Inexpert Sensor Fusion Using an Adaptive Kalman Filter Approach

Sensor fusion in complex systems often relies on well-calibrated, expert sensors to achieve accurate and reliable outcomes. Integrating untrained or inexpert sensors into such networks poses significant challenges such as the cost associated with manually training them for the overall network objective. This paper presents an adaptive Kalman filter (AKF) approach to seamlessly incorporate an inexpert sensor into an established sensor network. By leveraging the existing trained sensors, the proposed method dynamically adjusts the inexpert sensor’s parameters, enabling it to contribute effectively to the overall system. We evaluate the performance of this approach using absolute errors and Kullback-Leibler (KL) divergence. Preliminary results show that the proposed AKF achieves a KL divergence on the order of 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−4</sup> in a binary classification task, significantly outperforming a recursive density estimation method. It is also shown that in a diverging multivariate target tracking problem, after being trained, two inexpert sensors successfully contribute to a sensor network, even after temporarily removing the expert sensor.