Comparison of multivariate analysis methods with application to fault diagnosis for non-Gaussian process

Traditional fault diagnosis methods of multivariate analysis (MVA) usually require that sampling data of separated latent variables must be subject to normal distribution, which is usually difficult to meet the actual industrial processes. This paper firstly introduces a method of fault diagnosis based on Q statistic. It requires that sampling data must be subject to normal distribution. Then this paper introduces a method of fault diagnosis based on information incremental matrix (IIM), whose sampling data haven´t the limitation of normal distribution. The method is mainly composed of defining covariance matrix, calculating information incremental matrix, information incremental mean and dynamic threshold, and so on. Finally, this paper gives a example of numerical simulation and a example of Tennessee Eastman Process (TEP), to verify the detection performance of two fault diagnosis methods, i.e., Q statistic and IIM, in false and missed alarm. The results show that Q statistic method have poor detection performance in the case that sampling data are not subject to normal distribution, while the method of fault diagnosis based on IIM is better.