Improving airline safety

In an ongoing project for NASA, Pacific Northwest National Laboratory statisticians developed a system for analyzing flight data recorded aboard commercial aircraft.

"We developed an 'atypical flight detector' to detect anomalous flights that show patterns different from the norm. These flights set up red flags for airline staff to ask if there's something about the flight that would warrant further investigation," said Tom Ferryman, who directs PNNL's work for the NASA Aviation Performance Measuring System (APMS) project.

Many air carriers routinely record performance flight data, collecting measurements on between 50 and 3,000 variables per second. "This instrumentation collects flight information from beginning to end, gate to gate, flight after flight. We had to find a way to sift through it quickly to focus the limited time of domain experts on operationally significant problems," Ferryman said.

PNNL statisticians developed algorithms to make sense of the data by clustering flights into typical and atypical patterns. "We look for all those things experts say happen, but ideally should not happen, like slowing down by putting the flaps or gear down early, and we enable airline experts to find these envisioned things. But the experts can't envision all possible errors," Ferryman said.

PNNL developed algorithms that find unenvisioned conditions the experts did not think to look for in flight data. In one exercise these algorithms uncovered asymmetry in engine output for a set of flights involving a particular aircraft. This can indicate mismatched throttles, a condition that degrades aircraft fuel efficiency. The air carrier had not been looking for this problem, but PNNL's algorithms found the unexpected.

When the optical disc used to record flight data is swapped off the airplane and into a computer loaded with the APMS software, atypical conditions are quickly apparent to airline experts. Airline safety personnel can further investigate to determine whether they are unsafe conditions.

Although the analysis uses sophisticated statistical techniques, flight operations and safety staff are shielded from the statistics by user-friendly and intuitive visual representations.

Statistical algorithms developed for APMS have other applications, including monitoring health patterns in communities, use in the financial industry or, possibly, different approaches to air carrier security.