This paper presents a hybrid approach for online fault detection in nonlinear processes. To solve the possible monitoring difficulties caused by nonlinear characteristics of industrial process data, two applications of the Kernel Method: Hypersphere Support Vector Machine (HSSVM) and Kernel Principal Component Analysis (KPCA) are used as fault detection methods. On top of that, to obtain the adaptive models for online monitoring and fault detection in unsteady-stage conditions, instead of the static ones established by traditional HSSVM and KPCA, multiple methods are adopted, including Recursive KPCA, Adaptive Control Limit (ACL) and Online Sequential Extreme Learning Machine (OS-ELM), all of which update the detection model in real time with dynamically adjusting. The T2 control limit of Recursive KPCA, the classification hyperspheres of HSSVM and the single hidden layer feedforward network (SLFN) trained with OS-ELM work collaboratively in monitoring the real time process data to detect the possible faults. The proposed approach was tested and validated via a set of experimental data collected from a bearing test rig. Experimental results show that this approach is adequate for fault detection while meets the needs of real time performance.