A Radio Frequency Polarimetry Sensor for Aircraft Engine Health Monitoring

RF sensors are a non-contact technology that have the potential to provide on-wing aircraft engine health information. The RF sensor response to rotor whirl and compressor blade vibration was investigated through a series of experiments. Data processing methods involving neural networks were used to separate the desired RF response from the unwanted response.

The RF sensor response to whirl was experimentally investigated through static offset and rotor whirl tests. Static offset experiments showed a one-to-one invertible mapping between the rotor shaft location and the RF polarimetry data. A simple, multi-layer neural network was successfully used to perform a regression analysis to find this mapping. This work also showed that the RF sensors are sensitive to rotor whirl shape and size. Since the rotor shaft traces an ellipse once per revolution during rotor whirl, whirl was anticipated to be directly related to a once-per-revolution response in the RF signal. The RF signal was pre-conditioned by computing the variance of the RF signal at shaft rotational frequency. This pre-conditioned data was fed into a simple, multi-layer neural network. The neural network was able to predict the whirl major and minor axes lengths.

The RF sensor response to rotor rotation and blade vibration was experimentally investigated. The experiments showed that the RF sensors are sensitive to blade vibration for synchronous and asynchronous vibrations. A model of amplitude and frequency modulation was formulated in order to describe the RF sensor response to rotor rotation and blade vibration. It follows from this model that amplitude and frequency modulation causes a large number of asymmetrical sidebands to appear around the RF tone, and that the blade vibration response is in these sidebands. The RF results were consistent with this model. Data from the RF sensors were pre-conditioned based on the model of amplitude and frequency modulation and used as an input into a simple, multi-layer neural network. The neural network regression algorithm was successfully able to predict the blade vibration deflection by nodal diameter.