In the field of non-cooperative network topology inference for unmanned aerial vehicle (UAV) swarm networks, current causality discovery-based methodologies typically depend on critical assumptions that the data and acknowledgment (ACK) frames can be reliably distinguished and that the prior knowledge of transmission configurations, such as packet size, is available. However, in real-world scenarios, due to feature variations stemming from heterogeneous device configurations and undisclosed protocols, such assumptions are difficult to be satisfied. This paper proposes Frame-Agnostic Adaptive Causal Topology Inference (FAAC-TI), a novel network topology inference framework that overcomes the above limitations through two innovations. First, we propose Self-Informed Transfer Entropy (SITE) utilized as causality metric in FAAC-TI, which integrates the source node’s own past history contribution with cross-node causal influence, eliminating the conventional dependence on pre-classification of Data/ACK frames. Second, FAAC-TI contains a data-driven adaptive parameter selection scheme which determines parameters through run-length statistical characterization of binary time series when calculating causality metric, obviating the need for prior knowledge as previously mentioned. Experimental results show that the SITE metric achieves significantly higher topology inference accuracy on average than transfer entropy (TE) baselines across varying network scales and dynamics, while the adaptive parameter selection scheme maintains close deviation from the conventional parameter presetting scheme based on prior knowledge.

This paper focuses on non-orthogonal multiple ac-cess (NOMA) scenarios with emerging sensing targets, where beamforming is utilized to integrate sensing and NOMA communications (NOMA-ISAC). This scheme does not require ded-icated sensing signals, avoids further degradation of system performance in the overloaded regime, and facilitates massive device connectivity in future wireless networks. We investigate the performance region of NOMA-ISAC from an information-theoretic perspective and propose a learning-based design to approach the given performance bound. Specifically, the upper bounds for the sensing estimation rate and communication information rate are derived to evaluate the achievable rate region of NOMA-ISAC. Guided by mutual information, the proposed end-to-end learning architecture for NOMA-ISAC enables joint transmitter-receiver optimization to achieve bound-approaching performance. Furthermore, we design a sensing mutual information neural estimation unit (SMEU) for the proposed architecture to evaluate the available sensing mutual information. Simulation results show that the proposed scheme can approach the theoretical bounds of NOMA-ISAC.

Deep learning-based semantic communication is a promising research direction for next-generation communication systems. The emergence of large language models(LLMs) with remarkable semantic comprehension abilities leads us to consider whether LLMs can be used in semantic communication to enhance model's performance. In this paper, we discuss the main implementing details of the idea by proposing a general end-to-end learning semantic communication model with LLM, including subword-Ievel tokenization, a rate adapter based on gradients for matching the rate requirements of any channel codec and fine-tuning for possessing private background knowl-edge. By taking Bidirectional and Auto-Regressive Transformers (BART) and Generative Pre-trained Transformer 2 (GPT2) as examples, we demonstrate how we can utilize various structures of LLMs to design semantic codecs. In terms of semantic fidelity, generalizability to cross-scenario, and complexity, results reveal that the LLM-based semantic communication system achieves exciting performance. We hope this initial work can inspire more research devoted to this field.

Visual inspection is an effective approach for anomaly detection in the industrial Internet of Things. The inpainting-based strategy is widely adopted for industrial visual inspection. However, the training process of the semantic-based inpainting model requires heavy resource consumption due to its complex structure. Additionally, the inspection performance of the model degrades when it encounters an out-of-scope image. Image can be naturally deemed as a tensor with the global low-rank property. Therefore, we propose a data structure-based strategy to implement effective inpainting in this paper. Inspired by the tensor-tensor product, a patch-masked deep tensor factorization model is constructed to reconstruct the intentionally masked region. This model has a simplified structure with three feed-forward neural subnets sharing one trainable low-dimensional input. Besides, Laplacian regularization is imposed to improve the recovery accuracy based on the local smoothness in images. Above all, we apply a parallel approach to implement patch-masked visual inspection by comparing the structural similarity between the recovered and original patches. It is independent of the dataset and works as a sample-related pattern. Experiments conducted on the MVTec dataset demonstrate that, although our model has lower inspection resolution than semantic-based inpainting models, it has much higher training efficiency, which makes it more beneficial for practical deployment.

The recent proposed long-term spectrum prediction is a promising technology to predict a power spectral image from time-frequency dimensions in the next period, which plays a crucial role in spectrum management and electronic countermeasures. However, real-world spectrum data often suffer from data loss and noise interference, posing challenges of ensuring prediction accuracy. This paper focuses on the periodic prediction of the future spectrum based on corrupted measurements, enabling the generation of a spectral image that captures frequencies and time slots for the subsequent period. We formulate the task of recovering spectrum data from noisy observations as a tensor completion problem by minimizing the sum of the weighted tensor nuclear norm, the noise variance, and the deviation values at the missing position. For this optimization problem, we employ an adaptive non-convex relaxation to approximate the rank of the spectrum tensor. Additionally, a generic noise Frobenius term is adopted to effectively eliminate arbitrary noise with a certain distribution. Based on the model, we propose a joint method that integrates a pre-fill operation into the completion model to enable simultaneous prediction of future spectrum and recovery of corrupted historical data. Experiments on both synthetic and real-world spectrum data verify the effectiveness of the proposed method.

End-to-end learning communication provides an exciting new approach to physical layer design, which inspired us to design the Non-orthogonal multiple access (NOMA) system with superposition coding guided by information theory. Unlike assuming continuous Gaussian inputs in most NOMA systems, we consider actual finite-alphabet inputs at two transmitters. Due to the lack of capacity region under finite cardinality, optimizing design is challenging. In this work, we first give a closed-form expression for the conditional mutual information to define the capacity region of NOMA with finite-alphabet inputs. To simplify the conditional mutual information, we then derive a tight lower bound on mutual information, which is necessary to achieve an accurate and consistent mutual information estimator. Finally, we propose an end-to-end learning model for NOMA that considers maximizing mutual information and bit error rate (BER) constraints to achieve optimal encoders and decoders. Simulation results show that the proposed scheme achieves the capacity region of NOMA with finite-alphabet inputs. It also significantly improves the bit error performance of the system than prior work.

Time-frequency analysis is one of essential signal processing tool for wireless sensor signal in Internet of Things (IoT), but its traditional processing approaches, such as short time-frequency transformation (STFT) and discrete wavelet transformation (DWT), are challenged by the limitation of capability to self-learn from unknown environments and adjust parameters adaptively. To address this problem, we propose to build up the deep learning network to learn time-frequency analysis to instead of traditional STFT and DWT approaches. With using typical neural network layers to remodel STFT and DWT operations, the proposed models consider the efficiency on both training and processing procedures and show their parameter adaptability and capability of deep feature extraction from sensor signals. Moreover, we demonstrate how to integrate learning time-frequency analysis networks into practical IoT applications, signal detection in noisy environment, and classifying of various modulated wireless sensor signal, by which their performance are further evaluated in terms of computation complexity and efficiency.

In this paper, we investigate the problem of utilizing the Kalman Filter to reconstruct signals with sparse frequency content under a streaming CS framework. We develop a Gaussian Markov model of the sparse streaming signal under the Analog-to-Information Converter (AIC) hardware structure and propose an adaptive Kalman Filter for the reconstruction. Different from existing CS schemes for streaming signals, we exploit the correlations between the signals of two consecutive observation windows to model the process in the state transition form so that the Kalman Filter can be incorporated to obtain the convergent estimation of the input streaming signal. Simulation experiments show the feasibility of the proposed model and demonstrate the superior performance of the proposed algorithms.