Search

As industrial processes become more and more complicated and our ability to capture data continuously enhanced, it is becoming increasingly difficult for standard SPC either to recognize a state of statistical control or to identify departure from one. This is because (1) systematic non-random patterns may be present throughout the data in the forms of auto-correlation, cross-correlation and/or input-output correlation, (2) the complexity and automation. Diagnosis relies more and more on sensor synthesis supported by proper modeling strategy and information association at a subsystem or system level. As a new approach to multivariate process monitoring and diagnosis, designated component analysis (DCA) is proposed based on the concept of principal component analysis (PCA). Instead of estimating variation patterns solely from the process data as in PCA, DCA defines a series of mutually orthogonal vectors to represent fault patterns according to prior engineering knowledge, estimates their statistical significance from data, and analyzes the correlation among the designated components. Thus the engineering model is integrated into the statistical model and sensor integration can be implemented on a system level. This approach enables identification and isolation of multiple faults that exist simultaneously. It also facilitates tracking multivariate variation propagating through multiple stations. As a result, DCA improves monitoring and diagnosis in both single and multiple stations. Its effectiveness and unique advantages are demonstrated through simulation and application examples in automobile body assembly. For the multivariate processes with strong autocorrelation and input-output correlation, a MIMO model-based approach is proposed. An expanded time series model is utilized to exclude effect of autocorrelation and account for the influence of input on the output. The predictions and residuals based on this model enable separation of impacts from process inputs and those from within process disturbances. A cross-correlation model is then used to further account for changes in the predictions and residuals. Simulation and application results demonstrated the advantages of the proposed method in dealing with correlated multivariate processes.