Integrated Sensing and Communications
1. Clutter suppression for ISAC systems
Integrated sensing and communications (ISAC) is a key technology for B5G/6G wireless systems, as it combines communication and sensing in one system, saving space, energy, and spectrum. A major challenge in ISAC is dealing with unwanted signal reflections, known as clutter, from objects such as buildings and trees, which interfere with detecting targets of interest. Therefore, an important area of research for future ISAC systems is clutter suppression to enhance target detection. The existing works on clutter suppression were proposed by assuming that the Doppler frequency of static clutter is zero. However, these algorithms are not applicable to the mobile ISAC systems, where the clutter has a non-negligible Doppler shift due to the mobility of the ISAC transceiver.
To enhance target sensing performance in mobile ISAC applications, such as simultaneous localization and mapping and vehicle-to-everything (V2X) communication, this paper proposes a space-time adaptive processing for ISAC (ISAC-STAP) algorithm. The proposed algorithm suppresses clutter in the target's range by using the neighboring range cells, and the target's channel parameters are estimated in the mobile ISAC systems. Fig. 1 highlights the effectiveness of our algorithm. In Fig. 1a, without clutter suppression, it’s nearly impossible to distinguish the target. The benchmark algorithm by H. Luo suppresses the signal at zero velocity, while the clutter remains as shown in Fig. 1b. The proposed ISAC-STAP algorithm accurately removes clutter, leaving only the target signal as shown in Fig. 1c. This improvement demonstrates the robustness of the proposed algorithm in suppressing the clutter caused by the mobility of the ISAC transceiver.
Fig. 1. Normalized power spectrum [dB] at the range cell of target 1
[Related publication]
-
N. Lee, H. Park, H. Kim, K. Jung, and S. Kim, “ISAC-STAP: Space-time adaptive processing for ISAC systems,” in Proc. 2024 IEEE Glob. Commun. Conf. Workshops (GLOBECOM Workshops), Cape Town, South Africa, Dec. 2024.
2. Beam Tracking
In mmWave mobile communication systems, accurate beam tracking has been essential due to the highly directional signals and user mobility. Traditional Bayesian methods, such as Kalman and particle filters, have been limited by model mismatch and high complexity under nonlinear mobility or weak multipath. To overcome these issues, we have developed a beam tracking algorithm based on Deep Q-Network (DQN), a model-free reinforcement learning approach that learns online directly from received signals. Unlike supervised methods requiring labeled data and predefined models, our DQN-based approach has operated without prior knowledge of the channel or user trajectory. We have defined the state using current and past signals to capture angular dynamics, and constrained the action space based on realistic mobility-induced variations. A reward function encouraging strong signal strength has enabled stable learning. By focusing on angle-of-arrival tracking, our method has achieved robust and efficient performance, outperforming traditional filters and discrete Q-learning in both accuracy and adaptability.
Fig. 2. one connected MS is communicating with a BS over the mmWave channel. MS is equipped with single antenna element, and BS is equipped with N antenna elements.
Fig. 3. The overall structure of DQN
Accurate beam tracking has been critical in mobile mmWave and ultra-massive MIMO (UM-MIMO) systems due to narrow beams and user mobility. As systems have transitioned from far-field to near-field (e.g., terahertz), both angle and distance have become resolvable, enabling position-level tracking and allowing more precise beam control than conventional far-field methods. To leverage this capability, we have developed a deep reinforcement learning-based beam tracking algorithm using a Deep Q-Network (DQN) Our model-free approach has learned optimal beam steering directly from received signals without requiring prior knowledge of the channel or user trajectory. By using beamforming gain as the reward and modeling the state as a temporal sequence of received signals, the algorithm has adapted online to dynamic user behavior. Tailored for near-field beam control, the proposed method provides an intelligent and scalable solution for next-generation mmWave and terahertz systems.
Fig. 4. NF beam tracking procedure in thetime domain
Fig. 5. Combiner and action space for NF beam tracking
[Related publication]
3. Sensing and Communication time optimization
Integrated sensing and communication (ISAC) is a new paradigm in which previously competing for sensing and communication tasks can be jointly optimized. In future 6G systems, higher-frequency(mmWave to THz), wider bandwidths, and denser large-scale antenna arrays provide an opportunity to integrate wireless signal sensing and communication into a single system and the entire communication system to act as a sensor. Context information (localization, imaging, environment reconstruction, etc) obtained from sense can help improve communication performance, such as high-precision beamforming and less overhead channel state information (CSI) acquisition.
In the ISAC system, reducing or optimizing bi-static sensing/channel estimation overhead is a critical challenge to improve the data rate in mmWave/THz communication [1]. As shown in Fig. 4, we propose a scheme optimizing the per-user channel sensing duration in multi-user multiple-input single-output (MU-MISO) systems. For each user, the BS predicts the effective rate to be achieved after pilot transmission. Then, the channel sensing duration of each user is optimized by ending the pilot transmission when the predicted rate is lower than the current rate.
Fig. 6. An illustration of the optimizing channel sensing duration in the ISAC system
4. UAV Tracking
Unmanned aerial vehicle (UAV) is widely used for military, delivery, and disaster because of many advantages of network formation, high mobility, and low cost. It is expected that the UAV control technique for target tracking is essential because UAV collects environmental knowledge and provides an effective wireless network to ground users. We have been studying deep reinforcement learning (DRL) based on multiple UAV control for target (i.e., ground user) tracking. DRL is an emerging tool to select optimal action in a dynamic system, for example, autonomous driving, UAV control problem and wireless network, etc. By applying DRL to multiple UAVs system, UAVs learn how to take actions to track/localize the ground target in a complex 3D environment. Besides, the DRL-based UAV control technique mitigates computation time that becomes critical when the number of UAVs or targets increases.
Fig. 7. Multiple UAVs track ground users in challenging 3D environment
Multiple UAVs are assigned to tracking missions in the existence of multiple ground targets. Each of the UAVs is equipped with a range sensor that measures the distance between UAVs and targets. The measurements received by UAVs are blocked due to obstacles and structure.
Video1. The results of 3D trajectories of UAV and target (left), and localization error (right) during the tracking missions
While operating tracking missions, UAV’s flight direction is determined by the trained-DRL network. Tracking error measures MSE between the true target position and estimated target position. Applying DRL to target tracking enables UAVs to track and localize ground users accurately.
5. Direction Finding
Direction finding is an essential technology to counter enemy air infiltration and threats and is also widely used in array signal processing such as radar, sonar, and wireless communication systems. Direction of arrival (DoA) estimation using an array antenna to find the direction of unknown signals is a representative direction finding method.
During the last decades, subspace-based DoA estimation algorithms like multiple signal classification (MUSIC) and estimation of signal parameters via rotational invariance techniques (ESPRIT) were proposed. Also, with the rapid development of artificial intelligence (AI) technology, many machine learning-based DoA estimation algorithms were proposed.
Fig 8. A scheme of deep learning-based DoA estimation.
Our lab mainly focuses on various machine learning-based DoA estimation for practical applications. We are also using the testbed for real-world data collection and DoA estimation algorithm validation.
Fig 9. A picture of transmitter and receiver used for the experiment.