Two-Axis Accelerometer Calibration and Nonlinear Correction Using Neural Networks: Design, Optimization, and Experimental Evaluation

Currently, there is no robust method that could calibrate the accelerometer output without explicitly deriving the error model of the device and estimate the nonlinear parameters of the model. This article presents a methodology to approximate the output of two-axis thermal accelerometers based on neural networks (NNs) for calibration and nonlinear correction. This method uses the output of the accelerometer and the Earth's gravitational acceleration expected at a static position as data for training. The proposed method uses different optimization methods (adaptive moment estimation (ADAM), gradient descent, and gradient descent with momentum) to find the best solution using half mean squared error (HMSE) as the cost functions for evaluation. Experiments are conducted and presented to validate the NN-based calibration method using 2800 unseen data points.