This article first adopts a fuzzy mathematical analysis model and proposes a mathematical model of fuzzy characteristics of music categories. Second, the article extracts audio data characteristics and conducts music classification experiments. The article establishes a DNN-Bottleneck feature framework, extracts robust perceptual features from audio data, and then encodes to achieve fast audio perceptual hashing. Experiments verify the effect of different feature selections on the semantic space representation of music emotions and the impact on retrieval performance.