Abstract: In actual industrial processes, fault data usually exhibit strong nonlinear characteristics, and the types of nonlinear characteristics are diverse. Existing kernel strategy-based process monitoring methods usually employ a single kernel function to extract the nonlinear characteristics of a fault, but comprehensively describing nonlinear characteristics is difficult. Therefore, the classification effect of single kernel-function-based monitoring methods for different faults is limited. In addition, the kernel parameters in the common kernel method are usually determined by experience. Obtaining the optimal result of feature extraction is also difficult. In this study, an optimized multi-kernel local Fisher discriminant analysis (OMKLFDA) model is proposed to solve this problem. First, multiple kernel functions are integrated into a local Fisher discriminant analysis (LFDA) model through weight coefficients so that the multiple nonlinear characteristics of the fault can be extracted. Second, an improved particle swarm optimization algorithm is utilized to select the optimal kernel parameters and weight coefficients for the fault classification model. In this way, the model can adaptively match the nonlinear characteristics of different faults. Finally, simulation results on the Tennessee-Eastman process and real-world diesel working process verify that the proposed method has an enhanced fault classification performance.