Accurate health estimates from HUMS vibration data

Condition Based Maintenance (CBM) of military helicopters are tracked by Condition Indicators (CI) calculated from Health Usage and Monitoring Systems (HUMS) vibration sensors. Even though many CIs have been proposed and implemented, they remain highly variable and difficult to interpret, leading maintainers to become desensitized to their output. Here we show that a sequential Monte Carlo algorithm operating a stochastic non-linear model of fault evolution can circumvent the shortcoming of the CI approach. The algorithm estimates fault magnitude probability distribution functions, which were compared to tear down inspections of removed components. We obtained a high accuracy rate (~95%) over all available data thanks to the excellent artifact rejection afforded by the algorithm. Data spanned all transmissions and hanger bearings over a 6-year operational history of a portion of the US military helicopter fleet, including combat operations. This approach could empower the maintainer to detect faults accurately and prior to all other existing warnings, while simultaneously reducing or eliminating false positives.