Abstract: To address the challenges in the estimation of the state of health (SOH) for lithium-ion batteries, including strong nonlinearity, complex degradation mechanisms, and the insufficient generalization capability of traditional single-kernel Gaussian Process Regression (GPR) models, this paper proposes an SOH estimation method based on a multi-kernel GPR ensemble learning model optimized by the Catch Fish Optimization Algorithm (CFOA). First, six health indicators are extracted from the voltage–time curve and the incremental capacity curve during battery charging to characterize the aging characteristics of the batteries. Subsequently, a multi-base-learner GPR model consisting of linear, Gaussian, Matern, and rational quadratic kernels is constructed, and linear regression is employed as a meta-learner to achieve multi-kernel ensemble learning, thereby enhancing the model’s capability to jointly capture the global degradation trend and local fluctuation characteristics of the batteries. Meanwhile, CFOA is utilized to optimize the hyperparameters, improving the estimation accuracy of the model. Finally, the proposed method is validated on four lithium-ion batteries and compared with different single-kernel GPR models as well as methods such as SVR, RF, and XGBoost. The results demonstrate that the proposed method achieves optimal estimation performance across all four batteries, exhibiting excellent stability and generalization capability.