Machine Learning for Failure Prediction in Computer-controlled Devices

Sponsored by: Missile Defense Agency

It has been postulated that software entities are more complex for their size than any other human construction. The problem of monitoring the condition of computer controlled devices - determining when such a device is in need of maintenance—is correspondingly more complex. Furthermore, software evolution is pervasive, and condition-monitoring devices that work before control software is modified may not work afterwards. Together with rapidly growing reliance on increasingly numerous computer-controlled devices, this creates a need for technology that automatically synthesizes new condition monitors. We implemented machine learning algorithms that can synthesize condition monitors with a minimum of human domain knowledge.

Such a condition-monitor analyzes behavior data generated by computer controlled devices - data that can come from sensors on mechanical equipment as well as sensors within the control software - and identifies evidence of impending malfunctions.

Contact