A large number of false positives and false negatives may emerge when using traditional kernel principal component analysis to make fault diagnosis to dynamic systems. To solve the problem, a step auto-regression Kernel Principal Component Analysis (KPCA) algorithm is proposed. The method establishes a fault diagnosis model based on step dynamic strategy. Based on sliding window mechanism and exponential weighting idea, it updates the diagnosis model by continuously adding weighted, real-time data. T2and SPE statistic are used to detect whether the system has faults or not. The method is applied to fault detection of diesel engine valve, and the results indicate that the algorithm can not only make full use of original data and real-time dynamic information to update the model automatically, but also reduce the calculation cost, detect abnormal working status earlier, and improve the accuracy of fault diagnosis. Furthermore, the method can make diagnosis result more accurate and reliable by increasing the fault sensitivity.