Multivariate control charts for industrial processes fault detection, monitoring and diagnosis

This dissertation focuses on the development of better methods and tools for assessing when a multivariate process monitored with individual observations is in a state of statistical control, when it has shifted toward a faulty mode of operation, and recognizing a previously detected fault if its symptoms exist in a pre-built database. The main tool developed during this research is the soft sensor associated with a faulty mode of operation. The soft sensor is a metric able to quantify the process shift in the direction of that specific fault. It is understood that for each fault one soft sensor exists. It is presumed that a control system monitoring a number of soft sensors simultaneously can determine when one of them signals, and accordingly indicate that the associated fault may be present in the process.;The research developed a mathematical method to characterize faults and to build associated soft sensors, and determined their statistical distribution. The soft sensor is more effective in detecting the fault because it monitors only the process shift in the direction of that fault (it is fault specific). In this way it eliminates or diminishes the noise created by process shifts in other directions. The general T2 chart with individual observations can detect when a process is out of control without being able to associate any fault.;The techniques developed are illustrated on a number of simulation examples where the process shift was induced intentionally in different directions and with different magnitudes. In addition, three industry examples are analyzed with the methods developed in this dissertation.