To overcome the limitations of manual visual inspection for welding defects in spent fuel storage racks—including inefficiency and high missed-detection rates—this paper proposes YOLOv11-FSDA, an enhanced target detection algorithm based on YOLOv11. Specifically designed to address challenges such as blurred object boundaries in welding images and performance degradation under low-light conditions with severe noise interference, this paper introduces a Frequency Spectrum Dynamic Aggregation (FSDA) module. This module employs a dual-domain approach: (1) in the frequency domain, it dynamically filters multi-channel amplitude/phase spectra through Fourier-based decomposition and learnable channel-wise weighting to suppress high-frequency noise while preserving structural information; (2) in the spatial domain, residual connections maintain critical details. By jointly optimizing frequency-specific noise attenuation and spatial coherence, FSDA effectively handles inconsistent noise patterns. Experiments on an augmented welding defect dataset demonstrate significant improvements over the baseline YOLOv11, with YOLOv11-FSDA achieving a 4.8% increase in mAP@50 and a 5.7% boost in recall rate (R). These advancements substantially enhance detection accuracy and reduce missed-detection probabilities, validating the algorithm's robustness for real-world industrial applications.