Aeromagnetic attitude compensation for uninhabited aircraft systems without high-altitude calibration patterns using hybrid recurrent neural networks

Since the 1950s, Tolles-Lawson-based aeromagnetic compensation methods have been used to separate an aircraft's magnetic signal from signal associated with ground geologic and cultural features. This is done by performing a high-altitude figure-of-merit (FOM) flight and fitting the band-pass-filtered magnetic data to determine compensation parameters. This paper describes a supervised hybrid recurrent neural network (HRNN) algorithm trained on low-altitude survey data to perform aeromagnetic compensation. The proposed HRNN attitude compensation method can be employed for aeromagnetic surveys where traditional FOM and compensation are not possible. It has particular relevance for surveying via uninhabited aircraft systems (UAS). Firstly, the HRNN was tested on data from a fixed-wing airplane survey, and the results were compared to hardware-based compensation results. The standard deviation of the difference between the two methods for magnetic attitude correction (MAC) was 0.1 nT for the training region and 0.4 nT for the application region, respectively. Secondly, a UAS FOM flight at the highest permitted altitude in Canada, 120 m above ground level, showed similar improvement ratios for software-based least squares (LS) and the proposed HRNN algorithm of 3.5 and 2.6, respectively. The percent change and deviation in differences in MACs from LS to HRNN was 0.0% and 0.9 nT across small-box loops and –2.7% and 0.4 nT across large-box loops. Finally, LS and the proposed HRNN algorithm were applied to a 50 m altitude UAS data set for which no FOM flight was possible. LS did not successfully model aircraft noise, whereas the HRNN demonstrated effective removal of the magnetic signal due to aircraft attitude variations. The modeled HRNN MAC had a standard deviation of 2.4 nT.