Process Fault Detection Based on Skew Gaussian Distribution Transformation and Canonical Variable Analysis Method

When the process data have non-Gaussian distribution characteristics, the fault detection model based on the Gaussian distribution hypothesis method will cause high false alarm. For this problem, a novel fault detection method, which based on skew Gaussian distribution transformation and canonical variable analysis (SGDT-CVA), is proposed. First of all, the Gaussian distribution transformation function is adaptively determined by the skewness of the process data. After transformation, the process data is transformed into obedient or approximately obedient Gaussian distribution. Then, CVA method is used to analyze the maximum correlation between variables, and the statistics are constructed according to the relationship between variables. Finally, experiments on the Tennessee Eastman process are used to illustrate the effectiveness of the proposed method and the results show the performance of the SGDT-CVA method is better than PCA, kNN and CVA.