Research Center of 6G Mobile Communications, School of Cyber Science and Engineering, Huazhong University of Science and Technology, Wuhan 430074, China
Artificial intelligence for the intelligent Internet of Everything (I-IoE) establishes a four-dimensional interconnection among people, data, processes, and things, providing possibilities for the next generation of industrial revolution. Against the ultra-high power consumption brought by massive connectivity, passive backscatter communication has emerged as a promising paradigm for the I-IoE. However, even with a well-designed communication module, integrating sensing components significantly multiplies the overall system complexity and cost in the context of the {0,1} modulation paradigm. In this paper, we introduce a [0,1] modulated backscatter architecture that achieves the seamless integration of passive sensing and communication. Unlike the conventional backscatter of reading, encoding, and reflecting, the proposed [0,1] modulated backscatter directly converts environmental physical quantities into continuous frequency-modulated square waves. Building on this concept, we propose an intelligent detection method based on graph dimensionality reduction. This method achieves low computational complexity at the receiver by leveraging a pre-measured dataset. We then present a hardware communication system that performs data collection, transmission, and demodulation, providing experimental validation for the proposed architecture. Furthermore, simulation results verify the feasibility of a continuous frequency division multiple-access method for large-scale tags. The experimental results demonstrate the superior performance of the proposed scheme. This work provides a potential breakthrough for achieving ultra-low-power integrated sensing and communication in sixth-generation wireless communication networks.
Tao Jiang, Miaoran Peng, Yu Zhang, Xin Zhu, Shuqi Tang.
[0,1] Modulated Backscatter with Lower-Power Integration of Sensing and Communication for I-IoE in 6G.
Engineering, 2026, 61 (6) : 354-364 DOI:10.1016/j.eng.2026.01.008
Artificial intelligence for the intelligent Internet of Everything (I-IoE) is widely recognized as a cutting-edge vision devised to digitize the physical world through information collection, transmission, and processing technologies [1], [2]. By establishing ubiquitous interconnections for thing-to-thing (T2T) and thing-to-person (T2P) communications, the I-IoE represents the third wave of the world’s information industry development, following computers and the Internet [3], [4], [5]. At present, the massive device connectivity and diverse requirements of Internet of Things (IoT) applications are creating major challenges in energy consumption and deployment costs [6], [7], [8], which will only increase with the development of the I-IoE.
Backscatter technology is a transformative communication paradigm known for its ultra-low power consumption and minimal hardware cost. It has emerged as one of the most competitive low-power I-IoE solutions in recent years [9], [10], [11], [12]. Unlike traditional communication devices that actively generate radio frequency (RF) signals, backscatter devices omit power-intensive components such as analog-to-digital converters (ADCs) and power amplifiers. Instead, backscatter tags transmit their own information by modulating ambient electromagnetic waves [13]. Specifically, a backscatter tag uses a microcontroller to switch between different load impedances, altering the amplitude and phase of the incident signal. Consequently, the tag completes a modulation process that converts a digital baseband signal to a carrier from the RF source [14]. Traditional backscatter communication achieves information transmission by reflecting received RF signals, which substantially reduces power consumption and extends battery life by approximately ten years. However, problems with deployment and maintenance persist because batteries eventually deplete. Furthermore, the need for the periodic replacement or redeployment of tags in applications such as implantable medical devices or in hazardous environments remains a significant challenge [15].
In recent years, researchers have developed prototypes to validate passive backscatter systems in fundamental I-IoE scenarios. For instance, Zhang et al. [16] designed and implemented the FreeRider, which achieves backscatter communication with a data rate of 60 kilobits per second (kbps) and a range of 42 m in an 802.11g/n wireless fidelity (WiFi) communication system. By leveraging existing commercial radio technologies (e.g., 802.11g/n WiFi, ZigBee, and Bluetooth), FreeRider enables backscatter communication without disrupting the original data transmission. Moreover, researchers have conducted in-depth investigations to address the issues of low data rate and limited transmission distance. Firstly, long-range radio (LoRa)-based backscatter communication was proposed to realize kilometer-scale passive transmission [17], [18]. In addition, a long-range backscatter prototype-verification system was designed and built based on adaptive coding, demonstrating a communication range of up to 2 km [19]. High-order modulation and high-frequency transmission were subsequently considered in order to increase the transmission rate. Peng et al. [20] utilized backscatter tags with multiple reflection coefficients to approximate orthogonal frequency-division multiplexing (OFDM) signals, achieving a data transmission rate of 21 megabits per second (Mbps). Finally, millimeter-wave backscatter communication systems operate over higher bandwidths, with passive transmission rates reaching up to gigabits per second based on the reflection principles of Van Atta arrays [21].
However, prior research has primarily focused on the passive communication link itself. In practice, the information bits originate from peripheral sensors, whose power consumption and cost are frequently overlooked. Depending on the type of sensors (temperature, humidity, voice, video, etc.), the power consumption of the devices ranges from several milliwatts to several hundred milliwatts. Moreover, the ADC converts these analog readings into digital (0-1) signals—a process that itself consumes hundreds of milliwatts. Finally, the sensor information must be converted into modulated signals. In particular, when processing audio and video information, low-power-consumption micro controller units (MCUs) and field-programmable gate arrays (FPGAs) fail to manage the computational complexity. The core issue is that sensors capture continuous analog variations, but conventional backscatter tags require this information to be digitized into discrete “0” and “1” symbols (e.g., using 2-frequency-shift keying (FSK)) for transmission. To accomplish this process, high-power-consumption ADC modules and computational procedures are unavoidable.
Building on Ref. [22], the present paper proposes a low-overhead integrated passive sensing and communication (IPSC) system that transforms conventional digital {0,1} modulation into a continuous analog {0,1} modulation scheme. This system directly converts environmental physical signals into continuous modulation signals, eliminating the tag’s need for sensor processing, ADC, synchronization, and complex data processing. Against this background, the main contributions of this paper are as follows:
• We have designed a hybrid analog-digital sensing and modulation scheme that converts impedance variations into frequency changes and outputs them as RF switch control signals. Compared with traditional continuous pulse-width modulation (PWM), this scheme achieves a tenfold improvement in transmission distance.
• To address the complexity of electromagnetic signals and channel variability, we have developed an intelligent detector based on a discrete dataset. This detector effectively reduces computational complexity at the receiver and enables real-time online demodulation.
• To correct for spectral distortion at the receiver caused by the electromagnetic environment, we adopt a light gradient boosting machine (LightGBM) model, which learns to map the received spectral features to the tag’s input voltage from a training dataset, ultimately improving the detection accuracy.
• The hardware protocol and communication system of the proposed IPSC architecture have been implemented and validated, and the experimental results demonstrate its advantages of low overhead and high reliability.
The remainder of this paper is organized as follows. We review the principle of conventional {0,1} modulation method in Section 2, and the overall framework of the IPSC system is described in Section 3. The intelligent detector and error-correction scheme are presented in 4 Discrete dataset-based intelligent detector, 5 Spectrum-based intelligent error correction, respectively. The implementation and evaluation of the system are analyzed in 6 Hardware implementation, 7 Evaluation. Finally, Section 8 concludes the paper.
2. {0,1} modulated backscatter communication
Existing backscatter communication typically adopts {0,1} modulation, in which tags generate discrete passive constellation symbols through frequency and amplitude variation. Taking on-off keying (OOK) modulation as an example, when a backscatter tag switches between two different impendence networks, the frequency shift Δf combines with the incident signal frequency f0 to generate second harmonics at f0±Δf [23]. The modulated signal reflected by the tag can be formulated as follows:
where B denotes the modulated signal of the backscatter tag, t is the time variable, j denotes the imaginary part, i is time domain integral variable, fc denotes the frequency offset, c(t) signifies the backscatter symbol, and Γ represents the reflection coefficient with
$ \Gamma=\frac{Z_{l}-Z_{0}}{Z_{l}+Z_{0}}$
where Zl signifies the normalized impedance, l denotes the load, and Z0 denotes the matching impedance (usually set as 50 Ω). The receiver identifies the tag’s information by demodulating the amplitude of the received signal at f0±Δf [24]. More specifically, the starting position of each frame is first recognized through frame synchronization. Next, the time-domain signal is converted into a frequency-domain signal via fast Fourier transform (FFT). Since it is difficult for the receiver to separately estimate the channels from the transmitter to the tag and from the tag to the receiver, joint estimation of the cascaded channel is typically calculated by means of pilot signals from both the transmitter and the tag. Finally, coherent detection is applied to demodulate the tag signal.
Fig. 1 illustrates the coherent detection results of an OOK sequence after receiver sampling. The receiver generates a 1 symbol when it detects a frequency shift and a 0 symbol when no shift occurs, based on the received signal amplitude. Furthermore, switching between different reflection coefficient combinations produces distinct reflected signal amplitudes. In addition, flipping the delay of the control signal alters the phase of the reflected signal. By combining different frequency, amplitude, and phase variations, multidimensional passive constellation points are formed to enhance the transmission rate and reliability.
3. The overall framework of low-power sensing and communication integration: From {0,1} to [0,1]
As shown in Fig. 2(a), integrating sensors with {0,1} modulated backscatter multiplies the power consumption and cost. Firstly, depending on the function, different sensor components from temperature to audio and video consume between 1 and 100 mW [25], [26]. Secondly, the continuous voltage variation provided by the sensors is converted into high and low levels through an ADC [27] and transmitted to the microcontroller [28], which consumes over 300 mW. The microcontroller then converts the input data into modulated symbols. This process exceeds the capacity of a typical microcontroller, especially when handling audio or video data. Moreover, to achieve timing synchronization between the tags and RF source signals, envelope detectors consuming more than 50 mW will be unavoidable [29].
To overcome the practical limitations of {0,1} modulation, Fig. 2(b) presents a low-power integration of sensing and communication architecture based on [0,1] modulation, where the analog data produced by a sensitive resistor is directly converted into continuous modulated signals and transmitted to the receiver. Compared with sensor units, the sensitive resistor is a passive component that can output changing environmental information through a voltage divider without requiring a power supply. To implement this approach, we replaced the sensor with a sensitive element. The impedance variation of this element is mapped directly to changes in the reflection coefficient’s magnitude and phase. In other words, the impedance of the sensing element varies with environmental physical quantities, which manifests as continuous phase changes on the Smith chart. These phase variations can be further converted into duty cycle, frequency, or phase shifts of the RF switch control signals, thereby achieving diversified [0,1] modulation. The overall power consumption of the [0,1] modulated tag is approximately 115 μW.
To elaborate further, environmental information (temperature, humidity, voice, etc.) is first converted into voltage variations through the sensitive impedance and voltage divider circuit. Subsequently, the voltage signal is integrated and compared with the square wave signal generated by a crystal oscillator, resulting in a square wave with continuously varying frequency. When the square wave enables the RF switch to toggle between the open and load states, a continuous frequency-shifting signal is superimposed onto the incident wave. The [0,1] modulated signal can be formulated as follows:
where Δfi denotes the frequency shift of the tag, which varies over time.
At the receiver, the environmental physical quantities are reconstructed by demodulating the continuously modulated excitation signals. To be precise, since the signal frequency remains constant across channels, the detected frequency shift $ \Delta \widehat{f}_{i}$ is computed by extracting the corresponding frequency from the maximum spectral spectrum obtained through FFT on each data frame. First, a frequency dataset is generated for the receiver by traversing the tag input voltages. During the data transmission, $ \Delta \widehat{f}_{i}$is obtained through a comparison between the detected frequency shift and the dataset. Consequently, $ \Delta \widehat{f}_{i}$can be demodulated by
Compared with conventional {0,1} modulation, the proposed method strikes a balance between transmission reliability and receiver computational complexity for a lower tag overhead. Unlike the fixed Δf in {0,1} modulation, the Δfi in [0,1] modulation continuously varies with environmental physical quantities, thereby generating analog modulation sequences. For demodulation, the proposed [0,1] scheme requires the receiver to distinguish continuous spectral changes, which shifts the computational burden from the tags to the receiver. For hardware architecture, the [0,1] tag eliminates the requirement for ADC components and microcontroller encoding processes, achieving lower power consumption and a cost-effective integrated communication-sensing solution.
4. Discrete dataset-based intelligent detector
Due to the continuous variation in modulation symbols, the [0,1] modulated backscatter requires a mapping relationship to be established between the environmental physical quantities and the received time-frequency characteristics during detection, which is formulated as follows:
where G represents the circuit loss, fmax denotes the maximum frequency shift, and fmin indicates the minimum frequency shift. fmax and fmin are determined by the hardware characteristics of a voltage-frequency converter and are generally on the order of megahertz. R0 stands for the reference value of the sensitive impedance, Ri represents the actual value of the sensitive impedance, Vref refers to the voltage reference value, and V0 is the input voltage of the circuit. However, in practice, directly demodulating Δf to obtain Ri via this ideal mapping is impractical because it ignores real-world factors such as channel effects, receiver characteristics, and circuit losses.
Therefore, it is necessary to establish a dataset through pre-measurements to create a discrete relationship between the environmental physical quantities and the received spectrum. Ultimately, after frame synchronization at the receiver, by dividing each frame of data and extracting the maximum spectrum value, the demodulation of [0,1] data can be achieved by comparing the results with the dataset using machine learning (ML) algorithms. This approach effectively shifts the computational complexity from the tag to the receiver.
To increase the detection accuracy while reducing the computational complexity at the receiver, an intelligent detection algorithm based on graph dimensionality reduction is proposed, as shown in Fig. 3. Firstly, the received signal (y) can be formulated as follows:
$y(n)=h_{\tau} s_{j}+\omega(n)$
where $h_{\tau}=h_{\mathrm{d}}+h_{\mathrm{f}} \Delta f_{i} h_{\mathrm{b}}$ stands for the transmitting signal from the tag; τ is the backscatter symbol index; ωn represents the noise; n is frequency domain variable; sj is the transmitting signal of the transmitter; j is the index of modulation symbol; and hd, hf, and hb signify the channel coefficient from the transmitter to the receiver, the transmitter to the tag, and the tag to the receiver, respectively.
where αr and βr with r=1,2,...,R are weight vectors, R denotes the number of the lines of the matrix, and Y denotes the received signal matrix. The demodulation matrix $ \widehat{C}$ can be formulated after offline training, model pruning, and online reasoning methods. Fig. 3 presents the dataset construction and actual detection results. For an input voltage range of 0.5–1.5 V, the tag generates a frequency shift between 1.0 and 1.7 MHz. After accounting for channel superposition and hardware effects, the demodulated frequency range spans 0.5–2.5 MHz.
5. Spectrum-based intelligent error correction
Fig. 4 shows three distinct scenarios observed in the received spectrum. When a single peak is present, the receiver can accurately identify and demodulate the frequency shift. When the received energy falls below the recognition threshold, multiple faint peaks emerge, rendering demodulation impossible. In cases where two peaks with similar amplitudes appear, the receiver may incorrectly interpret either as the correct result, leading to partial demodulation errors. To address this issue, we develop an error-correction algorithm that trains on spectrum characteristics to reduce the probability of misinterpretation.
We focused on estimating voltage values from spectrum data using the LightGBM regression model. Our preliminary analysis revealed that errors primarily occurred for voltage values at the extremes of the range and during periods of high spectral fluctuation. To overcome these issues, we designed an error-correction mechanism that relies on enhanced feature extraction and robust training strategies.
5.1. Feature engineering
We extracted three groups of features to enable the model to capture both spectral variability and temporal dynamics.
• Spectral statistical features. Using FFT, we derived frequency-domain descriptors including the spectral centroid, bandwidth, flatness, and entropy. These features capture shifts in spectral distribution and help correct the bias introduced by frequency misinterpretation.
• Time-domain statistical features. We calculated the mean, variance, skewness, kurtosis, and root mean square (RMS) to characterize the signal dispersion and asymmetry. These descriptors provide complementary information on the shape of the signal, which reduces systematic underestimation or overestimation in certain intervals.
• Sliding window features. We adopted a rolling window to compute local mean and standard deviation. These features help the model track spectral variations over time. They are particularly effective in correcting errors when short-term spectral fluctuations affect the regression.
5.2. Model training and error correction
We configured LightGBM with robust hyperparameters to reduce the impact of outliers through adaptive balancing of the mean squared error and mean absolute error (MAE). Stratified sampling was further applied to ensure that edge voltage intervals were sufficiently represented, thereby directly correcting for the observed underfitting in these regions.
Thanks to this combination of targeted feature design and robust boosting strategies, the proposed pipeline effectively addresses the key sources of misinterpretation and improves the accuracy and stability of voltage prediction.
6. Hardware implementation
In this section, the hardware prototype design of the proposed [0,1] modulated backscatter tag is presented.
6.1. Tag design
Fig. 5 shows the hardware prototype of our frequency-based [0,1] modulation backscatter tag. Its key components are a sensing module, a modulation module, an FPGA for frequency band selection, and an RF front end. The modulation module consisted of a 6 MHz crystal oscillator and a voltage-to-frequency converter (AD7441). According to the characteristics of the AD7441, fmin=1MHz and fmax=6MHz. This square wave drives an RF switch (ADRF5020) to toggle between two impedance networks, which alters the spectral characteristics of the incident signal.
Additionally, an anti-collision mechanism was demonstrated by superimposing the square wave generated by the FPGA (AGLN060V2) onto the original signal. This FPGA operation consumed 32 μW. The FPGA module of each tag is designed for frequency selection, which can also be applied to isolate the frequency band of each tag. This flexible frequency band adjustment helps avoid interference between the tag and other IoT devices. Moreover, in multi-tag access cases, the FPGA superimposes a square wave of the same frequency on the voltage-to-frequency converted square wave, enabling different tags to perform frequency shifting in different frequency bands.
6.2. Experimental setup
We evaluated the proposed scheme in four different scenarios, as shown in Fig. 6. A [0,1] modulated backscatter tag was applied in each scenario. The transmitter and receiver were both universal software radio peripheral (universal software radio peripheral (USRP) 2954R), with the tag operating in the 3.2 GHz frequency band.
• Indoor small room. As shown in Fig. 6(a), we applied the [0,1] modulated tag in a small room with an area of 6.0m×4.3m to measure the transmission performance under strong multipath interference. The transmitter-tag distance was 0.5 m, and the tag-receiver distance ranged from 1 to 5 m. The transmitting effective isotropic radiated power (EIRP) of the transmitter was 23 dBm, and the sampling frequency was 3 Mbps.
• Indoor conference room.Fig. 6(a) also shows a 12m×9m indoor conference room scenario, where the transmitter-tag distance was set to 2 m, and the distance between the tag and the receiver was 2-12 m. The EIRP of the transmitter was 15 dBm.
• Corridor. As shown in Fig. 6(b), we measured the tag transmission performance at distances of 0.5 and 1.0 m between the transmitter and the tag, with the tag-receiver distance ranging from 2 to 30 m. The transmitter’s EIRP was set at 23 dBm. Due to corridor length constraints, a 20 dBm attenuator was applied to the system.
• Outdoor.Fig. 6(c) portrays an outdoor measurement scenario and the layout of the devices. The distance between the transmitter and the tag was 0.5 m, the modulation rate of the tag was 20 kbps, and the distance between the tag and the receiver was 6-80 m.
7. Evaluation
In this section, we verify and analyze the practice transmission performance of the proposed scheme through data error rate (DER), throughput, and received signal strength indicator (RSSI) calculations. We also verify the scalability of the proposed tag through a simulation experiment.
7.1. Code book analysis
First, with the transmitter-tag distance fixed at 0.5 m and the tag-receiver distance at 2 m, we swept the tag input voltage in 0.1 V increments. The receiver captured 200 frames of the signal at each voltage to build the training set and acquire the mapping code book. Table 1 presents the code book of different scenarios. Due to the hardware design, the relationship between the tag’s input voltage and the receiver’s detected frequency was piecewise monotonic but overall nonlinear. To further reduce the probability of detection errors, a voltage divider was utilized to control the input voltage within the range of 1.3-2.3 V. Consequently, the proposed solution demonstrated high stability across various environments including indoor, corridor, and outdoor conditions, while maintaining strong scalability.
7.2. Intelligent detector analysis
To thoroughly evaluate the proposed LightGBM model, we conducted comparisons with three categories of models: a simple linear baseline, a conventional ensemble method, and boosting-based models from the same family. This selection ensured that the evaluation would span from the most basic linear assumption to bagging-based ensembles, and finally to gradient boosting approaches directly comparable with LightGBM.
• Simple baseline (linear regression). A linear regression model was chosen as the simplest baseline, since it assumes a direct linear mapping between the spectrum and voltage. The results showed limited predictive power (MAE = 0.3425, root mean square error (RMSE) = 0.3923, and R2 = 0.1245 denotes the variance), confirming a highly nonlinear relationship between the spectrum and voltage that cannot be effectively captured by such models.
• Conventional ensemble model (random forest). We implemented a random forest model to test whether bagging-based ensembles could improve performance. The performance improved over the linear baseline, with MAE = 0.2265, RMSE = 0.3060, and R2 = 0.4675, but remained suboptimal. This result demonstrated that, while bagging reduced variance, it lacked the boosting mechanism required to fully exploit the feature interactions, resulting in weaker generalization ability.
• Models in the same family (CatBoostvsLightGBM). Within the family of boosting-based models, CatBoost was also evaluated alongside LightGBM. The results showed that CatBoost achieved substantial improvements over the linear regression and random forest models, with MAE = 0.1078, RMSE = 0.1510, and R2 = 0.8703. However, LightGBM performed better than CatBoost; it achieved the best overall results (MAE = 0.0883, RMSE = 0.1282, and R2 = 0.9065), demonstrating superior capability in capturing both time-domain variations and frequency-domain spectral characteristics.
As summarized in Fig. 7, these experimental results demonstrate that LightGBM consistently provides the best balance between accuracy and generalization, establishing it as the most suitable regression model for this study.
Building on these comparative results, we further examined why LightGBM achieved such a performance by analyzing the contribution of individual features. In this study, after a comprehensive evaluation and redundancy analysis, the final feature set used for LightGBM training included frequency-domain statistics, time-domain descriptors, and sliding-window descriptors. The effectiveness of this feature selection was further examined through an SHapley additive exPlanations (SHAP) analysis, as illustrated in Fig. 8. The analysis revealed that the rolling standard deviation, signal variance, and signal kurtosis were the most influential features. The rolling standard deviation captured local fluctuations of spectral energy within short windows, enabling the model to respond sensitively to transient changes; the signal variance reflected the overall spread of the spectral distribution, which provided a measure of long-term variability; and the signal kurtosis emphasized the presence of sharp peaks or heavy-tailed distributions, which were often associated with unstable or extreme voltage states. In addition, frequency-domain statistics such as the spectral flatness and spectral bandwidth made a significant difference by quantifying how uniformly energy was distributed across frequencies and how concentrated it was around the spectral centroid. These results confirmed that accurate voltage prediction emerged from a joint characterization of temporal dynamics and spectral structure.
7.3. Indoor performance
Fig. 9, Fig. 10 present the experimental results of the proposed [0,1] modulated backscatter tag in the indoor scenarios. In small rooms, due to severe multipath interference, the transmission throughput dropped to 15.6 kbps when the tag-receiver distance reached 5 m. In the meeting room, the DER performance of the tag was maintained at 10-3 while keeping the throughput at 19.95 kbps.
7.4. Corridor performance
Fig. 11 presents the throughput performance of the proposed [0,1] modulation scheme in the corridor scenario, which was jointly determined by the data error probability and the total transmitted data. The transmission performance was verified at transmitter-tag distances of 0.5 and 1.0 m, respectively. When the receiver-tag spacing reached 30 m, the transmission throughput exceeded 18.2 kbps.
Next, we investigated the impact of interference signals on the detection performance at the receiver. First, we identified the interference signal with the greatest effect on the detection of the received signal. Fig. 12 shows how the tag’s DER was affected by the power level of an interference signal. The distances from the tag to the receiver and from the interference source to the receiver were both set to 2 m. When the interference signal was below 12 dB, the DER remained at 0. When the interference signal exceeded 15 dB, the DER rose above 0.4, as the tag signal became buried in the interference.
7.5. Outdoor performance
Fig. 13 displays the transmission performance of the tag under outdoor conditions. When the tag-receiver distance moved from 12 to 80 m, the transmitted EIRP was 23 dBm, and the received RSSI was approximately 50.7 dBm. Within a 70 m range, the transmission DER of the tag was less than 0.2, with a throughput exceeding 16 kbps.
7.6. Multi-tag access simulation
The simplified circuit structure of [0,1] modulated backscatter tags prevents them from exchanging pilot signals with the receiver to obtain channel state information. To avoid introducing high-power-consumption envelope detectors while considering the frequency-domain characteristics of the designed modulation method, we propose a broadband continuous frequency-division multiple access (cFDMA) scheme, as illustrated in Fig. 14.
A communication system with K [0,1] modulated backscatter tags was adopted. First, the transmitter sends the excitation signal at a frequency of f0. The radiation frequency fk of the kth backscatter tag (k = 1, 2, …, K) is formulated as follows:
where ϕk denotes the phase shift of the kth backscatter tag. The amplitude of the second harmonic is much lower than the first harmonic and can be ignored at the receiver. The channel from the transmitter to the kth backscatter tag (gf,k) and the channel from the kth backscatter tag to the receiver (gb,k) can be characterized by
respectively, where df,k signifies the distance from the transmitter to the kth backscatter tag, and db,k represents the distance from the kth backscatter tag to the receiver. The information carried by each tag is demodulated separately after the frequency band division of the received signal. Consequently, the received signal can be written as follows:
where YFDMA is the received signal of K backscatter tags. N(t) denotes Gaussian white noise matrix. Since the signal of each tag is demodulated separately, the intelligent detector proposed in Section 4 is extended to a multi-tag case with a complexity of O(KN).
The commercial RF identification (RFID) access protocol is applied as the baseline. First, the reader sends query commands to surrounding tags, and each tag randomly selects a time slot number from the available options. Then, the tag with slot number “0” generates a 16-bit random bit sequence and reflects it back to the reader. If two tags independently select the same time slot, a collision occurs and both tags fail to access the channel.
Fig. 15 presents the simulation results of the proposed access method. As the number of tags try to access increases, the actual number of access tag in the commercial RFID show a declining trend due to increased collision occurrences. In contrast, the proposed scheme shows a linear increase in both the number of attempted tags and the number of successfully accessed tags.
7.7. Power consumption analysis
The power consumption of the [0,1] modulated backscatter tag primarily originates from three components: the voltage-frequency converter, the RF switch, and the microcontroller. The microcontroller contains two output ports and one input port, with a total power consumption of 18.15 μW. Additionally, the RF switch consumes 264 μW, while the voltage regulator has a static power consumption of 115.5 μW. With an integrated circuit design that reduces the control voltages, the total circuit power consumption can be reduced to less than 100 μW.
8. Conclusion and future work
This paper proposed an IPSC architecture that transits from digital {0,1} modulation to analog [0,1] modulation, featuring a complete hardware implementation and experimental validation. Under the proposed [0,1] modulation algorithm, a backscatter tag achieved a throughput performance exceeding 18.2 kbps within a 30 m range. Additionally, we designed an intelligent detection algorithm based on a discrete dataset to adapt to varying transmission environments and dynamic channels. Finally, we developed a broadband cFDMA method and validated the access performance for over 150 tags through simulation experiments. Future work is summarized as follows:
(1) Voice transmission. Unlike temperature, humidity, or pressure sensing, microphones convert sound frequency into voltage output, which demands high transmission rates. Achieving long-distance, highly reliable modulation and demodulation for such variable audio signals is therefore an important challenge.
(2) Application-specific integrated circuit (ASIC) design. ASIC design is necessary to reduce the power consumption and cost of individual tags in the future. The tag size can be reduced to below 1 cm2 through ASIC, which will enable broader application and promotion of the design.
(3) Diverse application scenarios. The beneficial characteristics and integrated sensing-communication capability of our IPSC system give it considerable potential for applications such as environmental monitoring, disaster warning, and logistics tracking.
CHANGE HISTORY
A Correction to this paper has been published: https://doi.org
The Retraction Note was updated to reflect reception of statements of disagreement from all authors.
This article has been retracted.Please see the Retraction Notice for more detail: https://doi.org
JagatheesaperumalSK, PhamQV, RubyR, YangZ, XuC, ZhangZ. Explainable AI over the Internet of Things (IoT): overview, state-of-the-art and future directions. IEEE Open J Commun Soc2022; 3:2106-36.
ZhangX, ZhuJW, CuiTJ. An ultracompact spoof surface plasmon sensing system for adaptive and accurate detection of gas using a smartphone. Engineering2024; 35:86-94.
[4]
ShahSHA, KoundalD, SaiV, RaniS. Guest editorial: special section on 5G edge computing-enabled Internet of medical things. IEEE Trans Industr Inform2022; 18(12):8860-3.
[5]
YangT, YiX, LuS, JohanssonKH, ChaiT. Intelligent manufacturing for the process industry driven by industrial artificial intelligence. Engineering2021; 7 (9):1224-30.
[6]
International Data Corporation. Worldwide Internet of Things spending guide. Report. Singapore: International Data Corporation; 2025.
[7]
SinhaS. State of IoT summer 2024. Report. Hamburg: IoT Analytics GmbH; 2024.
[8]
DuhovnikovS, BaltaciA, GeraD, SchupkeDA. Power consumption analysis of NB-IoT technology for low-power aircraft applications. Proceedings of the IEEE 5th World Forum on Internet of Things; In: Proceedings of the IEEE 5th World Forum on Internet of Things; 2019 Apr 15-18; Limerick, Ireland. New York City: IEEE; 2019. p. 719-23.
[9]
LiuV, ParksA, TallaV, GollakotaS, WetherallD, SmithJR, et al. Ambient backscatter:wireless communication out of thin air. In:Proceedings of the Association for Computing Machinery’s Special Interest Group on Data Communication; 2013 Aug 12-16; Hong Kong, China. New York City: Association for Computing Machinery (ACM); 2013. p. 39-50.
[10]
WanX, XiaoC, HuangH, XiaoQ, XuW, LiY, et al. Joint modulations of electromagnetic waves and digital signals on a single metasurface platform to reach programmable wireless communications. Engineering2022; 8:86-95.
[11]
LiuC, ZhangR, LiY, XuF, TaD, WangW. An ultrasonic backscatter instrument for cancellous bone evaluation in neonates. Engineering2015; 1(3):336-43.
[12]
JiangT, ZhangY, MaW, PengM, PengY, FengM, et al. Backscatter communication meets practical battery-free Internet of Things: a survey and outlook. IEEE Commun Surv Tutor2023; 25(3):2021-51.
PengM, ZhangY, XiaoL, ZhouJ, LiuY, JiangT. Design and implementation of sector-based beam alignment for backscatter communications. IEEE Trans Microw Theory Tech2025; 73(9):6892-904.
[15]
HeX, CaiJ, LiuM, NiX, LiuW, GuoH, et al. Multifunctional, wearable, and wireless sensing system via thermoelectric fabrics. Engineering2024; 35:158-67.
[16]
ZhangP, JosephsonC, BharadiaD, KattiS. FreeRider:backscatter communication using commodity radios. In:Proceedings of the 13th International Conference on Emerging Networking Experiments and Technologies; 2017 Dec 12-15; Seoul, Republic of Korea. New York City: Association for Computing Machinery; 2017. p. 389-401.
JiangJ, XuZ, DangF, WangJ. Long-range ambient LoRa backscatter with parallel decoding. In: Proceedings of the 27th Annual International Conference on Mobile Computing and Networking; 2021 Oct 25-29; New Orleans, LA, USA. New York City: Association for Computing Machinery; 2021. p. 684-96.
[19]
SongG, WangW, YangH, ZhangD, GaoP, JiangT. Exploiting channel polarization for reliable wide-area backscatter networks. IEEE Trans Mobile Comput2022; 21(12):4338-51.
[20]
PengY, HeS, ZhangY, XiaoL, JiangT. Towards software-defined backscatter modulation via signal emulation. IEEE Trans Wirel Commun2024; 23 (10):14836-47.
[21]
MazaheriMH, ChenA, AbariO. mmTag:a millimeter wave backscatter network. In:Proceedings of the Association for Computing Machinery’s Special Interest Group on Data Communication; 2021 Aug 23-27; online. New York City: Association for Computing Machinery; 2021. p. 463-74.
[22]
JiangT. Constructing air-interface links for mobile communications: from {0, 1} to [0,1]. Engineering2025; 46:16-22.
[23]
GuoX, ShangguanL, HeY, ZhangJ, JiangH, SiddiqiAA, et al. Efficient ambient LoRa backscatter with on-off keying modulation. IEEE/ACM Trans Netw2022; 30(2):641-54.
[24]
ThomasSJ, WheelerE, TeizerJ, ReynoldsMS. Quadrature amplitude modulated backscatter in passive and semipassive UHF RFID systems. IEEE Trans Microw Theory Tech2012; 60(4):1175-82.
[25]
Analog Devices, Inc. (ADI). DS60 analog temperature sensor[Internet]. Wilmington: Analog Devices,Inc.; [cited 2025 Sep 14]. Available from: https://www.analog.com/cn/products/ds60.html.
[26]
Melexis. MLX90640-D55 thermal camera [Internet]. Shenzhen: Waveshare Wiki; [cited 2025 Sep 14]. Available from: https://www.waveshare.net/wiki/MLX90640-D55_Thermal_Camera.Chinese.
[27]
Microchip. MCP4725 12-bit single output DAC w/EEPROM and I2CTM [Internet]. Shanghai: Microchip Technology Inc.; [cited 2025 Sep 14]. Available from: https://www.microchip.com/en-us/product/MCP4725.
[28]
Microchip. AGLN060V2 ultra low density FPGAs [Internet]. Shanghai: Microchip Technology Inc.; [cited 2025 Sep 14]. Available from: https://www.microchip.com/en-us/product/AGLN060V2.
[29]
Analog Devices, Inc. (ADI). LTC5536 600 MHz to 7 GHz precision RF detector with fast comparator output[Internet]. Wilmington: Analog Devices,Inc.; [cited 2025 Sep 14]. Available from: https://www.analog.com/cn/products/ltc5536.html.