Fault detection and diagnosis methods have to deal with large variable data sets encountered in complex industrial systems. Solutions to this problem require multivariate statistics approaches often focused on the reduction of the space dimension. In this paper we propose a fault detection and estimation approach using Multivariate Kullback-Leibler Divergence (MKLD) to cope with the negative effects due dimension reduction while using Principal Component Analysis (PCA). The obtained results show its superiority on the usual PCA-KLD based approach. An analytical model of the MKLD is proposed and validated for low severity fault (incipient fault) detection and estimation in noisy environment operating conditions.