In order to strengthen the detection and analysis of vulcanized rubber and its auxiliaries with similar appearance and odor or similar characteristics, the support vector machine modeling method based on improved particle swarm optimization was introduced into the qualitative analysis of terahertz spectrum. The experimental results show that the accuracy of the improved algorithm is larger than 81.25% for different data sets. Compared with support vector machine algorithm optimized by traditional particle swarm optimization, the algorithm also improves the recognition time, and the time spent overall is less than 9.40s. The method can be stably and accurately classified for different data sets, and provides a new research idea for the detection and classification of rubber and its additives.