An Unsupervised Machine Learning Framework for Anomaly Detection in Antenna Surface Currents to Minimize Beam Chromaticity

Abstract

Frequency dependent variations of beam are a critical limitation in the Ultra-Wideband (UWB) systems and precision Cosmology experiments, affecting the system sensitivity and fidelity. In this work, we present an unsupervised machine learning framework to address this challenge by analyzing spatial deviation patterns in antenna equivalent surface currents. The equivalent currents are reconstructed from near-field electric measurements using a source reconstruction technique across more than a hundred frequencies within the antenna's operational band. A non-centered, weighted Principal Component Analysis (PCA) is employed to derive an idealized current distribution corresponding to minimized beam chromaticity. By comparing the magnitude of the reconstructed and idealized currents, a differential current map is generated. Uniform Manifold Approximation and Projection (UMAP) is then applied to reduce the dimensionality of this map in a frequency-indexed form, enabling clustering of frequencies using techniques such as K-means and Density-Based Spatial Clustering of Applications with Noise(DBSCAN). This allows automatic identification of critical frequency bands and associated spatial zones where the antenna structure requires modification. The proposed framework efficiently identifies coherent physical patterns across large, high-dimensional datasets, demonstrating the utility of machine learning in mitigating beam chromaticity and improving antenna sensitivity.