Spectrum Learning-Aided Reconfigurable Intelligent Surfaces for ‘Green’ 6G Networks

Video analytics relying on AI becomes an indispensable technique in support of autonomous driving, where a large amount of information generated by diverse onboard sensors (such as cameras, LiDAR sensors and radar sensors) has to be processed in near-real-time for collision avoidance in extremely complex surroundings. However, promptly exploiting a deluge of delay-sensitive sensors’ data requires substantial onboard computation resources and incurs high energy consumption [6]. To maintain safety for self-driving, RISs usher in a paradigm shift from connected-vehicles to pervasive connected-intelligence by enhancing the link quality of cellular vehicular-to-everything (V2X) communications in complex urban environments at a low power.

In recent years, UAVs, also commonly known as drones, have been extensively used for civilian, commercial, and military services in combination with AI [7]. However, the provision of sufficient energy for propulsion is a challenge, which may be mitigated by joint energy and data networking and laser charging [8]. In this context, RISs show significant promise in terms of reducing the communications-related power of UAVs.

At the time of writing, MEC is considered as a promising technique for offloading the processing of data to edge servers having much higher computing capabilities, thus enabling low-latency joint processing and communications. As a result, the AI processing is relocated from the central cloud to the network edge [9]. This implies that a large amount of data has to be transferred from the end devices to the edge server via band-limited uplink channels, which may become the bottleneck in MEC. In this context, RISs play a crucial role in mitigating the propagation-induced impairments with the aid of signal reflection.

The smart industry, or “Industry 4.0”, integrating AI and the industrial IoT, paves the way for intelligent production by interconnecting isolated industrial assets [10]. The wireless links are typically of non-line-of-sight (NLOS) nature, thereby leading to vulnerable links in dynamic environments. To circumvent this challenge, RISs improve the propagation environment by constructing signal reflections, thereby enhancing the robustness of wireless communications.

We collected RF traces by building a universal software radio peripheral (USRP2) based testbed, which is connected via the Gigabit Ethernet to a host computer, as illustrated in Fig.  2. The RF trace collection relies on a transmit unit and a receive unit that communicate over a $1$ MHz channel at a carrier frequency of $2.4$ GHz. The traces are collected over the signal-to-noise ratio (SNR) range spanning from $0$ to $20$ dB. Since each USRP2 unit has one transmit antenna, multiple interfering signals are obtained by letting multiple USRP2 units transmit at the same time. In particular, the transmit unit includes a transmitter host computer performing the baseband signal processing and a transmitter USRP2 for up-conversion, digital-to-analog (D/A) conversion, and wireless transmission. Upon receiving the RF signal via the wireless channel, the receiver USRP2 first processes the signals ahead from the radio interface and then performs A/D conversion and down-conversion. Finally, the required baseband processing is carried out by the host computer, where the Inphase (I) and Quadrature (Q) sequences are stored for a wide range of SNRs to account for different interference scenarios. The acquired I/Q samples capturing the features of the incident RF signal are used for training the CNN model.

In the proposed framework, spectrum learning is considered as a classification task, which is fulfilled via an appropriately trained CNN model. Specifically, based on the acquired RF traces that amount to approximately one billion I and Q samples, the CNN is trained offline based on a GPU cluster (e.g., the NVIDIA Tesla P100 GPU computing processor illustrated in Fig. 2) by using the Adam algorithm that utilizes the cross-entropy as the loss function. The CNN model trained consists of four layers: two convolutional layers that use the rectified linear unit activation function, two dense fully-connected layers and the output layer that uses the softmax activation function.

For the spectrum learning module, we can directly infer the spectrum access information related to the desired and interfering users by extracting the resultant I/Q samples from a copy of the impinging signals and feeding them into the appropriately trained CNN that is deployed at the RIS controller. Specifically, as soon as the RF signal arrives at the RIS elements, it undergoes A/D conversion and frequency down-conversion. Then, baseband processing is executed and the extracted I/Q sequence is fed into the CNN model for spectrum access inference.

The ON-OFF RIS control module is trained for avoiding undesired reflections by dynamically setting the binary ON or OFF status of the RIS elements. In particular, we assume that both the CSI and phase shifts of the RIS can be perfectly estimated and optimized at the BS, which feeds this information back to the RIS controller via a dedicated control channel³³3Channel estimation and phase shifts optimization of RIS have been widely investigated, hence they are not the main focus of this paper.. By disabling and enabling the RIS (i.e., RIS is turned OFF and ON), the SINR that is achieved at the BS can be calculated respectively. If the calculated SINR with disabling the RIS is no worse than the corresponding SINR when the RIS is enabled, the entire RIS is turned OFF. Otherwise, the entire RIS is turned ON.

The correct spectrum inference probability of the trained CNN model is demonstrated by considering the experimental scenario of Fig. 2 supporting $U_{D}$ contaminated by $U_{I}$. During the CNN model training, the number of time series I/Q samples is set to $32$, $128$, and $512$. The experimental results show that the correct inference probability obtained via the CNN model for the Class-2 (“$U_{D}$”), the Class-3 (“$U_{I}$”) and the Class-4 (“$U_{D}\!+\!U_{I}$”) is around $97\%$. Notably, the trained CNN model attains a perfect accuracy for the Class-1 (“Idle”) prediction due to the pattern distinguishable from the rest of the classes. The relatively high inference probability of the trained CNN model results in typically feeding an accurate spectrum identification estimate to the ON-OFF RIS control module.

In the considered scenario, there is one desired user ($U_{I}$), one interfering user ($U_{D}$), and $K$ RISs. We assume that the RISs are equally spaced by $5\ m$ in the vertical direction, each having $256$ elements. The amplitude reflection coefficient is $1$ and the impact of interference is inversely proportional to $\theta$. As illustrated in Fig. 3, the RIS₁-$U_{D}$, RIS₁-$U_{I}$, and RIS₁-BS distances are $60\ m$, $10\ m$, and $80\ m$, respectively. Furthermore, the angle of incidence between the BS and $U_{D}$ at the RIS is $150\rm{{}^{o}}$, while between $U_{D}$ and $U_{I}$ is $\theta\in[0,150\rm{{}^{o}}]$. To demonstrate the energy-efficiency of the proposed SL-aided ON-OFF RIS control scheme, we evaluate a pair of benchmark schemes: 1) RIS always ON, and 2) RIS always OFF, where the transmission power of $U_{D}$ is $23$ dBm and the noise power is -$94$ dBm.

The achievable SINR (in dB) versus the angle of incidence $\theta$ is depicted in Fig. 4, where a single RIS is considered, and the transmit power $P_{I}$ of the interfering user is $10$ dBm and $15$ dBm. We observe from Fig. 4 that the benchmark schemes exhibit an opposite trend as $\theta$ increases. In particular, as $\theta$ increases, the interference contaminating the desired receiver through a reflection from the RIS is reduced. Therefore, it becomes more likely that the RIS is ON, thereby increasing the achievable SINR of the conventional ‘RIS always ON’ scheme. By contrast, the achievable SINR of the ‘RIS always OFF’ scheme reduces with $\theta$ due to the reduction of the distance between $U_{I}$ and BS. In this case, the interference reflected by the RIS becomes more pronounced. It can be observed that the SINR obtained by the proposed SL-aided RIS control scheme is consistently higher than that of the two benchmark schemes, which is the benefit of intelligently controlling the status of the RIS according to the dynamically fluctuating environment. The improvement of SINR via the proposed scheme is at the expense of increasing the RIS controller’s complexity, i.e., the complexity introduced by the CNN online inference and the ON-OFF RIS state configuration. However, the associated complexity only grows linearly with the number of RISs, which is acceptable in practice.

The impact of the number of RISs ($K$) on the achievable SINR is evaluated in Fig. 5, where we set $P_{I}=10$ dBm and $\theta$ is randomly selected from $[30\rm{{}^{o}},120\rm{{}^{o}}]$. Intuitively, we observe that the SINR achieved by the ‘RIS always OFF’ benchmark scheme remains unchanged, while that of the other two schemes increases with the number of RISs. Moreover, the SINR obtained by the proposed SL-aided RIS control scheme is significantly higher than that of the two benchmark schemes. In particular, when we use the same transmission power, compared to the ‘RIS always OFF’ and ‘RIS always ON’ case studies, our proposed solution improves the received SINR from about $4.8$ dB and $2.9$ dB to $6.3$ dB for $K=5$. This demonstrates the potential of the proposed SL-aided RIS framework in terms of improving the energy efficiency of future 6G networks.