field of freight train fault detection. However, constructing high-quality training datasets remains challenging due to the complexity and diversity of train fault types. Existing crowdsourcing platforms fail to ensure the accuracy of data annotations, and inaccurate bounding box annotations are common in training datasets. This inevitably leads to a significant decline in the performance of object detectors.To address this issue, we have developed a hierarchical guided interconnected fault detector, which significantly improves the robustness and detection accuracy of models in noisy label environments. Specifically, we utilize a deep interconnected feature pyramid network that progressively builds features through a multi-level feature enhancement mechanism. At the same time, a cross-layer direct feature bridging mechanism is employed to achieve effective complementarity between low-level and high-level features, reducing the negative impact of noise on the feature extraction process.In addition, we propose a Hierarchical Dynamic Kullback–Leibler Divergence Guided Assignment mechanism, using Dynamic Kullback–Leibler Divergence as a distance metric between candidate and ground truth boxes. Through a dynamic threshold adjustment strategy and a two-step hierarchical assignment strategy, the selection of candidate boxes is optimized. This mechanism enables the model to adapt more flexibly to sample allocation in noisy environments, reducing the interference of noisy annotations during training.We conducted extensive experiments on a freight train fault dataset with noisy bounding box annotations and achieved optimal performance. The code is available at <a href="https://github.com/panxingyu666/-HDKLGA-">https://github.com/panxingyu666/-HDKLGA-</a></p>