There are over-fitting phenomenon and prone omissions when moving window adaptive Kernel Principal Component Analysis (KPCA) is utilized to deal with sensitive parameters or slow degradation problem. In order to solve the problem, a step dynamic auto-regression KPCA was proposed. Firstly, the initial model was established step by step drawing on dynamic data matrix. Then, the exponentially weighting rule was introduced to process real-time data and update the model based on the moving window adaptive KPCA. Finally, the algorithm complexity was analyzed and specific steps were given. The simulation data was utilized to analyze the impact of decomposition coefficient and weighting factor. The results show that, compared with the moving window adaptive KPCA, the proposed algorithm efficiency was improved by nearly 90% and the number of false positives was almost 0 in the case of appropriate parameter selection; and it could also control the adaptive ability to solve a variety of dynamic problems by adjusting the value of weighting factor. The algorithm was applied to the experimental data analysis of compressor surge and bearing fault, the result verified its ability to deal with the problem of sensitive parameter and slow degradation.