machine learning techniques for failure prediction in computer-controlled devices

We propose to implement an array of machine learning algorithms thatcan synthesize failure-prediction monitors with a minimum of human domain knowledge.Such a condition-monitor analyzes behavior data generated by computercontrolled devices --- data that can come from sensors on mechanicalequipment as well as sensors within the control software --- andidentifies evidence of impending malfunctions. The main challenge isto determine what behavior signals a possiblefailure. Traditionally this has been done by hand, but recent advancesin machine learning make it possible to automate that process. Ouremphasis will be on machine learning algorithms that can handle thenon-numerical data generated by software monitors as well as thenumerical data generated by sensors place on mechanical equipment. Failure prediction (or condition monitoring) is the basis of condition-basedmaintentance, which preempts equipment failures while potentially reducingmaintenance costs. Thus, our proposed technology can be beneficial in all areas ofindustry and government where computer controlled equipment is used.In addition to permitting condition-monitoring for computer-controlleddevices, our proposed technology may also allow condition-monitoringof information systems and executing software. This is a crucialelement in the development of increasingly autonomous, self-monitoringsoftware systems.