A novel framework that combines canonical variate analysis (CVA) with Global-Local Preserving Projection (GLPP) is proposed for fault monitoring in dynamic processes, referred to as canonical GLPP analysis. This method constructs a Laplacian matrix using a Hankel matrix, effectively capturing temporal correlations and underlying patterns. To solve the problem, Cholesky decomposition is applied to the covariance and cross-covariance of the Laplacian matrix, transforming it into a generalized eigenvalue problem. The resulting canonical GLPP analysis identifies an optimal projection matrix, uncovering dynamic variations by integrating both self-correlations and cross-correlations, while simultaneously preserving global and local structures as dictated by manifold learning. The performance of this method in fault detection and classification is demonstrated using a case study on the Tennessee Eastman process. Case study shows that the canonical GLPP analysis provides a more reliable and accurate framework for fault detection in dynamic processes.