Previous work on process monitoring has shown that chemical processes can be modeled by data-based models such as principal component analysis (PCA) and neural network models. However, it is difficult to train a model that has good generalization capabilities in, fault detection, especially for nonlinear processes. On the basis of the idea of making the trained model robust with respect to the noisy training data, this paper intends to develop a unified training method for PCA and the autoencoder model. A unified training model called the robust self-supervised model is first proposed. Then theoretical analysis shows that the models trained by the proposed methods are more sensitive to fault occurrence in the process. The corresponding statistics are introduced for the proposed methods. According to the simulation results for three case studies, the efficiencies of both robust autoencoder and robust PCA monitoring are evaluated.