Open Dataset Shows How Multisine Waveforms Carry Data and Power Wirelessly at Once

Imagine a battery-free sensor that receives both its instructions and its electricity from the same radio signal. That is the promise of simultaneous wireless information and power transfer, or SWIPT, a technology that has moved rapidly from theory toward laboratory reality as the Internet of Things swells with billions of low-power devices. Now a team at the University of Cyprus, working with a colleague at LUT University in Finland, has released an openly licensed experimental dataset that documents exactly how such a system behaves in the real world, and they hope it will become a standard benchmark for a growing community of researchers teaching machines to decode and optimize wireless power.

The dataset, published as a FAIR²-certified data article in Frontiers in Communications and Networks, grew out of a point-to-point SWIPT testbed built around multisine waveforms. In a multisine signal, the transmitter does not encode information in the conventional way, by modulating a single carrier with amplitude or phase changes. Instead, the message is carried in the number of sine-wave tones that make up the signal itself. The Cyprus team used tone counts of 2, 4, 8, 16, and 32 as five discrete signal classes, so that a receiver able to infer how many tones are present can recover the transmitted symbol while simultaneously harvesting the energy that the waveform carries.

The hardware behind the measurements reflects the kind of low-cost, low-complexity architecture that SWIPT researchers care about. On the transmit side, an NI USRP 2920 software-defined radio, driven by LabVIEW on an ordinary laptop, generated the multisine waveforms at 915 MHz within a total bandwidth of 200 kHz. A DA-915-01 dipole antenna with 1.7 dBi of gain in the GSM-900 band radiated the signal, deliberately chosen for its simplicity and fixed properties so that no adaptive beamforming degrees of freedom would confound the results. Every waveform was normalized to its amplitude envelope before transmission to prevent digital-to-analog converter saturation and front-end nonlinearity, keeping the signals inside the usable dynamic range of the radio.

The receiving chain is where the dataset becomes unusual. Rather than a conventional demodulator, the receiver uses a Powercast P21XXCSR RF energy harvesting module, a diode-based rectifying antenna or rectenna that converts incident radio waves into direct current. The harvested output feeds a WSN-EVAL-01 wireless sensor board and a Microchip XLP PIC24F access point, and the resulting power readout, expressed in milliwatts, is logged in real time through a HyperTerminal interface. Crucially, the same received signal that excites the harvesting circuit also produces the measurement feature, meaning the entire receiver performs its inference on energy-domain observations alone. That makes the data especially relevant to passive or nearly passive IoT receivers that have no budget for conventional signal processing.

The measurement campaign was systematic to a fault. The researchers swept a fully crossed experimental matrix of four transmitter gain settings, 65, 75, 90, and 100 percent of the USRP gain, corresponding to 20.15, 23.25, 27.9, and 31 dB, and six transmitter-receiver separations from 10 to 35 centimeters in 5-centimeter steps. At every one of the twelve resulting gain-distance combinations, all five multisine tone classes were transmitted repeatedly, with 750 samples recorded per condition and each sample labeled with its tone class. The antenna geometry, orientation, and line-of-sight placement were held fixed throughout, so the dataset cleanly characterizes the effects of transmitted power and path length on the harvested energy.

The resulting numbers trace a physically intuitive story. At the highest gain and shortest distance, mean harvested power reached 2.50 milliwatts with a maximum of 5.6 milliwatts, while at the lowest gain and longest distance the mean fell to just 0.26 milliwatts. Mean power declined monotonically with distance for every gain level, consistent with ordinary propagation loss, and higher gain consistently mitigated that loss. Perhaps more striking is the pattern tied to the waveforms themselves: median harvested power decreased monotonically as the tone count increased, from the highest values at two tones to the lowest at thirty-two, and this class separation persisted across every gain and distance setting. Because the tone classes remained well separated with limited overlap, even as absolute power fell, the measurements carry genuine discriminative information usable for classification.

To demonstrate that usability, the team ran baseline machine-learning experiments using multinomial logistic regression and support vector machines, with linear kernels at C equals 1 and radial basis function kernels at C equals 25 and gamma equal to 0.2. The authors are explicit that these classifiers serve as evidence that the data contain class-discriminative structure, not as records of the best achievable performance. What elevates the resource methodologically is its partitioning strategy. Dataset A, used for training, validation, and internal testing, contains measurements at the two higher gains and the three shorter distances, while dataset B is reserved entirely for external generalization testing at lower gains and longer distances. Models never saw dataset B during training, feature scaling, hyperparameter tuning, or model selection, so performance on B measures how well a classifier trained in one power-distance regime transfers to a different one.

The release follows FAIR² principles with unusual thoroughness. The data ship as machine-readable CSV files with a data dictionary, partitioning metadata, a JSON schema, provenance documentation, a reproducibility guide, and analysis scripts in Python 3.12 using scikit-learn, NumPy, pandas, and matplotlib. Metadata aligns with schema.org, W3C PROV-O, and the MLCommons Croissant specification, checksums allow third-party integrity verification, and the ODC-By v1.0 license permits unrestricted reuse with attribution. The certification process also documents known biases candidly: a single hardware platform, indoor short-range line-of-sight conditions, a fixed dipole geometry, and a single acquisition session without timestamps all constrain how far results can be generalized.

Those limitations are worth understanding before downloading. The stored values reflect the complete receiver chain rather than calibrated incident RF power, because the diode-based rectenna performs a nonlinear RF-to-DC conversion that is itself part of the measured response. At the lowest gain setting, where readings sit near the bottom of the receiver dynamic range, sensitivity limits and finite readout resolution may loom larger than at higher gains. There are no raw IQ captures or time-domain waveforms, so the dataset cannot support channel estimation or end-to-end deep learning directly from baseband samples, and environmental logs of temperature and humidity, though the harvesters carried the sensors, were excluded by design.

For the SWIPT community, the significance lies in reproducibility. Waveform design for simultaneous information and power transfer has produced a rich theoretical literature, but rigorous, openly shared experimental datasets that couple signal features to harvested energy under controlled hardware conditions remain scarce. By encoding information in tone count, capturing the nonlinear rectenna response, and partitioning the data for in-distribution and out-of-distribution evaluation, the Cyprus and Finnish team has created a resource intended for benchmarking tone-count classifiers, studying energy harvesting efficiency, and developing AI-enabled wireless systems that must do two jobs at once: talk and power their listeners.

Subject of Research: Experimental dataset from a multisine SWIPT system measuring tone-count classification and RF energy harvesting across transmitter gain and distance settings

Article Title: Multisine signal tone and power harvesting measurements across gain and distance settings

Article References: Stylianou, P., Faddoul, E., Selim Korium, M., & Krikidis, I. (2026). Multisine signal tone and power harvesting measurements across gain and distance settings. Frontiers in Communications and Networks, 7, Article 1911508. https://doi.org/10.3389/frcmn.2026.1911508

Image Credits: AI Generated

DOI: 10.3389/frcmn.2026.1911508

Keywords: SWIPT, wireless power transfer, multisine waveform, RF energy harvesting, rectenna, software-defined radio, machine learning, tone-index modulation, IoT, FAIR data, signal classification, University of Cyprus