Artificial Neural Network Calibration of Wide Range of Motion Biaxial Inclinometers

Effective calibration of sensors with fine-scale irregularities in their residual space requires large and dense calibration datasets. For the case of biaxial electrolytic inclinometers, these irregularities are not evenly distributed and concentrate in small regions that drive data resolution requirements. Using evenly spaced sampling for look up tables results in less irregular regions being over-sampled, burdening the calibration process. Artificial neural networks have the capability to optimally distribute a limited number of trainable parameters to minimize the residuals. This can have the benefit of reducing data collection requirements as well as reducing memory requirements. In this paper, we compare the residual model accuracy of a neural networks and look up tables for biaxial inclinometers with temperature variability. We control for neural network size by equating trainable parameters to lookup table data and control for data acquisition by the number of sample points. To avoid biasing the neural network, we introduce random perturbation to otherwise uniform data sampling locations. For temperature dependent validation, we found that the neural network reduced the difference in performance between the orthogonal measurement channels by 99% as compared to a look up table.