A causal saliency enhancement and Mamba-based multi-scale feature fusion encoding framework for railway fastener fault diagnosis

Fault detection of railway fasteners is vital for ensuring train operation safety within intelligent rail maintenance systems. However, existing Transformer-based detectors often incur high computational overhead and exhibit limited adaptability to multi-scale variations and complex background interference. To address these limitations, we propose SaM-DETR, a novel object detection framework that deeply couples causal saliency enhancement with a Mamba-based Multi-scale State Space Module (MSSM). Specifically, we introduce a differentiable Top-K selection mechanism combined with a causal invariance loss to adaptively learn the optimal foreground–background ratio, effectively suppressing spurious correlations and enhancing salient feature extraction. Furthermore, we replace the vanilla transformer encoder with a linear-complexity Mamba-based MSSM encoder, enabling efficient long-range feature fusion while dramatically reducing giga floating-point operations (GFLOPs). Extensive experiments on two real-world railway fastener datasets with different resolutions demonstrate that SaM-DETR achieves a balanced trade-off among detection accuracy, computational efficiency, and inference speed, outperforming mainstream DEtection TRansformer (DETR)-based models. These results validate the robustness and lightweight advantages of its novel causal-saliency—driven multi-scale design in multi-resolution industrial scenarios.