Oxygen-rich side blow bath smelting (OSBS) technology was developed in 2000 based on Vanyukov process. The technology has been widely adopted in China in recent years. However, matte grade and the slag composition are difficult to detect efficiently and accurately in real time. Currently, most smelters rely on offline manual testing, causing significant delays in obtaining results, which hampers timely feedback on matte grade during smelting, hindering efficient control. This paper proposes a dynamic soft sensor framework for matte grade prediction. This framework involves sampling unequally spaced matte grade, applying a time-lagged correlation analysis-based method for variable selection and dynamic information exploration, expanding and serializing dynamic features, and utilizing Gaussian process regression whose hyperparameters are optimized by a particle swarm algorithm to predict matte grade. Real data validation confirms the effectiveness of this method, which has been successfully implemented in a copper smelter in northeast China.