Dataset of "Power Line Communication as a Sensor: Machine Learning-Based Spectral Fingerprinting for Non-Invasive Device Detection"

Broadband power line communication (BB-PLC) modems measure the noise spectrum of the power-line channel as part of their normal channel estimation. This dataset captures that opportunistic spectral readout for a controlled set of electrical loads, in order to test whether the noise a BB-PLC modem already reports can be used to fingerprint and identify the devices connected nearby. Contents. Noise spectra of 75 device configurations (38 physical devices in various operating states) plus one empty-grid baseline, recorded over a live 230 V, 50 Hz grid with a pair of G.hn BB-PLC modems acting as distributed spectral sensors. Each configuration was measured in three cable variants of the same controlled setup (a galvanically isolated sub-network behind an isolation transformer, a direct connection, and a 50 m cable), with 20 repetitions, on both MIMO channels, over the 1.8-80 MHz band (3201-3202 frequency points, approximately 24.4 kHz step). The archive contains 9,120 two-column noise-spectrum CSV files, dual-channel spectrum charts (PNG and SVG), GPS-stripped device photographs, a machine-readable device table (metadata.csv), pre-computed spectral statistics (summary_statistics.csv), and a Python loader with usage examples. Benchmark results. Using engineered spectral features and standard classifiers (gradient-boosted decision trees as the primary model), device identification is reliable when the target cable configuration is represented in training (per-installation use): up to 94% accuracy for the exact device and operating state, and up to 99% for the device category. When the model must generalise to an unseen cable configuration (transferable use), accuracy drops to 18-34% per device and 42-62% per category; only spectrally distinctive, strong emitters transfer, and a per-site empty-grid calibration helps only when it is measured on site. Operated as a per-installation anomaly detector trained only on the normal state, the method flags an unauthorised device with ROC-AUC of about 0.9. Note on amplitude values. The amplitudes in the CSV files are relative to the BB-PLC modem internal noise floor and are not calibrated to absolute physical power. They represent the spectral shape and relative magnitude of each device fingerprint and should not be interpreted as absolute noise-power measurements. Analysis and reproduction code (feature extraction, model evaluation, critical analysis, anomaly detection, and figure generation) is available at: https://github.com/musilp94/plc-device-detection-analysis