To address the prevalent electrical contact failure of high-voltage direct current relays, which is attributed to repeated exposure to high-voltage and large-current impacts in applications such as energy storage and electric vehicles, this article proposes a method for evaluating the electrical contact state of relays based on physically interpretable acoustic features. Traditional monitoring methods relying on the direct current circuit method are difficult to achieve online detection. Meanwhile, existing acoustic and vibration signal monitoring methods yield features with limited interpretability, as they lack grounding in degradation mechanisms. To address this issue, a finite element method (FEM)-boundary element method (BEM) coupled vibro-acoustic simulation model is formulated to elucidate the effects of collision velocity, contact pressure, and contact surface morphology on acoustic signals. Based on this, highly discriminative acoustic signal feature sets strongly related to multiple physical quantities are extracted. Subsequently, dynamic-structural interaction features are established by integrating physical prior knowledge, explicitly representing the implicit coupling relationship as interpretable indicators, and achieving feature enhancement under the constraints of physical laws. Then, a global-local fusion method grounded in statistical correlation is proposed. At the global level, the contribution degree for each of the feature sets is comprehensively quantified based on four dimensions, namely information content, diversity, complementarity, and stability. At the local level, weights are dynamically adjusted to accommodate temporal heterogeneity, effectively capturing the feature differences in the degradation stages. Finally, an ensemble classifier is constructed using random forests, decision trees, and K-nearest neighbors to realize recognition of electrical contact states. The results show that the features proposed in this study exhibit a significantly stronger correlation with contact resistance compared to Mel frequency cepstrum coefficients, and the average accuracy of the proposed method reaches 90.65%. This method enhances the analysis capability of relay degradation process based on acoustic signals, strengthens its adaptability and interpretability of evaluation, and provides effective technical support for relay health management.