Fault diagnosing of multivariate processes based on data mining

In modern industrial plants, large numbers of process measurements are stored in historical databases providing a potentially valuable source of process information. One potential use for historical plant data is as an aid in fault diagnosis. However, the information contained in these databases has been underutilized for several reasons. First, the volume of data that must be analyzed is enormous. Second, the data are multidimensional. Third, the variables are interrelated and need to be considered simultaneously in the analysis. In this thesis, a new data mining framework combining principal component analysis (PCA) and modern data mining techniques (k-Means clustering and decision tree induction techniques) is developed to exploit multivariate process data to detect and identify process faults. An extensive simulation study for a three-tank benchmark system demonstrates that this strategy is more effective than existing PCA methods in detecting system faults. It can also successfully distinguish between 20 different system faults.