In this article, a novel maximized mutual information based non-Gaussian subspace projection (MMI-NGSP) method is proposed for process monitoring and fault detection by searching for the low-dimensional subspace of measurement variables that retains the maximal statistical dependencies with quality variables. The basic idea of MMI-NGSP approach is to optimize the latent directions corresponding to the process measurement and quality variables respectively so that the maximized mutual information between the latent scores of measurement and quality variables is obtained. In our study, the gradient descent algorithm is developed to estimated the latent directions numerically. Further, both the geometric properties and fault detectability of the proposed MMI-NGSP method are investigated. The computational results of a simulation example demonstrate the validity of the proposed approach.