The marine lysozyme fermentation process is a highly nonlinear, multi-stage, strongly time-varying system, making it hard to ensure model stability and prediction accuracy in the global scope by a conventional single global soft sensor model. To effectively solve the above problem, this study innovatively proposed a soft sensor modeling method based on an improved seagull optimization algorithm (ISOA) combined with Gaussian process regression (GPR) weighted ensemble learning. First, the sample data set is divided into multiple local sample subsets by the improved density peak clustering algorithm (ADPC). Second, the Gaussian process regression model is optimally altered with an improved seagull optimization algorithm for the purpose of establishing the corresponding sub-prediction model. Finally, the prediction model's fusion strategy is ultimately determined depending on the degree of connection between the test samples and a subset of local pieces. Simulation results show that the proposed soft sensor model can predict the key biochemical parameters of the marine lysozyme fermentation process well with less prediction error through fewer training data, which can be extended to soft sensor modeling of general nonlinear systems.