Multisine signal tone and power harvesting measurements across gain and distance settings

1 Introduction

Wireless power transfer technologies enable energy delivery to remote and mobile devices without the need for physical connectors (Psomas et al., 2024). In applications such as the Internet of Things (IoT), simultaneous wireless information and power transfer (SWIPT) allows the concurrent transmission of data and energy to support low-power sensors and embedded systems (Krikidis et al., 2014; Perera et al., 2018). The efficient exploitation of SWIPT requires not only innovative waveform design (Krikidis and Psomas, 2019; Ayir et al., 2021) but also rigorous, reproducible datasets that describe both physical layer signal delivery and resulting energy harvesting to facilitate transparent performance benchmarking and advanced analytics (Dhull et al., 2024; Pan et al., 2018; Shariati et al., 2018).

Recent developments in integrated wireless systems further motivate the need for reusable SWIPT datasets. For example, multi-functional wireless architectures have been proposed to integrate sensing, communication, and wireless power transfer within the same system, including recent research on multi-functional reconfigurable intelligent surface (RIS)-assisted integrated sensing, communication, and powering (Magbool et al., 2025). These developments highlight the importance of experimentally grounded datasets that document communication-oriented signal features and receiver-side power transfer responses under controlled hardware conditions.

The dataset characterized in this study documents experimental measurements from a multisine-waveform-based SWIPT system, acquired through a laboratory testbed that integrates software-defined radio (SDR) tools, energy harvesting hardware, and supervisory data acquisition. In the study, the transmitter encodes information by varying the number of carriers (or tones) in the multisine signal, while the receiver is equipped to harvest incident radio frequency (RF) energy using a diode-based rectifying antenna (rectenna). This configuration is designed to reflect practical scenarios in which energy and data must be transmitted simultaneously to passive receivers in settings representative of future low-power wireless networks.

Structured experimentation was conducted across multiple parameter regimes, systematically varying the USRP transmitter’s gain and the physical separation distance between the transmitter and receiver. For each combination of gain and distance, signals with distinct tone counts were transmitted, corresponding to discrete experimental classes. The receiver’s power readouts—including received signal strength or harvested output—were captured in real-time via HyperTerminal interface and logged for each repeat measurement. This process yielded a tabular dataset that comprised individual sample identifiers, receiver-side power measurements for each experimental setting, and indicator variables denoting the transmitted multisine waveform class. The design explicitly distinguishes between datasets used for model training/validation and those reserved for external testing/generalization, thereby facilitating the assessment of generalization performance and supporting robust machine-learning benchmarking.

To ensure interoperability and AI-readiness, the dataset was structured as machine-readable CSV files, with detailed data dictionaries and partitioning metadata describing the setup, measurement context, experimental factor levels, and intended use of each subset. Each numeric measurement variable encodes a specific transmitter gain and distance scenario, with corresponding units (milliwatts) and explicit mappings to multisine tone count indicator classes. The dataset includes comprehensive information to support traceability from the device and protocol level through to individual sample acquisition, enabling reproducible workflows in signal classification and SWIPT energy analysis.

FAIR² principles are here explicitly implemented, emphasizing the following: findability through richly described variable and file naming; accessibility via standardized open formats; interoperability through alignment to schema.org and PROV-O metadata; reusability by providing comprehensive documentation of experimental design, variable semantics, and context of data use. Air-readiness and responsible AI requirements are addressed by including detailed metadata on experimental partitioning, model development inclusion, and variable definitions, ensuring transparency with respect to generalization and potential downstream bias considerations.

This data publication summarizes the development, structure, and descriptive characteristics of the SWIPT multisine signal dataset and its associated testbed metadata. By making clear the measurement conditions, experimental configurations, and annotation standards, this resource is intended to support reproducible engineering, communications, systematic benchmarking, and ethical reuse. Potential applications include the development and evaluation of supervised learning algorithms for waveform recognition, energy harvesting characterization, and the establishment of reproducible and transparent signal processing pipelines for AI-enabled wireless systems.

2 Methods summary

2.1 Study design

The experimental study was designed to investigate simultaneous wireless information and power transfer (SWIPT) in a controlled point-to-point radio frequency environment. Information transmission was realized by encoding data into the number of tones present within multisine signals. The receiver, equipped with a diode-based rectenna, was responsible for converting incident RF signals into harvested energy for analysis. Key experimental parameters were systematically varied: the number of signal tones, the transmitter gain settings of the software-defined radio (SDR) device, and the distance between the transmitter and the receiver. This design enabled the resulting dataset to support both the classification of signal types and the evaluation of energy harvesting performance across differing physical and operational conditions. Throughout, the study focused exclusively on scenarios achievable within the implemented SDR-based testbed and did not attempt to generalize findings to wireless systems or applications beyond those explicitly detailed.

2.2 SWIPT system implementation and signal generation

The implementation phase encompassed hardware selection, integration, and the procedural generation of multisine signals. The primary transmission hardware was an NI USRP 2920 SDR, operated through NI LabVIEW installed on an HP Probook laptop featuring 16 GB RAM and a Core i5 processor. This SDR platform was coupled with a DA-915-01 dipole antenna, offering 1.7 dBi of gain in the GSM-900 frequency band. The physical testbed allowed for adjustable transmitter-to-receiver distances, supporting the systematic interrogation of wireless propagation effects.

The implemented receiver corresponds to a low-complexity integrated RF energy harvesting receiver architecture. The receiver chain is based on the Powercast P21XXCSR RF energy harvesting module, followed by the WSN-EVAL-01 wireless sensor board and the Microchip XLP PIC24F access point. In particular, the same received RF signal that excites the energy harvesting circuit produces a receiver-side RSSI/power readout in milliwatts. This readout is logged through HyperTerminal and used as the measurement feature for machine learning-based inference of the transmitted multisine tone class. Therefore, the dataset is particularly relevant to low-complexity SWIPT receivers in which information recovery is based on energy domain observations.

Each multisine signal was defined by a set number of tones, with parameters including the center frequency (f c = 915 MHz), a total bandwidth of 200 kHz, and the tone count N drawn from the set 2 , 4 , 8 , 16 , 32. Each tone in the multisine waveform was constructed according towhere the amplitude A n for each tone was set to 1 / N to enforce normalization, f n was calculated as f 0 + n − 1 Δ f, and θ n denoted the phase of the n-th sinusoid. Prior to transmission, all waveforms were normalized to their respective amplitude envelopes to mitigate the risk of digital-to-analog converter (DAC) saturation and front-end non-linearities. These normalization steps ensured that the transmitted signals remained within the operable dynamic range of the USRP hardware, promoting the fidelity and repeatability of signal delivery.

x ∼ t = ∑ n = 1 N A n ⁡ cos 2 π f c + f n t + θ n ,

2.3 Receiver chain interpretation and rectenna nonlinearities

For each generated waveform, all relevant configuration parameters—including tone count, amplitudes, center frequency, frequency spacing, transmitter gain, and the spatial layout of the nodes—were documented. These records formed the foundational metadata, enabling traceability from raw transmission to data analysis.

The receiver-side values reported in the dataset should be interpreted as the output of the implemented receiver chain. The receiver chain consists of the Powercast P21XXCSR RF energy harvesting module, the WSN-EVAL-01 wireless sensor board, the Microchip XLP PIC24F access point, and the HyperTerminal logging interface. Since the RF-to-DC conversion is performed by a diode-based rectenna module, the relationship between the incident RF waveform and the reported power readout is nonlinear. This nonlinearity is part of the physical receiver response captured by the dataset.

2.4 Antenna choice and spatial layout

The DA-915-01 dipole antenna was selected because it provides a simple, fixed-gain transmission element operating in the GSM-900 frequency band. This choice supports repeatability and avoids introducing additional adaptive antenna or beamforming degrees of freedom into the dataset. The objective of the experiment was not to optimize antenna geometry but to collect receiver-side SWIPT measurements under controlled variations of multisine tone count, transmitter gain, and transmitter-receiver distance.

The recorded energy harvesting response depends on the spatial layout of the testbed, including the transmitter–receiver separation, antenna orientation, antenna gain, and line-of-sight placement. These factors determine the RF power delivered to the rectenna and therefore affect the total harvested power readout. During the measurement campaign, the antenna configuration was kept fixed while the distance and transmitter gain settings were varied systematically. Consequently, the dataset characterizes the implemented fixed-geometry SWIPT testbed.

The dataset reports the final receiver-side RSSI/power values obtained from the Powercast-based readout chain and logged through HyperTerminal. Therefore, the stored values reflect the complete receiver and digitization chain used in the experiment, rather than raw ADC codes or baseband samples. At the lowest transmitter gain setting of 65% USRP gain, corresponding to 20.15 dB, the received power values are closer to the lower part of the receiver dynamic range. In this regime, finite readout resolution, receiver sensitivity limits, and possible floor effects may have a larger relative influence on the recorded values than in the higher gain cases.

Consequently, the dataset should be interpreted as containing receiver-side power readouts from the implemented SWIPT receiver chain rather than calibrated samples suitable for independent digitization-resolution analysis. Future extensions of the dataset may include raw ADC/readout values, SDR/IQ captures, or calibration measurements to enable a direct assessment of quantization effects and their impact on classification at low received-power levels.

2.5 Measurement protocol and dataset formation

To ensure data representativeness and completeness, the measurement protocol followed a fully crossed experimental matrix. All combinations of transmitter gain settings (31, 27.9, 23.25, and 20.15 dB), separation distances (10, 15, 20, 25, 30, and 35 cm), and number of multisine tones (N ∈ 2 , 4 , 8 , 16 , 32) were explored. At each unique experimental condition, multiple power samples were acquired to support both statistical analysis and machine learning requirements.

Power harvesting measurements were captured using the WSN-EVAL-01 sensor board and digitized through the associated Microchip XLP PIC24F interface. Each measurement was transferred to a host computer via HyperTerminal, allowing for consistent annotation and storage into a tabular structure. All samples were uniquely identified and accompanied by field metadata that detailed the experimental parameters for reproducibility.

Subsequent data annotation protocols assigned sample-specific labels, indicating the number of transmitted tones, gain setting, distance, and a group identifier corresponding to either training/validation/internal testing (dataset A) or external/generalization testing (dataset B). Variable names and units were standardized according to the dataset schema (e.g., power expressed in milliwatts). These measures facilitated downstream partitioning and enabled consistent machine learning processing. Initial quality checks were conducted at this stage to verify the presence of all required samples, appropriate numerical formatting, and the resolution of missing or anomalous values.

To improve the traceability of the data generation process, Figure 1 summarizes the complete workflow followed to produce the final dataset. The process begins with the selection of the multisine waveform parameters, including the number of tones, carrier frequency, bandwidth, and transmitter gain. The waveform is then generated in NI LabVIEW and transmitted using the NI USRP 2920 at 915 MHz. The transmitted signal is radiated through the DA-915-01 dipole antenna and received by the Powercast P21XXCSR RF energy harvesting module. The resultant receiver-side RSSI/power readout is transferred through the WSN-EVAL-01 wireless sensor board and Microchip XLP PIC24F access point and logged through HyperTerminal in milliwatts. The recorded values are then organized into a wide-format CSV table, annotated according to the gain-distance condition and multisine tone class, and partitioned into datasets A and B. Finally, the data is packaged with metadata, documentation, schema files, and reproducibility material for FAIR² publication.

2.6 Machine learning evaluation and partitioning

Supervised machine learning protocols were employed to assess the classification potential of harvested power data and evaluate model generalization. Each data sample was allocated to datasets A or B based on its transmitter gain and physical distance conditions. Dataset A is used for model development, including training, validation, and internal testing, and contains measurements collected at 100% and 75% USRP transmitter gain over distances of 10, 15, and 20 cm. Dataset B is reserved for external testing and generalization assessment, and contains measurements collected at 90% and 65% USRP transmitter gain over distances of 25, 30, and 35 cm. This condition-based separation was selected to evaluate whether models trained under shorter-distance and selected high-gain conditions can generalize to different gain and longer-distance conditions. Dataset B was not used during model training, feature-scaler fitting, hyperparameter tuning, or model selection. A summary of the dataset partitioning is given in Table 1.

TABLE 1. Dataset classification.

A parsing and tagging routine systematically assigned group variables, annotated usage (training, validation, testing, or generalization), and enabled traceability for each combination of gain and distance. These annotations ensured clarity regarding the experimental scenarios under which each measurement was obtained and its analytical role.

Classification was conducted using multinomial logistic regression (MLR) and support vector machine (SVM) algorithms. MLR was applied for multiclass classification with the signal index as the target, while SVMs were assessed on both binary and multiclass signal class identification. Model evaluation encompassed multiple metrics, including classification accuracy, precision, recall, F1-score, and operational measures such as learning curve analysis and execution timing. Internal validation was executed on dataset A, while external generalization capacity was quantified by measuring the classifier performance exclusively on the held-out dataset B, which was not exposed during training or parameter selection. Inclusion and analytical use for each record were stringently documented, precluding any leakage of test data into model development.

The SVM classifiers are included as baseline dataset validation tools. Their purpose is to demonstrate that the receiver-side power measurements contain class-discriminative information that can be exploited by standard supervised classifiers. The reported SVM results should be interpreted as baseline evidence of dataset usability, not as the best achievable performance. The linear SVM uses (C = 1), while the RBF-SVM uses (C = 25) and (gamma = 0.2). These values were selected as fixed baseline parameters for reproducible evaluation and visualization. No claim is made that they are globally optimal.

2.7 Software and computation environment

The software used throughout the experimental acquisition and analysis are summarized in Table 2 below.

TABLE 2. Software used throughout the experiment.

2.8 Data finalization, formatting, and documentation

After all measurements and analyses, dataset finalization focused on ensuring data integrity, adherence to schema definitions, and comprehensive metadata capture. Each tabular dataset and associated schema was rigorously reviewed for alignment with source documentation and experiment logs. Field names were verified for consistency, and numerical values and descriptors were cross-checked for conformity with the experimental matrix. Data type integrity was validated, and entry counts were reconciled against sample records.

Final datasets were exported in CSV format, with each variable annotated for unit, domain, and semantic function. Special columns, such as indicator variables for class labeling and sample IDs, were explicitly described in accompanying datasheets. Measurement units (e.g., milliwatts for power) were systematically captured for all numeric fields.

Complete metadata were compiled for each record, including transmitter gain, distance, dataset group, analytical inclusion status, and processing completeness. Dataset integrity was assessed through presence–absence checks, duplication review, and analysis of missingness. The prevalence of distinct field values within each variable was computed and documented to support quality assessment and provide potential end users with relevant statistical context.

Machine learning results, partition assignments, and analytical outcomes were recorded in the datasheet, ensuring that users could clearly identify the training, validation, or test/generalization role of each sample. Inclusion and intended-use codes were assigned following schema guidelines.

Finally, the preparation for public dataset release included computation of checksums or digital signatures to enable third-party integrity validation. All data tables, partitioning records, and datasheets were assembled into a machine-readable package and deposited in a public repository. This approach promotes accessibility, transparency, and full traceability, supporting robust reproducibility and reuse within the scientific and engineering communities.

3 Data overview

3.1 Data summary

The dataset comprises systematically collected receiver-side power measurements that support the evaluation of simultaneous wireless information and power transfer (SWIPT) via multisine signals in a controlled point-to-point radio frequency testbed. It contains digitally annotated records that reflect all combinations of three primary experimental variables: number of tones in the synthesized multisine waveform (N = 2 , 4 , 8 , 16 , 32), transmitter gain (65%, 75%, 90%, 100%; corresponding to 20.15, 23.25, 27.9, and 31 dB), and transmitter–receiver separation distance (10–35 cm, in 5 cm increments). Receiver-side power is measured in milliwatts (mW) for each configuration using a software-defined radio (SDR) transmitter, dipole antennas, and a diode-based energy harvesting receiver, with all signal generation and acquisition parameters algorithmically logged.

The dataset is distributed in tabular CSV format and organized with explicit variable naming (e.g., Gain100_Distance10) to denote each gain–distance condition, accompanied by a unique sample identifier (ID) and a class indicator for multisine tone count. Individual measurement matrices provide 750 annotated examples per variable for a total of 750 × 12 numerical records. Explicit group assignment metadata is included, partitioning records into training/validation/internal testing (dataset A) and external testing/generalization (dataset B), based on transmitter gain and distance scenarios, to support downstream machine learning reproducibility and model validation.

Accompanying datasheets document field definitions, value ranges, measurement units, and all experimental conditions, thus ensuring traceability and facilitating reuse. All power measurements are accompanied by transmitter configuration, distance, dataset group, and tone count indicator variables. Dataset finalization includes count and type integrity checks, absence of missing values, and documentation of potential reuse scenarios for signal classification and energy harvesting analysis. The resulting resource enables standardized benchmarking, partitioned evaluation, and AI-readiness for reproducible experimental research in SWIPT and RF-based energy harvesting systems.

3.2 Quantitative summary of the dataset

3.2.1 Total number of records

The multisine_signals_v1.0.0.csv dataset comprises 750 records, each corresponding to a single experimental instance. Every record is uniquely identified by an integer-valued ID field ranging from 1 to 750. For each measurement sample, receiver-side power is recorded simultaneously across 12 distinct experimental conditions defined by transmitter gain and transmitter–receiver distance parameters. This structure results in 13 columns: 12 measurement fields and one indicator column that labels each sample according to the multisine tone count applied during that event. The dataset_partitioning_v1.1.0.csv provides 12 records that annotate dataset partitioning and measurement field mapping for each unique gain–distance combination.

3.2.2 Coverage

Receiver-side power is measured under 12 unique transmitter gain and distance configurations. These configurations comprise four gain settings (65%, 75%, 90%, and 100%, corresponding to 20.15, 23.25, 27.9, and 31 dB, respectively) and six spatial separation distances (10, 15, 20, 25, 30, and 35 cm), producing a cross-factorial experimental design. Each configuration is encoded as its own column (e.g., Gain100_Distance10, Gain75_Distance15), with each having exactly 750 measurement values. Samples are assigned to one of two mutually exclusive groups for machine learning: group A (training/validation/internal testing) or group B (external testing/generalization), which are distributed evenly across the configuration space as documented in the partition file. The multisine signals are further categorized by an Indicator variable, representing five possible signal classes that correspond to the number of waveform tones (N = 2 , 4 , 8 , 16 , 32).

3.2.3 Attribute diversity

Across the dataset, 13 attributes are provided: 12 numeric measurement fields recording harvested power for each gain-distance condition, and a single categorical attribute, “Indicator”, denoting the multisine tone count class. The variable names explicitly encode the transmitter power and separation parameters, supporting granular traceability between measurement and experiment configuration. The Indicator attribute is uniformly distributed, with each of its five possible integer values (0–4, mapping to N2, N4, N8, N16, and N32) represented by exactly 150 records, supporting balanced class assignments for downstream analytical workflows.

3.2.4 Quantitative attributes

All 12 measurement columns record receiver-side power in milliwatts (mW) as floating point values. Descriptive statistics for these attributes are as follows.

• Gain100_Distance10: mean 2.50 mW, standard deviation 1.76 mW, min 0.17 mW, max 5.6 mW.
• Gain100_Distance15: mean 1.47 mW, standard deviation 0.95 mW, min 0.15 mW, max 3.18 mW.
• Gain100_Distance20: mean 1.18 mW, standard deviation 0.77 mW, min 0.13 mW, max 2.88 mW.
• Gain75_Distance10: mean 1.14 mW, standard deviation 0.97 mW, min 0.1 mW, max 3.32 mW.
• Gain75_Distance15: mean 0.78 mW, standard deviation 0.63 mW, min 0.09 mW, max 2.08 mW.
• Gain75_Distance20: mean 0.59 mW, standard deviation 0.43 mW, min 0.09 mW, max 1.63 mW.
• Gain90_Distance25: mean 0.64 mW, standard deviation 0.45 mW, min 0.09 mW, max 1.63 mW.
• Gain90_Distance30: mean 0.45 mW, standard deviation 0.28 mW, min 0.09 mW, max 1.42 mW.
• Gain90_Distance35: mean 0.43 mW, standard deviation 0.26 mW, min 0.09 mW, max 1.01 mW.
• Gain65_Distance25: mean 0.38 mW, standard deviation 0.28 mW, min 0.09 mW, max 1.04 mW.
• Gain65_Distance30: mean 0.29 mW, standard deviation 0.18 mW, min 0.09 mW, max 0.7 mW.
• Gain65_Distance35: mean 0.26 mW, standard deviation 0.15 mW, min 0.09 mW, max 0.63 mW.

The distributions are unimodal for each field, with the modal power value varying by configuration and exhibiting reduced mean and maximum values at lower gain and greater distance settings.

3.2.5 Data completeness

The dataset demonstrates complete coverage without missing values in all 13 fields. Each measurement variable contains 750 non-null numerical values as expected, and all indicator/class assignments are present and balanced across the defined signal classes. There are no duplicate IDs or inconsistencies in field assignment, as verified by data type and range checks. Top-ten value frequencies for each measurement field indicate consistent representation and no anomalies, such as outlier repetition or absent intervals.

3.3 Dataset availability

The dataset is FAIR²-certified and publicly available under the Open Data Commons Attribution License (ODC-By v1.0), permitting unrestricted reuse with appropriate attribution. Access is provided via (1) the interactive FAIR² Data Portal, which supports the visual inspection of waveform classes, feature distributions, and data completeness, and (2) the downloadable FAIR² Data Package, containing raw and processed signal data, structured metadata, code for data generation and analysis, and detailed methodological documentation. The dataset is archived and indexed in an external repository with persistent identifiers, ensuring long-term access and integration with relevant data ecosystems. All experimental protocols and data processing procedures are fully described in the accompanying metadata and methods files. Access and documentation are available at: https://doi.org/10.71728/senscience.gk8g-g8p8 and https://github.com/EliioF/Multisine-signal-dataset. A tabular summary of the repository files is provided in Table 3.

TABLE 3. Summary of repository files.

3.4 FAIR² compliance introduction

The FAIR² compliance certification table provides a structured assessment of the dataset’s alignment with the principles of findability, accessibility, interoperability, and reusability (FAIR), together with extensions for AI-readiness (AIR) and responsible AI (RAI). This table offers a high-level summary of key dataset attributes, covering data identification, availability, metadata standards, interoperability protocols, transparency of provenance, and readiness for ethical and responsible AI research. Each compliance dimension is evaluated with reference to the supporting documentation, implementation measures, and supporting infrastructure as documented in the associated datasheets, metadata files, and provenance records. The certification enables users and stakeholders to rapidly evaluate the adherence of the dataset to community-endorsed data management principles and AI requirements, supporting downstream integration, traceability, and reuse. This structured overview ensures that all critical aspects for machine learning and advanced analytics applications, as well as responsible and transparent data handling, are systematically addressed.

3.5 FAIR² certification

Table 4.

4 Dataset visualizations

This section presents a visual summary of the SWIPT multisine signal dataset, characterizing the relationships among transmitter parameters, separation distance, and harvested power, and highlighting features relevant for machine-learning-based signal classification. The visualizations delineate how the experimental design systematically modulates transmitter gain and spatial separation to generate a matrix of receiver-side power measurements under controlled conditions. Plots of mean harvested power versus distance, stratified by transmitter gain, demonstrate consistent signal attenuation with increasing separation and elevated power levels under higher gain conditions, reflecting physically plausible propagation effects and the capacity of the dataset to capture differential energy harvesting efficiency across the tested parameter space.

Distributions of power measurements for each transmission scenario, illustrated via histograms and kernel density estimates, reveal variations in both central tendency and dispersion across gain-distance pairs. These distributions highlight how hardware-induced variability and experimental tuning affect received signal profiles, with high-gain and short-distance conditions producing greater power and spread while lower gain and longer distance compress readings toward lower values. Such systematic coverage supports direct comparison of power variability and supports transparent benchmarking across SWIPT conditions.

Class-conditional analyses, using violin and box plots of harvested power by multisine tone count, establish a clear separability of tone classes based on receiver power features. The monotonic decline in median power with increasing tone count is robust across all transmitter settings, underscoring the suitability of the collected measurements for signal classification tasks. The limited overlap between class distributions—particularly in high-gain, short-distance scenarios—suggests favorable signal-to-class assignments for machine learning models, while persistent separation across all conditions enables the assessment of model performance under domain shifts.

The dataset structure and partitioning are further visualized through schematic matrices and partitioning diagrams, which clarify the allocation of samples to model development (internal) and external generalization (test) groups. Training and validation are restricted to data obtained at higher gains and shorter distances, while external testing is confined to lower gain/longer distance conditions, thus ensuring the strict separation of development and evaluation regimes. This experimental stratification facilitates robust assessment of model generalization and supports reproducible benchmarking.

Overall, the visual overview demonstrates that the dataset provides a structured, context-rich foundation for the evaluation of SWIPT energy harvesting and data demodulation, the benchmarking of machine learning algorithms for signal classification, and transparent assessment of generalizability across parameter regimes.

Mean harvested power versus transmitter-receiver distance has four gain settings in the SWIPT experiment, as shown in Figure 2. Each line represents a gain level, with shaded regions indicating the standard deviation across repeated measurements. Results illustrate systematic attenuation with increasing distance and demonstrate how higher transmitter gain mitigates signal loss, thus supporting the analysis of energy harvesting efficiency and experimental coverage.

Mean harvested power exhibits a monotonic decrease with increasing transmitter-receiver distance for all gain levels, demonstrating systematic attenuation consistent with physical signal propagation. Higher transmitter gain settings (100% and 90%) yield elevated power levels at all measured distances relative to lower gain settings (75% and 65%). Standard deviation error bands indicate greater measurement variability at shorter distances, particularly under high gain conditions, whereas variability is reduced at larger separations and lower gains. The overlap and separation among gain curves allow the assessment of the differential impact of transmitter power on energy harvesting performance, with higher gain mitigating losses attributable to increased path length. Experimental coverage is uniform across both distance and gain axes, enabling rigorous evaluation of signal loss and generalization across the tested SWIPT scenarios.

The distribution of harvested power measurements for each gain and distance scenario is visualized in Figure 3 by histogram and kernel density estimate. The panels illustrate how transmitter gain and separation distance affect received power to reveal shifts in central tendency, spread, and potential skew or multimodality across multisine wireless energy transfer trials.

The distribution of receiver power measurements varies systematically across panels representing different combinations of transmitter gain and distance. Panels with higher gain and shorter distance (e.g., Gain100_Distance10) exhibit higher central tendency and broader spread in harvested power, with values extending toward the upper range of the x-axis. As transmitter gain decreases and separation distance increases, the power distributions shift leftward, showing reduced mean and peak power, greater concentration near lower values, and diminished spread. Several panels display multimodal or skewed characteristics, particularly under intermediate gain/distance configurations, indicating the presence of subpopulations or hardware effects in signal detection. The kernel density estimates closely mirror the histogram shapes, highlighting the effect of experimental parameters on power variability and overall signal distribution. No measurement variable shows evidence of missing values or outlier-dominated tails within the visible axes’ limits, and all panels use consistent axis scales, facilitating the direct comparison of central tendency and spread across scenarios.

Distributions of receiver-side harvested power for each multisine tone class are shown in Figure 4 as violin and box plots across all transmitter gain and distance combinations. Power distributions shift systematically with tone count, highlighting clear class-dependent patterns suitable for machine-learning-based signal classification in SWIPT experiments.

Receiver-side harvested power distributions exhibit a monotonic decrease in median and spread as the number of multisine tones increases, with the highest power observed for signals with two tones and the lowest for signals with thirty-two tones. This class-dependent separation is consistent across all gain and distance configurations, including both high and low transmitter gain, and short and long distances. The power distributions for each tone class remain well-separated, with minimal overlap, especially at higher gain and shorter distance settings, supporting the discriminative suitability of harvested power as a feature for multisine signal classification. Additionally, the absolute value of harvested power decreases systematically as transmitter distance increases and gain decreases, yet class separation persists. The observed patterns indicate that the dataset provides balanced and reproducible distributions that motivate machine-learning-based classification of multisine tone index in SWIPT applications.

In Figure 5, the experimental partitioning of transmitter gain and distance conditions in the SWIPT multisine dataset illustrates the split between internal model development (A) and external generalization testing (B) scenarios to support transparent machine-learning benchmarking and reproducibility.

The partitioning of transmitter gain and distance conditions in the dataset is clearly demarcated by dataset group. Internal model development samples (A) are concentrated at shorter distances (10, 15, and 20 cm) with higher transmitter gain settings (100% and 75%), while external generalization test samples (B) are obtained at longer distances (25, 30, and 35 cm) with moderate-to-lower gain settings (90% and 65%). This spatial and power stratification supports the transparent benchmarking of model transferability across distinctly partitioned experimental regimes. No overlap is present between internal and external groups in the gain–distance parameter space, ensuring strict separation for reproducibility in machine learning evaluation.

A schematic overview of the experimental design and dataset partitioning is given in Figure 6 for the SWIPT multisine study. Each tile represents a unique combination of transmitter gain and spatial distance, with cell color encoding inclusion in model development (A) or external test (B). Variable labels identify measurement conditions, highlighting the modularity and transfer potential for benchmarking and cross-condition analyses.

The schematic tile plot illustrates the experimental matrix formed by transmitter gain and spatial distance, with each tile representing a distinct measurement variable. Model development samples (group A) are concentrated at higher transmitter gains (31 and 23.25 dB) and shorter distances (10, 15, and 20 cm), while external test samples (group B) correspond to lower gains (27.9 and 20.15 dB) and longer distances (25, 30, and 35 cm). All measurement variables are uniquely labeled, confirming full coverage across partition groups and supporting modular reuse for benchmarking and transfer learning analyses. Cell color coding provides a clear distinction between internal and external evaluation conditions, with no apparent overlap between partition assignment.

5 Discussion

5.1 Value and applications of the dataset

This dataset provides a structured collection of receiver-side RSSI/power measurements in the context of controlled USRP transmitter gain and distance experiments using multisine waveforms. By systematically varying transmitter gain (65%, 75%, 90%, and 100%) and separation distance (10 cm–35 cm) and annotating each sample with a class label indicating the number of transmitted tones, the dataset enables the analysis of signal propagation, attenuation, and classification under repeatable hardware configurations. The tabular organization and granular documentation of gain–distance scenarios facilitate the precise replication and extension of experimental protocols in the wireless communications research domain.

The dataset supports investigation into the discrimination and modeling of multisine waveform classes in the presence of physical channel variability, thus enabling research into signal detection, classification, and regression tasks associated with receiver-side power levels. Potential new research directions include benchmarking machine learning models for tone count classification in multisine communications, evaluating the effect of hardware gain and spatial separation on model generalization, and analyzing bias or drift introduced by varying experimental parameters. In addition, the explicit partitioning for model development and external validation permits the evaluation of algorithmic performance in both in- and out-of-distribution contexts.

Relative to existing publicly available datasets that focus on single-tone or narrowband signal characterization, this dataset captures multidimensional parameter spaces associated with transmitter gain, distance, and waveform complexity. It differs in explicitly providing a harmonized schema for gain–distance variation and integrates detailed metadata fields that support machine-actionable provenance and FAIR² compliance. As a result, it offers complementary support for the development, comparison, and benchmarking of algorithms in wireless physical-layer experiments, particularly where a systematic exploration of hardware and environmental parameters is required. Its metadata design further enables integration with other signal processing datasets through standardized variable naming and unit declarations.

5.2 Limitations of the dataset

Several limitations should be considered regarding the structure and composition of this dataset. The measurements are restricted to a discrete set of transmitter gain and -receiver distance scenarios, potentially limiting the applicability of the dataset to untested gain settings or separation distances. The dataset exclusively features power measurements collected with defined USRP transmitter gains (65%, 75%, 90%, and 100%) and fixed distances (from 10 cm to 35 cm), which may not reflect the full range of operational conditions encountered in other radio frequency or wireless communication studies.

The dataset does not include raw SDR/IQ recordings or time-domain waveform captures. Therefore, it is not intended for studies that require the reconstruction of the received baseband signal, channel estimation from IQ samples, or end-to-end deep learning directly from waveform samples. The dataset focuses on receiver-side RSSI/power readouts in milliwatts from the implemented energy harvesting receiver chain.

The measurements were collected in a controlled laboratory setting with fixed hardware placement for each distance/gain configuration. Environmental parameters such as temperature and humidity were monitored during the measurement campaign, since the Powercast energy harvester contains built-in temperature and humidity sensors. However, the purpose of this study was to focus on the energy harvesting performance through the received power. Hence, the dataset does not include environmental sensor logs. Future extensions may include temperature, humidity, and ambient RF spectrum scans.

The measurements were collected during a single experimental session/campaign, although the full acquisition process required time to complete because all gain, distance, and tone-count conditions were measured systematically. The released CSV file does not include timestamps or session identifiers. Thus, temporal variation during the acquisition session cannot be independently analyzed.

All receiver-side measurements were recorded using a single hardware and software configuration—HyperTerminal connected to the USRP system—which may introduce device- or platform-specific influences that are not characterized within the current metadata. No explicit batch or session variable is provided, so any temporal or environmental variation during data acquisition is not captured, which may affect generalizability. Each measurement condition (gain–distance combination) has an equal number of samples (750 per variable), but the data structure does not permit an analysis of repeated measures over time or across replicates from independent acquisition runs.

Class labeling for the target variable (Indicator, denoting multisine tone count) is categorical and fixed to five discrete levels, representing specific synthesized waveform scenarios. The dataset does not offer raw waveform data or details on the underlying signal processing pipeline, limiting its application for studies requiring access to the original or intermediate signal forms.

The diode-based rectenna introduces nonlinear RF-to-DC conversion behavior. This is part of the measured SWIPT receiver response, but it means that the reported values should not be interpreted as linear, independently calibrated incident RF power measurements. This limitation is particularly relevant at the lowest transmitter-gain setting, where receiver sensitivity, threshold effects, and finite readout resolution may have a larger relative effect.

The measurements are also specific to the fixed dipole-antenna geometry, orientation, and laboratory layout used in the experiment. Different antennas, polarization alignments, propagation environments, or rectenna placements may lead to different harvested power responses.

While the associated partitioning file clarifies intended use for model development and external testing, it does not capture potential covariates such as channel conditions, hardware calibration drift, or operator effects, which may introduce unexpected variance. Furthermore, the dataset does not contain negative controls or measurements with intentionally perturbed hardware settings, restricting its use for robust outlier detection or fault diagnosis studies.

5.3 Opportunities for future research

The dataset here presented is relevant for research on wireless signal processing, multimodal waveform recognition, and machine learning applications involving physical layer sensing or classification. By systematically varying transmitter gain and transmitter–receiver distance, the dataset provides structured observations suitable for algorithm benchmarking, data-driven modeling, and the comparative evaluation of RSSI-based receiver performance. The explicit encoding of physical parameters and the inclusion of a class indicator for transmitted multisine tones enable research focused on waveform identification, multi-condition classification, or sensitivity analysis to controlled hardware and propagation changes.

Future directions for dataset expansion could include increasing the diversity of environmental conditions—such as different ambient noise levels, obstacle placement, or antenna orientations—owing to the recognized significance of external factors in wireless experiments. The inclusion of additional transmitter hardware, wider gain and distance ranges, or other signal modulation schemes would further support generalizability studies and broader benchmarking scenarios. Providing raw complex baseband or time-domain signal captures alongside processed power readouts could facilitate research into feature learning, end-to-end deep learning, or unsupervised representation modeling. Supplementing metadata with detailed provenance information, hardware calibration records, and contextual sensor data (e.g., temperature and humidity) could strengthen reproducibility and external validation.

6 Conclusion

The dataset described herein establishes a transparent and reproducible foundation for experimental research on simultaneous wireless information and power transfer (SWIPT) in point-to-point radio frequency systems that employ multisine signals. The data resource is rigorously organized with field-level provenance, linking each measurement to its corresponding experimental protocol step, thus enabling traceability and facilitating both human and machine-actionable access.

Key variables—including independent control parameters such as transmitter gain, tone count, and spatial separation—are systematically varied according to a fully crossed experimental matrix. All receiver-side power measurements are annotated with explicit metadata and assigned unique variable names, while group membership, model development status, and intended analytical use are consistently labeled for each record. This arrangement supports robust partitioning into training/validation and external generalization subsets, directly addressing methodological demands for internal and external reproducibility in machine learning workflows.

The dataset’s formatting, schema annotation, and measurement documentation adhere to the FAIR² principles, encompassing not only findability, accessibility, interoperability, and reusability but also AI-readiness and responsible data governance. Extensive datasheet metadata accompanies the raw measurement tables and partition records, documenting units, value domains, statistical summary metrics, and field definitions. This ensures that end users can accurately discern the experimental scenarios represented by each measurement, reproduce pre-processing, and apply appropriate downstream machine learning or statistical analyses.

The release protocol further encompasses completeness and integrity checks, schema validation, and the computation of digital signatures for public dissemination, thus providing verifiable assurance regarding the dataset’s fitness for reuse in a range of scientific and engineering contexts. All elements of the resource are deposited in machine-readable formats, with sufficient metadata to support deployment in automated or semi-automated workflows.

By systematically recording parameter variation, documenting measurement provenance, and enforcing quality standards throughout acquisition and finalization, the dataset provides a reliable reference for the investigation of SWIPT signal classification and energy harvesting phenomena. Its alignment to community and AI interoperability standards maximizes the potential for secondary analysis, benchmarking, and methodological extension while maintaining transparency and responsible stewardship of experimental data.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. This document and the accompanying data package were prepared by the authors with assistance from the SENSCIENCE FAIR2Data Publishing platform (v0.11β), which uses generative AI technology.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Besma Smida, University of Illinois Chicago, United States

Reviewed by: Ahmad Bazzi, New York University Abu Dhabi, United Arab Emirates

Ajin R. Nair, Manipal Institute of Technology Bengaluru, India