A new method of double-level local kernel principal component analysis (DLKPCA) was proposed to overcome challenges that traditional kernel principal component analysis (KPCA) method has encountered in incipient fault detection of nonlinear process. In this method

local information was obtained from data in variable-wise and sample-wise viewpoints to improve fault detection performance. First

all process variables were diced into several local variable blocks according to similarity of mutual information correlation among different variables and kernel principal components. Then

residual function was constructed from score vector and eigenvalue to mine local sample-wise information. Finally

the Bayesian fusion strategy was used to integrate results of each block. Simulation results on Tennessee Eastman standard process show that the proposed method can effectively detect incipient faults and has better fault detection performance than traditional KPCA method.