This research addresses the prevalent challenge of low diagnostic efficiency in soft fault diagnosis within DC-DC circuits, which is exacerbated by high feature dimensionality and an abundance of redundant information. To tackle this issue, we introduce a novel feature selection method grounded in the SHAP (SHapley Additive exPlanations) algorithm. Initially, our method computes the SHAP values for each feature across 16 distinct fault categories, thereby quantifying their contributions to the model's predictive capabilities. Following this computation, a thorough evaluation of feature importance is conducted through rigorous analysis. Based on this analysis, redundant features are selectively eliminated, yielding an optimized feature subset tailored for re-modeling. The experimental outcomes demonstrate that our proposed method not only effectively reduces feature dimensionality but also significantly enhances the accuracy and efficiency of fault diagnosis in DC-DC circuits.