Fault Diagnosis Based On Multivariate Statistical Analysis

With the development of science and technology, the scale and complexity ofsystem increase rapidly. The failure of equipment and system will not only bring greateconomic losses to the society, even can cause personal casualty and pollution to theenvironment. Thereby it is critical for fault diagnosis timely and effective. Faultdiagnosis methods based on the analytical model has been widely used in the researchof fault diagnosis, however its monitoring performance in a large part depends on theaccuracy of the mechanism model. It is often difficult to establish accurate mechanismmodels in practice. Fault diagnosis methods based on data driven don’t need toestablish precise mechanism model, just analyze the running data, therefore it gainsmore and more attention.Fault diagnosis methods based on multivariate statistical analysis are importantbranch of fault diagnosis methods based on data driven. This dissertation has mainlystudied multivariate statistical analysis methods based on the theory of multivariatestatistical analysis. The main research contents are as follows:(1) A fault diagnosis method based on relative transformation of informationincrement matrix is proposed. The method solves the problems that some faults of thevariables in the system can’t be detected performance in that although the absolutevalue of the variable exceeds the threshold value is small, but compared with thevalue of the variable itself is relatively large because of different dimension. Besides,this method can be better and more accurate for fault diagnosis, and can reduce thefalse positive rate and false negative rate effectively at the same time.(2) Inspired by the improvement of total projection to latent structures (T-PLS) topartial least squares algorithm (PLS), the method proposed by YIN is improved. Anew autoregressive total projection to latent structures (AR-TPLS) is proposed in thisdissertation. The new method not only avoids the complex solving process ofnonlinear iterative partial least squares algorithm (NIPALS) in partial least squares(PLS) and total projection to latent structures(T-PLS) proposed by ZHOU, but alsoovercomes the problem that process residual of modified PLS proposed by YIN stillhas large variations which are not proper to be monitored using Q-statistic. Besides,this method is simple and has less computation.(3) Introduced the ideas of relative principal component analysis (RPCA) to total projection to latent structures algorithm (T-PLS), to solve the problem because of thedifferent data dimension, and the effect of correlation coefficient matrix on faultdetective ability has been studied through simulation. In the process of fault detection,finding out the process variables whose effects on quality variables is larger accordingto the correlation degree of process variables and quality variables, then focus themon monitoring, has great significance.