The data in the chemical process is often complex in structure, and the coupling and correlation between variables are strong. Hence, it is difficult for conventional monitoring models to learn the essential characteristics of process data, which greatly affects the performance of the model. Aiming at the above problems, this paper proposes an optimized multi-kernel dictionary learning method, which is suitable for the industrial processes with strong nonlinearity and complex data structure. Firstly, the synthetic kernel fused by multiple kernel functions is used to map the process data to a new high-dimensional feature space to mine the multivariate distribution characteristics of the data. Secondly, the grid search method is used to optimize the model parameters in the kernel function for better model performance. Finally, the effectiveness and practicability of the proposed method in fault detection are verified on TE benchmark platform.