In long-term Internet of Things (IoT) sensing systems, such as environmental monitoring, industrial inspection, transportation surveillance, and structural health monitoring, inter-sensor relationships often evolve over time because of sensor displacement, device aging, sensor failure, device replacement, and environmental variations.As a result, historical observations may become structurally outdated, thereby biasing graph topology learning and degrading both topology estimation accuracy and graph signal recovery performance. To address this challenge, we propose a two-stage framework for dynamic graph topology learning that explicitly mitigates the adverse effects of outdated observations. First, an anomaly-detection-based grouping strategy identifies observations associated with topology transitions and partitions the data stream into structurally consistent groups. Second, a dynamic group-based topology learning model estimates or updates the graph topology for each group using only structurally relevant observations.The proposed method establishes a theoretically grounded relationship between topological perturbations and graph signal recovery error in the graph frequency domain.Under the graph signal smoothness assumption, we further derive the analytical probability distribution of a frequency-domain anomaly-detection operator, thereby enabling principled identification of topology change points and observation segmentation. Building on this analysis, we formulate conditions linking anomaly detection decisions to topology changes and develop a dynamic grouping strategy for time-varying graph learning. Experiments on synthetic and real-world datasets demonstrate that the proposed method effectively captures topology changes, mitigates the adverse effects of outdated observations, improves topology estimation accuracy, and achieves strong graph signal recovery performance.

To address the channel estimation challenge in semi-passive Intelligent Reflecting Surface (IRS)-aided communication systems, a graph-based active element deployment and channel estimation method is proposed, combining graph smoothing with atomic norm minimization. The semi-passive IRS channel is modeled as a graph signal, and the element deployment is reformulated as a graph sampling problem. Then, spectral clustering is applied, and the vertex with minimum total variation in each cluster is sampled. Based on that, a channel estimation method is developed by incorporating the smoothness priors of the graph signal with atomic norm constraints, resulting in an atomic norm minimization under smoothness constraints. Simulations reveal the performance improvement of the proposed scheme in channel estimation for semi-passive IRS-aided systems.

In this paper, a new echo signal detection method, to the best of our knowledge, for pseudo-random single-photon counting ranging (PSPCR) LiDAR systems is proposed, which is applied for long distances, low repetition rates, and system cost reduction. First, in order to achieve a comparable temporal resolution as that in time-correlated single-photon counting (TCSPC) systems, we extend the pseudo-random code to discriminate the minimal time slot in time correlation. Second, we use the full width at half maxima (FWHM) in the duration of each pseudorandom code for correlation to reduce the impact of pulse width variation and timing jitters on ranging accuracy. Third, we study the bias and errors caused by using synchronous signals as the "START" signal, and propose to use the time of flight (ToF) at half energy to reduce the walk error. Simulation results show that, compared with existing PSPCR methods, the proposed method improves ranging accuracy with a lower repetition rate and lower peak and average power-centimeter-level ranging accuracy over tens of kilometers can be achieved using a laser with a repetition rate of 400 kHz, peak power of up to 1 kW, and average power of up to 1 W. (c) 2024 Optica Publishing Group

AbstractThe high spatial search complexity of the direct positioning method in passive positioning systems leads to long positioning time and high computational resource consumption. In response to this issue, this article proposes a fast localization scheme based on composite convolutional neural networks (CCNN), which can effectively explore the correlation between the position of the radiation source and the characteristics of the received signal. CCNN is a 20‐layer composite network based on fully convolutional network layer, which is composed of convolutional layers, batch normalization (BN) layers, and ReLU activation function layers with unidirectional connections. Then, CCNNs are adjusted and trained for positioning single and multiple radiation sources, respectively. Simulation results show that the computational time of the proposed method can be reduced by nearly 98% compared with the direct positioning scheme. Meanwhile, about 71.2% of positioning error's reduction is achieved.

In the face of increasingly dynamic, varied, and so-phisticated jamming attacks, conventional anti-jamming solutions are insufficient to ensure the reliability of satellite communication systems in a hostile jamming environment, leading to the vulner-ability of communication links. In this paper, a new space-time anti-jamming algorithm based on satellite collaboration is proposed. Firstly, an anti-jamming model for the uplink of satellite communications is developed, in which the inter-satellite collaboration is proposed for adaptive interference cancellation. The proposed method enables efficient joint space-time anti-jamming through the exchange of jamming information between reliable inter-satellite links. Secondly, we formulate the collaborative and joint space-time anti-jamming problem under the constraint of different Doppler shifts of jamming signals between the receiver and its collaborator. To solve this problem, we further propose a phased-array space-time anti-jamming algorithm which is jointly designed by the Doppler shift compensation to the jamming signal. Simulations over four types of jamming signals show the performance improvement and the robustness of the proposed method in anti-jamming compared with conventional phased-array beamforming in satellite communications.

Global Navigation Satellite System (GNSS) signals are inherently weak and highly susceptible to jamming. Traditional signal analysis-based detection methods struggle with accurate identification in scenarios involving multipath propagation, weak interference, or complex interference types, leading in high missed detection rate (MDR). To solve these issues, a convolutional neural network (CNN) based interference detector by using carrier-to-noise density ratio (CNR) parameters from large-scale observation station is proposed in this paper. By learning the change principle in CNR features at lots of receiving stations under interference, a deep learning model is designed to effectively detect interference. In order to automatically extract the feature behind CNR data well, a CNN consisting of two convolutional layers, two max-pooling layers, and a fully connected dense block is built and trained. After preprocessing of normalization and dimensional transformation, the CNR observation data are fed into CNN to produce detection results. Simulation results demonstrate that a significant improvement in detection capabilities can be provided by the proposed method under both weak and complex interferences. Across the mixed dataset including 4 types of suppression jamming and 3 JSR values, an detection accuracy of 98.49% can be achieved by the CNN scheme, while only 95.12% is got by the baseline method, i.e RFA.

Indoor localization plays an essential role in enabling location-based services (LBSs) for sensor networks wherever global navigation satellite systems (GNSSs) are unreachable. For deep-learning-based indoor localization methods, radio data are extremely important for accurate fingerprinting-based localization but are often incomplete or are not temporally and spatially sufficient for data acquisition. Motivated by the goal of leveraging existing radio fingerprints to further improve localization accuracy, in this article, we propose a graph-based fingerprint augmentation method for deep-learning-based indoor localization. In this method, by modeling each reference point (RP) as a vertex of a graph and its radio data as graph signals, we develop a graph signal model where virtual RPs are introduced as missing vertices with missing radio data. Then, we explore the underlying spatial structure among all the real and virtual RPs to find the graph Laplacian with which to reconstruct the radio fingerprints by a semisupervised graph interpolation. On this basis, some deep neural networks and convolutional neural networks (DNNs and CNNs) trained by the reconstructed radio fingerprints are developed for indoor localization. Experiments on real datasets reveal the performance gain of the proposed fingerprint augmentation method in localization accuracy, showing the potential of graph-based data augmentation for deep-learning-based indoor localization.

In recent years, the development of automatic driving is trending and drawing extensive attention both from industry and academia. One of the key elements of automatic driving is the accurate perception of the automatic driving environment. This is critical to planning and control. As the diversity and number of sensors become more and more complicated in equipping the self-driving vehicle, representing features from different views in a unified perspective comes to vital importance. The well-known bird’s-eye-view (BEV) is a natural and straight-forward candidate view to serve as a unified representations. In this paper, we propose an end-to-end birds-eye-view perception based on Homography matrix, relying only on the camera and without post-processing. Our network architecture achieved by two stage network: the image-view network and the Bev-View network, information flows to the Bev-View network through Homography matrix. Extensive experiments demonstrate that our method achieves a new performance.

Data acquisition in the large-scale Internet of Things (IoT) demands high-cost sensing and communication. In this paper, we develop a data-driven deep learning method for low-complexity sensor selection and high-accuracy data reconstruction in large-scale IoT. The proposed deep learning model is designed with a sampling network for sensor selection and a deep neural network (DNN) for data reconstruction, where the Gumbel-max trick is applied to ensure that the loss function of the model is differentiable. With the design of an interconnected sampling network and reconstruction network, the sampling matrix for sensor selection and its corresponding data reconstruction can be trained simultaneously. The proposed method is data driven and does not require the graph structure in advance; hence, it can avoid the high-complexity computation required by graph sampling-based methods. Experiments on both synthetic and real-world datasets reveal the performance improvement of the proposed method in computational complexity and data reconstruction accuracy.

To get higher utilization efficiency of spectrum and orbits in low earth orbit (LEO) satellite system, satellite communication and navigation integration has become a search hotspot recently. Because orthogonal time frequency space (OTFS) modulation can reduce the effect of time-frequency doubly selective channels between the LEO satellite and earth, an OTFS-based communication and navigation integrated scheme is studied in this paper. First, a hybrid modulation constellation by superimposing quadrature amplitude modulation (QAM) and dual-code alternating binary offset carrier (AltBOC) modulation is proposed to realize signal integration of two systems. Second, a serial interference cancellation (SIC) detector is designed to separate two signals. In addition, a direct path time delay estimator based on pseudo-range code measurement is designed. What's more, the BER performance of the integrated signal is derived. Simulation results demonstrates that the proposed scheme has the lower BER and the higher localization accuracy than the baseline method under a LEO scenario with doubly selective channel.

针对利用直接定位(DPD)方法对全球导航卫星系统(GNSS)终端面临的干扰源进行定位时,在线计算复杂度高的问题,提出了一种利用神经网络逐级缩小定位区域的低复杂度多级干扰源直接定位方法.该方法首先使用多级全连接神经网络(FNN)逐级缩小干扰源所在的区域范围,每一级处理将目标区域均分为两个子区域,并利用预训练的以接收信号功率作为输入特征的神经网络选出干扰源所在的子区域,从而大幅缩小干扰源所在的目标区域;然后在锁定的最终子区域内划分网格并使用DPD方法对干扰源进行精细定位.由于逐级二分处理已将干扰源可能存在的区域大幅缩小,因此有效降低了 DPD方法的网格搜索集合大小.仿真结果表明:在较高干噪比条件下(对于压制式干扰通常大于20 dB),所提方法能获得与DPD方法接近的定位性能,而在线计算复杂度相比于DPD方法可降低大约2M倍(M为神经网络级数),定位误差相比于传统基于到达时间差(TDOA)的Chan氏定位方法可降低90％以上.

In a low earth orbit (LEO) satellite communication system with massive multiple input multiple output and non-orthogonal multiple access (MIMO-NOMA) transmission, how to reduce the inter-cluster interference as the same time to ensure the receiving performance of each user within a cluster is an important problem in downlink design. Hence, a beamforming scheme to maximize the weighted sum of signal to leakage plus noise ratio (SLNR) and signal to interference plus noise ratio (SINR) under the transmit power constraint is proposed in this paper. First, a user grouping strategy is adopted based on the spatial angle according to the characteristics of the channel from satellite to ground. Then, an alternating direction method of multipliers (ADMM) based solving procedure is designed to handle the beamforming problem by utilizing the advantage of ADMM in solving a multi-objective problem. The simulation shows that the proposed scheme gets a better sum rate than the baseline linear beamformers due to its more efficient optimization objective.

To avoid the high calculating requirement of the central node in the centralized positioning system, a distributed scheme with multiple localization sub-networks to localize an unknown radiation source (RS) is discussed in this paper. Furthermore, in order to reduce the adverse effect of the potential information loss in the parameter estimation of the traditional two-step localization algorithm, a neural network to deduce the objective position from the hybrid localization parameters is designed for each sub-network to replace the solving of positioning equations involved in the two-step localizer. Finally, in order to obtain more accurate localization results, two fusion methods based on the weighted sum and the neural network are given at the final control center to fuse the localization results of all the sub-networks. The experimental results show that the proposed scheme significantly outperforms the traditional single-parameter based two-step localization method even though the latter one adopts the centralized positioning strategy. Furthermore, the neural network based fusion method can improve the final positioning precision obviously for signal to noise ratio (SNR) higher than 10dB.

The explosively-growing Internet of Things (IoT) is generating massive amount of multivariate time series data. Anomaly detection for multivariate time series is of great importance in detecting and locating system failures, device malfunctions and malicious attacks in IoT. In this paper, we propose a new anomaly detection model, called Graph Attention-based Gated Recurrent Unit (GAGRU), to exploit the complex spatiotemporal features in multivariate time series. The proposed GAGRU is basically an attention-based Gated Recurrent Unit (GRU), which embeds the graph attention mechanism into the GRU to integrate the spatial features while exploiting the temporal dependency of data using GRU. Moreover, the GAGRU model is also designed to automatically learn the graph structure and score anomalies under their own TopK criterion. Experiments on two real-world datasets demonstrate the performance improvement of the proposed GAGRU model.

Orthogonal frequency division multiplexing (OFDM) is one of the key technologies in the physical layer of the internet of things (IoT).Pilot design and channel estimation are key issues in OFDM systems.In view of the problem of performance loss by fixed pilot pattern due to the complexity and variety of IoT communication scenarios, a pilot design and channel estimation scheme based on graph signal processing (GSP) was proposed.Firstly, the time-frequency resource block was modeled as a graph signal, and the channel estimation problem was reformulated into a sampling and reconstruction problem of the graph signal.Then, considering the influence of time-frequency fading, a weighted graph adjacency matrix was designed to construct a graph topology structure based on the time-frequency position.On this basis, the pilot position is selected based on the graph signal sampling theory, a greedy pilot pattern design algorithm based on weighted graph topology was proposed.At the same time, signal reconstruction was performed based on the graph signal reconstruction method, and a channel estimation method based on the graph smoothness constraint was proposed.Compared with the conventional scheme, simulation results show that the proposed method achieves higher channel estimation accuracy in high-speed scenarios of double selective channels, and effectively reduces pilot overhead in low-speed scenarios.

In this work, the signal recovery problem regarding incomplete and noisy spatio-temporal signals is studied. A spatio-temporal signal is considered as a time-varying graph signal and a diffusion-induced first-order Markov signal model is developed to incorporate both the spatial structure and temporal correlation into the underlying graph. With this model, prior knowledge on spatial smoothness and temporal correlation is revisited, and the connections between the graph structure and differential temporal smoothness are revealed. The authors then accordingly formulate a spatio-temporal signal recovery method by jointly exploiting the spatial smoothness, low rank and refined differential temporal smoothness. The formulated recovery problem is solved by a block coordinate descent-based algorithm, which iteratively optimises the recovery accuracy and temporal correlation matrix. The experiments on three real-world datasets reveal the high signal recovery accuracy of the proposed algorithm.

In this paper we develop a clustering-aided signal sampling and reconstruction method for data acquisition in large-scale sensor networks. Using the localization feature of a large network, we exploit the vertex-domain locality by the localized operator of each vertex on the graph, and develop a clustering method that sequentially selects cluster heads and their corresponding members by the use of the overlap factor of each vertex. On this basis, we apply greedy sampling set selection for each cluster in a distributed manner. By combining all local sampling sets, the global sampling set is selected and signals over the whole graph is then efficiently reconstructed. Simulation results over various large networks show that compared with existing sampling set selection methods, the proposed method can reduce the computational complexity while achieving acceptable reconstruction accuracy.

In order to improve the bit error rate (BER) performance of a MIMO-NOMA system with successive interference cancellation (SIC) detector at user terminals, a nonlinear precoding scheme to directly minimize the inter-cluster interference is proposed in this paper. It jointly optimizes the weighted coefficients in the equivalent channel information of each user cluster and the precoded results under a total transmission power constraint. Furthermore, a solution procedure based on the alternating direction method of multipliers (ADMM) algorithm is designed for the above optimization problem. In addition, a user clustering scheme is given to coordinate with the proposed precoder to get the better performance. Simulation results demonstrate that the system whole BER performance of the proposed scheme is improved obviously compared with other precoders, such as the traditional zero force (ZF) precoder and so on. Especially, the average BER of the users with stronger channel in each cluster is reduced by nearly three orders of magnitude at signal to noise ratio (SNR) of 25dB.

Massive user activity detection is a challenging task for massive Internet of things (mIoT). In this paper, we propose a new deep neural network, named concentrated layers convolutional neural network (CLCNN), for user activity detection in mIoT. We firstly propose three basic rules in the design of residual network specifically for mIoT scenarios. Secondly, with the rules above we develop a new improved residual network block which includes integrated convolutional layers with activation functions, by which the residual convolutional network is constructed. Moreover, the regularization and its corresponding hyperparameter for the proposed network are also investigated against overfitting. Simulation results show that the proposed CLCNN network outperforms the existing deep learning algorithm and conventional compressive sensing solutions in user activity detection and corresponding channel estimation.

In this letter, a deep Q-learning network (DQN) based resource allocation (RA) scheme is proposed for the massive multiple-input multiple-output (MIMO)- nonorthogonal multiple access (NOMA) systems. The reinforcement learning (RL) frame is developed to build an iterative optimization structure for user clustering, power allocation and beamforming. Specifically, a DQN is designed to group the users based on the reward item calculated after power allocation and beamforming. The objective is to maximize the reward item, i.e., the system throughput. Then, a back propagation neural network (BPNN) is used to realize the power allocation. During the training of BPNN, the exhaustive search results in the quantized power set are taken as the output labels. Simulation experiments show that the proposed scheme can achieve high system spectrum efficiency approximating to the exhaustive search based on user clustering and power allocation.