Anticipating when a bridge needs replacing, when a nuclear power generating plant is going out of kilter and, perhaps, when the stock market is due to take a dive all might be possible with variations of the same basic computer program.

Indeed, researchers at Argonne National Laboratory near Chicago have developed software that can analyze patterns from just about any kind of complex signal and detect subtle variations associated with big changes before they happen.

The software was developed to help engineers discern whether warnings triggered by sensors monitoring steam pressure, flow and temperature in nuclear plants are legitimate causes for concern or mere false alarms.

In an industry where safety concerns dictate three and four levels of sensor redundancy, false alarms from failing sensors can cause costly, unneeded equipment shutdowns. Telling real alarms from false ones is a huge problem, one important enough for the federal Energy Department to sponsor a national contest to see if anyone could solve it.

Researchers were given 18 months worth of digitized signals from a nuclear plant in Florida, where 3,400 sensors fed information about the state of processing equipment. Buried within this mountain of data were 10 simulated sensor failures and the contestants were invited to find them.

Argonne's team not only spotted all 10 simulated sensor faults but its program also found two subtle disturbances in the data, suggesting potential failures that the Florida plant's engineers didn't know about.

None of the other contestants discovered any of the planted faults.

"We won that competition in 1996 and it really sparked a lot of interest from the industry," said Kenny Gross, an Argonne researcher.

The approach taken by Gross and his Argonne colleagues, including Ralph Singer and Stephan Wegerich, is to not only look at signals from monitors but also to study patterns in the so-called "noise," the part of the data thought to be meaningless variations in the signal associated with vagaries in the apparatus.

Gross and his colleagues have come to believe that this "noise" harbors useful information about the relationships among components in systems under scrutiny.

According to this theory, if one valve in a nuclear plant becomes sticky and is likely to fail eventually, subtle pressure changes measured directly by monitors near the valve might not spot the trend. But when combined with other subtle changes throughout the system that stem from the sticky valve, the problem might be diagnosed.

Called the Multivariate State Estimation Technique, or MSET, the software begins by looking at the signal patterns generated when everything in a process is working normally, and it then keeps looking at the patterns over time. Using its mathematical formulas, the computer can decide whether an alarm produced by one sensor is genuine or if the sensor itself is faulty, not the valve it's monitoring.

When MSET spots a faulty sensor, the computer can generate a virtual sensor signal to replace the faulty one, keeping the system running smoothly.

MSET's value to nuclear plant operations is why it was developed at Argonne, which has a mission of providing technology to the nuclear power industry. Once they proved MSET's value, the researchers decided their sophisticated signal analyzer probably had applications elsewhere.

Others agree.

The National Aeronautics and Space Administration has provided funding to MSET's developers to produce a version of their software that might help NASA avoid costly delays when it launches the space shuttle.

Developers believe even enterprises with lower profiles might benefit from MSET and they've found backers to provide capital.

The first target market has undertakings similar to nuclear plants, said Alan Wilks, vice president of development and operations for Smart Signal, which is based in Mt. Prospect, Ill.

"The chemical processing industry has lots of sensors and complex processes," Wilks said, "so that's a natural. At petroleum refineries, for instance, accurate monitoring is a life-and-death matter."

Some fatal refinery explosions have been caused by pressure sensor failures that went unnoticed because the sensor continued to register in the normal range when, in reality, pressure was building to explosive levels. By scrutinizing noise in the signal, MSET can detect such failures.

Published 9/23/1998