Data-driven multivariate statistical analysis methods have been widely used in fault monitoring of large-scale and complex industrial processes. The condition Gaussian network (CGN) provides a way of probabilistic reasoning for continuous process variables, which has gained increasing attention. In this paper, a backward exponential filter is introduced into the discrimination rule and a CGN based on an adaptive threshold scheme is developed, which can effectively avoid process variables being misclassified because of small fluctuations caused by noise or disturbances. The purpose is to enhance the performance of the CGN method for process monitoring while maintaining a low misclassification rate and false negative rate. The performance of the proposed method is evaluated at the Tennessee Eastman Process and Intelligent Process Control-Test Facility. The results show that the proposed method performs better than the existing CGN-based methods and three conventional classification methods.