Abstract: Principal component analysis (PCA) is a kind of commonly used fault detection method, but because of the uncorrected feature extraction, there are higher rates of false and missed alarm by using it in fault diagnosis. Thus, this paper firstly introduces the method of fault diagnosis based on the information incremental matrix obtained by the global covariance matrix. It can effectively reduce the rate of false and missed alarm as compared to PCA. But when the number of samples increases, the calculated threshold value is more unrepresentative and a much larger amount of calculation is required, they influence the performance of this method. Then, in order to overcome these shortcoming of the above method, one new fault diagnosis method is proposed by the local information incremental matrix obtained by the covariance matrix of moving the window, which comprises partial samples. This new method is mainly composed of defining the local covariance matrix, calculating local information incremental matrix, local information incremental mean, local dynamic threshold, and detecting abnormity and diagnosing fault, and so on. Finally, through two examples of numerical simulation to verify the detection efficiency of three fault diagnosis methods, i.e., PCA method, the method of fault diagnosis based on the information incremental matrix obtained by the global covariance matrix, and the proposed method, in false and missed alarm. The results show that the new method possesses the best detection performance.