07 December 2001
Improved online calibration
Manufacturers may soon be able to cut sensor calibration costs while enjoying near-real-time fault detection, thanks to two decades of inferential sensing research ready to begin yielding benefits, according to a paper presented at the September 2001 Condition Monitoring and Diagnostic Engineering Management Conference in Manchester, England.
Three data modeling algorithms, the researchers wrote, are in wide use and "have progressed to become viable options for sensor calibration monitoring." One of the algorithms, the Multivariate State Estimation Technique (MSET), developed by Argonne National Laboratory and commercialized by Lisle, Ill.-based SmartSignal, now augments Delta Air Lines' maintenance program with postflight evaluation of in-flight data for 351 aircraft.
Nuclear power plants now use three data modeling algorithms: MSET, the Nonlinear Partial Least Squares (NLPLS), and the Autoassociative Neural Network (AANN), said Dr. Wesley Hines, a University of Tennessee researcher. There are similarities among them in operation, if not in the underlying mathematics.
Deploying any of the systems in an industrial environment is fairly straightforward, Hines said, and consists of installing an ordinary PC as an additional node on the LAN. Then the modeling software on the PC conducts analyses of a preselected data set to train itself to recognize normal and anomalous conditions. Custom software might be necessary to prepare proprietary data formats for use.
Once trained, the data collected by the in-situ monitoring systems goes to the software for continuous monitoring and interpretation. Operators receive near-real-time reports of anomalous conditions or sensor failures.
"Each system," Hines said, "must be developed using a data set that is assumed to contain error-free observations from all expected operating states. The time and oversight required to design and train each type of system vary significantly."
Each of the systems is theoretically capable of handling an unlimited number of sensor inputs. The MSET system runs best if it handles uncorrelated data sets separately, however, and AANN degrades noticeably when monitoring output from more than about 30 sensors. Each becomes unreliable when the input lies outside the training set. IT
Return to Previous Page