Automated Feature Selection for Embeddable Prognostic and Health Monitoring (PHM) Architectures

This work presents novel approaches for feature selection and alarm settings that can be exploited by automatic health monitoring systems that use vibrations of industrial machinery as a primary source for detection of failures and incipient faults. For any feature extracted from a sensor signal, a baseline is created that is accepted or rejected according to its statistical properties and the largest time constant of the system. The proposed framework determines alarms using an alarm coefficient that is motivated by established engineering norms, heuristics, and acceleration models. The operation of the architecture and the system performance are tested with industrial failure data.