Home > Innovations 2013 > MACHINE LEARNING BOOSTS HUMAN EXPERTISE TO PREVENT PLANE CRASHES

MACHINE LEARNING BOOSTS HUMAN EXPERTISE TO PREVENT PLANE CRASHES

Working closely with Honeywell engineers, ISIS researchers are mining regional airline data to build the Vehicle Integrated Prognostic Reasoner (VIPR), which uses knowledge derived from advanced data mining and machine-learning techniques to diagnose and detect potential problems in an airplane before an accident or emergency landing.

machine learning

The VIPR project aims to find evolving faults in aircraft systems, such as the engine and the avionics system, as well as anomalies that occur due to pilot actions and unusual environmental conditions, such as inclement weather or the orientation of a runway in a particular airport.

"We are one of the first projects that has taken this data and tried to apply intelligent analysis to help isolate, detect, and prevent adverse events," said Gautam Biswas, professor of computer science and computer engineering, who leads the NASA-funded VIPR project.

Even though plane crashes are rare, the growing complexity of aircraft systems has increased the chances for unexpected occurrences; hence the need to combine machine-driven exploration with human expertise to understand these situations, said Biswas.

"We're reaching limits in terms of how effectively human experts can analyze unusual situations," he said. "We must therefore find ways to use all our resources—human expertise and research in data mining and machine learning—to enhance existing knowledge. Analyzing the huge amounts of operational data that we have collected over the years will improve decision-making during flight operations, maintenance, scheduling, and overall airline management. The most important goal is improved airline safety and efficiency."

machine learning

VIPR uses advanced data mining and machine-learning techniques to explore and analyze large amounts of flight data to derive new and useful knowledge. Human experts then use that knowledge to improve diagnostic monitors and reasoning systems available on today's aircraft.

Biswas, graduate student Daniel Mack, and associate professor of computer science and computer engineering Xenofon Koutsoukos sift through the data collected by numerous sensors and monitors for up to fifty flights before an adverse event occurred. They consider mitigating factors, such as weather conditions, degrading equipment, and pilot error, and look for sequences of events that might have been overlooked, such as an evolving degradation in a fuel injection system that caused an engine to overheat and eventually shut down. More recently, they've generalized their approach to exploration methods that search for anomalies in terabytes of flight data.

"We found that in many cases, you could have reliably detected the likelihood of a particular problem occurRing by thorough and careful analysis of available data," Biswas said.

The ISIS team's activities go beyond conventional machine learning. Two sets of experts—aircraft engineers from both Honeywell and NASA are presented with diagnostic knowledge that can yield potential  innovations and safety features. For example, ISIS researchers working with Honeywell experts have discovered new monitors and more accurate diagnostic knowledge to detect faults in fuel supply lines, the fuel injection systems, and the engines themselves. Their results show that faults can be detected more quickly and accurately, allowing the initiation of maintenance actions in a timely manner to avoid compromising aircraft safety.

machine learning"Our task is to find ways to help experts do what they do better," Biswas said.

Biswas said real airline disasters and averted disasters alike motivate his work. For example, data analysis of Air France Flight 447 en route from Rio de Janeiro to Paris when it crashed into the Atlantic Ocean in 2009 indicated a junior pilot failed to understand sensor data and alarms that went off in the aircraft. On the other hand, the now famous Captain Sully had the experience and training to safely crash-land a U.S. Airways flight in the Hudson River that same year with no casualties.

The data mining technologies developed by ISIS may also help inform training methods, improve software-integrated design, and find systematic ways of analyzing the vast reams of data the FAA requires airlines to collect to inform decisions, rather than relying on current ad hoc methods for identifying problems.

Biswas and Koutsoukos came into the VIPR project with extensive experience in analyzing complex cyber-physical systems. "It is the interaction between software and hardware that really determines how the system behaves," said Biswas. "The depth of our expertise—combined with years of experience in this field—enables us to analyze interactions between different parts of the system and understand how these systems behave."

Phase one of VIPR received top honors in the 2011 Associate Administrator Awards for Technology and Innovation by NASA's Aeronautics Research Mission Directorate. "Receiving this award was particularly gratifying because ensuring flight safety has such a broad impact on both individuals and society," Biswas said.