Abstract
<jats:p>This paper investigates classification features of five static Wi-Fi radio link obstruction scenarios based on Channel State Information (CSI) extracted using ESP32-S3 microcontrollers in promiscuous mode. A comparative analysis of the multidimensional CSI profile (114 frequency subcarriers) and the classical RSSI metric was conducted on a dataset of 22,500 packets. The averaged CSI amplitude provides high inter-class separability (Fisher Ratio = 2.65). Although the interquartile ranges of for specific scenario pairs do not overlap, the Silhouette Score (−0.003) in the PCA space indicates a partial overlap of semantically similar clusters. Machine learning evaluation (k-NN, SVM) demonstrated that while the windowed RSSI vector [, ] at W=50 achieves the highest overall accuracy due to temporal aggregation (~0.5 s delay), the multidimensional CSI spectral analysis enables instantaneous packet classification (SVM accuracy 76.9%, F1-score 0.756) and provides better structural differentiation of adjacent classes with similar attenuation levels. The findings, obtained under a specific laboratory configuration (ESP32-S3, 24 m² room, 28 active APs), establish an experimental basis for developing cooperative spatial monitoring methods under similar deployment conditions. Keywords: channel state information, classification algorithms, dimensionality reduction, embedded systems, feature extraction, microcontrollers, OFDM, principal component analysis, received signal strength indicator, wireless LAN.</jats:p>