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

BPL modems deployed in power grid segments for communication purposes possess an underutilized capability: they continuously measure the noise frequency spectrum of the power line channel as part of their operation and channel diagnostics. This paper investigates whether this opportunistic sensing function --- available at no additional hardware cost --- can serve as a basis for non-invasive spectral fingerprinting and machine learning-based identification of connected electrical devices. Electrical devices introduce characteristic noise signatures into the 1.8--80 MHz BPL band, forming spectral fingerprints that can enable identification of both the device and its operating state. We first validate BPL modem measurements against a calibrated spectrum analyzer, confirming that the spectral shape is reliably preserved despite a systematic amplitude scaling of approximately 2×. A dataset comprising 75 device configurations of 38 physical devices --- spanning light sources, electric motors, switched-mode power supplies, and uninterruptible power supply units --- was recorded across three cable configuration variants of the same controlled measurement setup, with 20 repetitions per variant. The noise spectra are characterized by statistical descriptors computed over 16 contiguous 5 MHz sub-bands and classified using the CatBoost gradient boosting algorithm. The system achieves a device-specific detection accuracy of 86% and 98% when classifying by device category, provided that training data include measurements from the target cable configuration. Potential application domains include automated detection of devices exceeding EMC emission limits, identification of unauthorized or anomalous loads for cybersecurity purposes, device activity tracking, and predictive maintenance of power grid infrastructure.