A Novel Machine-Learning-Based Batch Selection Method in Sparse Near-Field Scanning

This article presents a novel and efficient batch data selection method based on active and unsupervised learning in real-time near-field scanning. The new approach shows a remarkable advantage over random sampling in reducing the number of scanning data samples and the total scanning time. Moreover, careful hyperparameter tuning is unnecessary, as the performance is insensitive to the hyperparameter values. After randomly selecting some initial points, three key steps are sequentially and iteratively implemented. First, the query-by-committee (QBC) method is adopted to evaluate the uncertainties of the unobserved positions using different interpolation functions and select an "uncertain" group with the largest variances. Second, the weighted K-means clustering (WKMC) method divides the "uncertain" group into multiple clusters and selects the representative samples by choosing the most uncertain point in each cluster to ensure diversity. Finally, an estimated variance change (EVC) method is proposed to select the most representative samples to enhance diversity further. The proposed approach of this article, referred to as the QWE (QBC + WKMC + EVC) method, can well balance uncertainty and diversity with high efficiency. The newly proposed QWE approach is validated in both simulations and measurements of near-field scanning. The batch data selection method of this article can be extended to other multidimensional regression modeling problems in which data acquisition is expensive.