In this paper, a novel probabilistic kernel partial least squares (PKPLS) method and the corresponding process-monitoring method are proposed. The contributions are as follows: (1) PKPLS defines an appropriate probability model from a probabilistic viewpoint. The probabilistic characteristics and the relationship between input variables and output variables are analyzed using Gaussian latent variables. (2) PKPLS solves the nonlinear issue in process monitoring and addresses the problem of missing values, especially in big data processing. (3) Based on the data distribution structure, the monitoring model is established using PKPLS, which overcomes the limitation of the lack of a relevant probability density or generation model in traditional methods. (4) We qualitatively analyze the problem of parameter determination. A Bayesian rule and expectation-maximization algorithm are developed to estimate the parameters of the probabilistic model. The PKPLS algorithm is applied to process monitoring. Numerical examples and the Tennessee Eastman process are used to evaluate the performance of the PKPLS method. Extensive experimental results verify the satisfactory performance of the proposed new approach.