9713518 Runger This award provides funding to develop guidelines for a data structure for multivariate statistical process control. A process control algorithm should detect a process anomaly as quickly as possible, without incurring excessive false alarms. However, in practice, (1) one can often select from among many variables (even hundreds) that are available, (2) the time order of the original data is not suitable for direct input into the algorithm, a synchronization of variables is needed, (3) variables have different measurement frequencies, (4) some variables are more responsive to process upsets than others, (5) some variables can be computed as functions of the others, and (6) a partition of selected variables is important to reduce complexity, improve performance, and simplify the interpretation of the control algorithm. A mathematical description of process data and process upsets will be used to develop recommendations for preprocessing the data to address these issues. Process data will be used to validate recommendations. If successful, this project will affect the type and structure of process data that is collected for process control. This will include the process measurements that are recorded, the frequency of measurements, the time ordering of measurements, as well as how the measurements will be allocated (or partitioned) to particular control algorithms. Preliminary research has shown that a data structure is an important element of a process control strategy that becomes a critical element as the number of variables under analysis increase. Recommendations from this project will provide a synergy between the data structures and the control algorithms to increase the usefulness, application, and performance of control methods. The results will also apply to data structures in areas such as neural networks, data mining, and signal processing, as well as process control.
