Fault detection technique plays an important role in industrial systems. However, conventional model-based methods have unavoidable troubles in obtaining physical models of complex industrial processes. Data-driven fault detection approaches provide effective tools to solve them. Nevertheless, most of data-based approaches are focused on linear systems. It is still a challenging research direction for nonlinear processes. In this paper, a data-based fault detection scheme for non-linear systems is proposed. It provides a potential tool to fault detection of complex industrial systems in data-based manner. The presented algorithm uses the idea of model-based approaches for reference, employs the just-in-time learning (JITL) scheme to estimate the system output, and receives final fault detection results according to the decision rule of residual analysis. The algorithm owns inherent online adaptation and is easy to implementation. Three experiment examples, a numerical nonlinear system, a wastewater treatment system benchmark and a DTS200 three-tank system, are employed to prove the high accuracy, strong applicability and practical significance of the proposed method.