A novel data-driven approach for process fault detection, called Fisher Discriminant Global-Local Preserving Projection (FDGLPP), is introduced. This method combines Fisher Discriminant Analysis with Global-Local Preserving Projection to determine the optimal projection direction, then transforms the original data using the projection matrix. This approach preserves the dataset’s global manifold structure, local neighborhood relations, and the discriminative power of the projection subspace, thereby enhancing fault identification accuracy. Additionally, to mitigate the impact of noise and measurement errors, Kernel Density Estimation (KDE) is used to calculate the control limits for monitoring indices, which reduces false alarms and improves detection efficiency, enhancing monitoring effectiveness. Numerical simulations validate the dimensionality reduction advantages of the FDGLPP algorithm, and a case study on the Tennessee Eastman process demonstrates that this method significantly outperforms GLPP, LPP, and PCA in fault detection and identification. Finally, t-SNE is used to visualize FDGLPP’s discriminative capability.