<p>The accurate detection of stress states in bridge steel welds is of critical significance for ensuring structural safety and durability. In this study, metal magnetic memory testing (MMMT) technology is innovatively introduced to detect and evaluate stress states in steel welds. By conducting magnetic detection experiments on two categories of welds (comprising eight specimens) under three-point bending, the effects of variables such as detection paths and specimen differences (structure type and initial magnetic state) on the distribution of the magnetic field were analyzed, and the evolutionary patterns of magnetic field under varying stress conditions were elucidated. Drawing upon the shared characteristics observed in the magnetic field variations of both weld categories, the relative entropy and the gradient Euclidean norm of magnetic memory signals were extracted as key magnetic characteristic parameters in the study. The results show that relative entropy changes gradually during the elastic stage but increases sharply upon transition into plastic deformation, with its rate of rise markedly exceeding that in the elastic phase. This finding indicates that relative entropy may serve as a reliable indicator for accurately identifying the stress states of the structure. Furthermore, there is a high linear correlation between the Euclidean norm of the magnetic signal gradient and elastic stress. By employing optimized feature parameters, structural stress can be effectively quantified, with most measurement errors maintained within 20%. These findings provide a viable technical approach and methodological reference for the stress detection of bridge steel welds.</p>