In-Situ Array Calibration Using Neural Networks

Array sensors are subject to various perturbations, e.g., due to imperfections or the sensor platform, and the nominal array model often fails to precisely describe the actual array response. Any mismatch results in systematic errors and performance degradation. This work presents an in-situ calibration technique using a neural network (NN). The calibration of the array sensor within its operational environment enables capturing platform-induced perturbations, which may not be fully reproducible in controlled chamber measurements. Using an NN to model the mismatch between the actual and nominal array response requires no assumptions about the underlying perturbations, such as mutual coupling. We propose a subspace-based loss function for training the NN using in-situ measurements, and show in numerical experiments that the proposed approach outperforms classical array calibration methods. We demonstrate the superior performance of the NN-based method using practical measurements from an experimental array sensor. The proposed method significantly reduces systematic errors.