A Nonlinear Gaussian Belief Network (NLGBN) based fault diagnosis technique is proposed for industrial processes. In this study, a three-layer NLGBN is constructed and trained to extract useful features from noisy process data. The nonlinear relationships between the process variables and the latent variables are modelled by a set of sigmoidal functions. To take into account the noisy nature of the data, model variances are also introduced to both the process variables and the latent variables. The three-layer NLGBN is first trained with normal process data using a variational Expectation and Maximization algorithm. During real-time monitoring, the online process data samples are used to update the posterior mean of the top-layer latent variable. The absolute gradient denoted as G-index to update the posterior mean is monitored for fault detection. A multivariate contribution plot is also generated based on the G-index for fault diagnosis. The NLGBN-based technique is verified using two case studies. The results demonstrate that the proposed technique outperforms the conventional nonlinear techniques such as KPCA, KICA, SPA, and Moving Window KPCA. (C) 2015 Elsevier Ltd. All rights reserved.