Fault Detection Based On Multivariate Statistical Analysis

In order to ensure the reliability of product quality, and the stability ofindustrial systems, online monitoring of the production process is necessary. Modernindustrial system contains a large number of sensors, and the condition of thesesensors directly determines whether the system functions well or not. Therefore, thesensor fault detection has become an important research topic. The detecting methodwhich is based on data of the sensors does not apply for accurate modeling of theprocess. In this way, fault detection will be realized totally based on the datameasuring and controlling of the sensors. Research on this topic is described in thispaper. Based on the multivariate statistical analysis, series of sensor fault detectionmethods and applications are given in the paper, including Principal ComponentAnalysis (PCA) and Independent Component Analysis (ICA). In order to adjust tothe nonlinearity of industrial process, model based on Kernel Principal ComponentAnalysis (KPA) is raised, which is used in fault detection of blast furnace processand achieves high efficiency than the traditional PCA method. ICA method is widelyused in default detection of non-Gaussian process, while it can’t satisfy the statisticalneed of process monitoring. Therefore this paper propounds a modifying way basedon maximum mean deviation of statistics theory that subsets of the testing samplesare mapped to higher-dimensional kernel space. Then compare the data of the normalsamples and the testing samples in the space, and output the results of the detection.Experiments have been done in fault detection of gasifier section and the resultsdemonstrate that the new method is more efficient and convenient than the ICAmethod.