Abstract: Aiming at the challenges posed by the dense overlap of electromagnetic signals and the development of digital radars, which lead to difficulties in radar target identification and a sharp increase in unknown targets, a two-stage fusion inference method for radar usage based on feature subset optimization is proposed. First, a framework for feature subset optimization is established through feature analysis to achieve adaptive generation of the radar data knowledge base. A two-stage fusion inference method is then designed, the concepts of similarity and grey relational analysis are introduced to construct a knowledge inference mechanism, which enables the classification of electronic signals in a high-dimensional feature space, and an evidence fusion mechanism is designed based on subjective Bayes theory to achieve precise inference by integrating information from multiple signal dimensions. Experimental results indicate that this method offers higher inference accuracy for common radar usage and maintains robustness with a sufficient sample size.