To address the issues of insufficient generalization ability in traditional single-index diagnosis methods under complex operating conditions, and the opacity of the decision-making process in conventional machine learning models (the “black box” problem) for transformer DC bias condition assessment, this paper proposes an accurate assessment model based on multi-dimensional feature fusion and Shapley additive explanations (SHAP) interpretability analysis. First, based on the analysis of the vibration response mechanism of transformers under different degrees of DC bias, a 14-dimensional feature system fusing time-domain statistics, frequency-domain harmonic components, and specific physical mechanism features (such as weighted odd-even harmonic energy ratio and spectral entropy) is extracted and constructed. Next, a recursive feature elimination (RFE) strategy based on SHAP importance is utilized to adaptively screen the feature set, effectively eliminating redundant features. On this basis, the extreme gradient boosting (XGBoost) algorithm is adopted to build the assessment model, and Bayesian optimization is employed to adaptively tune the model hyperparameters, achieving precise identification of DC bias severity levels. Experimental verification based on multiple measurement points indicates that the optimized model can adaptively select the optimal feature subset for different sensor locations, achieving an assessment accuracy of over 96.2% across all tested points. Finally, the SHAP framework is introduced to quantitatively analyze the model’s decision-making mechanism from both global and local perspectives. The interpretability results demonstrate a high consistency between the model’s decision basis and the physical laws of vibration transmission and DC bias. This method demonstrates excellent interpretability while ensuring high diagnostic accuracy, providing a transparent and reliable solution for the engineering application of data-driven models in the field of transformer condition assessment.