Calibration of IMUs using Neural Networks and Adaptive Techniques: Targeting a Self-Calibrated IMU
Examensarbete för masterexamen
Systems, control and mechatronics (MPSYS), MSc
This thesis presents an investigation of different sensor models for calibrations of IMUs and suggests an approach for adaptive calibration which is intended to be utilized during use of the sensor. The results show that using a neural network as sensor model can reduce the calibration error compared to if a linear sensor model is used. The results also show that the suggested adaptive algorithm can be used to approximate the sensor model but is not sufficiently accurate. An IMU is a set of inertial sensors that can, among other applications, be used to estimate position and orientation of a body. An accurate calibration of the IMU improves the performance of the estimations and is hence an essential operation. The most commonly used calibration algorithms today use linear sensor models although the characteristics of the sensor are often argued to be nonlinear. The thesis contains an investigation of whether neural networks efficiently can be used for nonlinear sensor model approximation. It was found that, for offline calibration, neural networks can approximate the sensor model accurately but needs more data than the linear model. More precisely, using a neural network trained on 400 data points, the calibration error was halved compared to a calibration using a linear model. A problem with MEMS gyroscopes and accelerometers is that their characteristics change over time and are affected by parameters in their surroundings. To maintain an accurate approximation of the sensor model, the sensor would need to be re-calibrated. Most commonly used calibration methods today are time consuming and require expensive hardware, which make them complicated to perform in field. An algorithm for adaptive online calibration of IMU sensors has been introduced. This method is designed to be used continuously to update the calibration and thereby adapt the sensor model to changes in the sensor’s characteristics. The algorithm is iterative and alternates between calibrating the accelerometer using the gyroscope as reference and calibrating the gyroscope using the accelerometer as reference. To approximate the sensor model, neural networks are used. The result from evaluation of the developed adaptive algorithm shows that it can be used to approximate the sensor model but does not obtain as high accuracy as the offline algorithms. Hence it needs further development to be useful in practice.
Transport , Annan data- och informationsvetenskap , Transport , Other Computer and Information Science