IET Signal Processing最新文献

In the field of signal processing, modulation signals, including phase shift keying (PSK) and quadrature amplitude modulation (QAM), can significantly enhance the signal-to-noise ratio (SNR) through aliasing transmission following clustering and sorting. This article presents two novel approaches to compressed time difference of arrival (TDOA) estimation, leveraging amplitude-phase clustering signals. A carefully designed compression matrix is constructed based on the unique amplitude and phase characteristics of the signals. The study then analyzes the Cramér–Rao lower bound (CRLB) under full-sampling conditions. Finally, TDOA estimation is performed using the approximate maximum likelihood (AML) method. Simulation results demonstrate that the proposed compressed sampling TDOA estimation methods, based on amplitude-phase clustering, achieve accuracy within an order of magnitude of full-sampling performance. Additionally, this article explores the application of OFDM-QAM signals, which exhibit amplitude-phase convergence in the frequency domain, for time difference estimation in compressed sampling. A novel frequency-domain aliasing time difference estimation algorithm based on amplitude-phase convergence is proposed. Experimental results indicate that under high SNR conditions, the algorithm incurs only a minor SNR degradation of ~4 dB compared to time difference estimation in uncompressed transmission.

Hand gesture recognition using mmWave radar has emerged as a promising technology for human–computer interaction (HCI), smart home systems, and the Internet of Things (IoT). However, the practical application of this technology is often constrained by the high computational complexity and significant storage demands of contemporary deep neural networks, which impede their deployment on resource-limited embedded devices. To address this limitation, we present a novel approach that combines an improved MobileViT model with a knowledge distillation (KD) framework. The proposed method consists of three main stages. First, raw radar signals are captured and restructured into a three-dimensional format (Chirps × Samples × Frames, a 3D tensor) and processed to generate range-time maps (RTMs) and Doppler-time maps (DTMs). Second, an improved MobileViT network is designed, incorporating fewer redundant blocks, a lower input resolution, and a dual-branch input structure to effectively fuse features from the RTM and DTM. This enhanced architecture serves as a robust teacher model, excelling at extracting both local and global spatiotemporal features for accurate gesture recognition. Finally, KD is applied to transfer knowledge from the teacher model to a compact student network, thereby achieving model compression. Experimental results demonstrate that the final distilled student model, evaluated on the test set, has only 0.018 M parameters—~10% of the teacher model’s size—while still achieving a high recognition accuracy of 99.16%. Consequently, the resulting model is highly compact and accurate, demonstrating its suitability for real-world embedded deployment.

Brain-computer interface (BCI) plays an important role in various fields, such as neuroscience, rehabilitation, and machine learning. The silent BCI, which can reconstruct inner speech from neural activity, holds great promise for aphasia patients. In this paper, we design an imagined Chinese speech experimental paradigm based on initials and finals and collect raw signals from eight healthy participants by using 64-channel scalp electroencephalograms. Linear predictive coding (LPC) and mel frequency cepstral coefficients (MFCC), which are classical algorithms in the field of speech recognition, are used to extract distinguishing features for speech classification and reconstruction. Besides, the phase-lock value (PLV) is introduced to enrich the feature information. We choose support vector machine (SVM), linear discriminant analysis (LDA), decision tree (DT), and LogitBoost (LB) for binary classification in several different cases. Two-channel selection (CS) based on Broca’s area and Wernicke’s area of the brain is also introduced in the paper. The highest imaginary speech decoding accuracy reaches 84.38%, which demonstrates the effectiveness of the feature engineering. In addition, the comparative analysis is conducted with deep learning methods specifically designed for small sample scenarios. This study offers a novel systematic approach for the research of Chinese speech imagination BCI.

Obstructive Sleep Apnea (OSA) is a prevalent and underdiagnosed sleep disorder that can lead to serious cardiovascular and cognitive complications if left untreated. This study presents a novel deep learning architecture based on convolution, LSTM, short-time Fourier transform (STFT), attention, and transformer modules for automated OSA detection using single-lead electrocardiogram (ECG) signals, aiming to improve diagnostic accuracy. The proposed model integrates six parallel branches for feature extraction, combining convolutional layers, recurrent units, STFT, and residual connections to capture multiscale temporal and frequency-domain patterns. A three-path feature refinement that incorporates sequential, convolutional, and transformer encoder is considered as the second part of the model. Channel attention-based feature fusion modules are employed in the first and second parts to enhance feature relevance and suppress noise. Experimental evaluations on the PhysioNet Apnea-ECG dataset demonstrate that the proposed model achieves superior segment-level classification performance with 91.97% accuracy, 91.28% sensitivity, and 92.41% specificity. These findings suggest that the proposed method offers a robust, and scalable solution. Regarding the small number of parameters of the model, it can potentially be considered for real-time and wearable-based OSA monitoring applications. All codes and the trained model are released at https://github.com/mziaratban/OSA.

Currently, gas turbines are finding increasingly widespread applications. To ensure their safe and stable operation, condition monitoring and anomaly detection are crucial. However, traditional anomaly detection methods for gas turbines often rely solely on a single learning approach. To address these issues, this paper proposes a novel semisupervised learning framework that synergistically combines the large-scale automatic labeling capability of unsupervised learning with the precise classification power of supervised learning. First, an unsupervised learning algorithm is employed to hierarchically label anomalies in real operational data as suspicious anomalies, high-probability anomalies, and actual anomalies. Next, oversampling techniques are applied to address class imbalance issues by augmenting underrepresented classes in the dataset. Finally, supervised learning methods are utilized to train models on the labeled samples, with their performance compared against other machine learning (ML) approaches. Through comparative analysis of multiclass classification evaluation metrics, the feasibility of the proposed semisupervised learning framework is demonstrated, and the optimal monitoring model is identified. The core contribution is providing a semi-supervised learning framework that categorizes operational data into a multitier hierarchy to enable a nuanced early warning mechanism.

Formation control of multi-autonomous underwater vehicle (AUV) systems is essential for collaborative ocean tasks such as environmental monitoring and resource exploration, but it is hindered by challenges like limited flexibility in formation transformation, model uncertainties, and unknown ocean current disturbances. To address these issues, this paper presents a finite-time flexible formation control strategy based on the affine image framework for second-order multi-AUV systems with AUV model uncertainties and external disturbances. The proposed strategy incorporates a finite-time disturbance observer to accurately estimate composite disturbances (including model uncertainties and ocean currents) and a sliding mode control (SMC) approach to ensure robust finite-time convergence of the formation. A key innovation is the extension of the affine image framework to second-order dynamics, enabling flexible transformations between affine and rigid formations while requiring only d + 1 leaders (in d-dimensional space) to know the target shape, thus, minimizing global information needs. The proposed method achieves finite-time stability and strong robustness against uncertainties, providing a unified solution for flexible formation control in uncertain underwater environments. Numerical simulations validate that the strategy effectively guides AUVs to converge to desired formations within finite time.

Underwater images serve as one of the most intuitive media for human perception of underwater environments. However, the limited bandwidth and susceptibility to noise in underwater acoustic (UWA) channels pose significant challenges for traditional image encoding and transmission methods, thereby hindering high-quality image reconstruction. Semantic communication aims to shift from pixel-level to semantic-level transmission, enhancing both reliability and efficiency. This paper proposes a scene-guided generative communication method for underwater images with semantic alignment. We decouple and transmit only the essential layout information of underwater scenes. This enables highly efficient compression. At the receiver, we employ a graph convolutional network (GCN) to correct layout distortions and a context-aware diffusion model to generate realistic underwater images that preserve high semantic fidelity to the original. Experimental results demonstrate that, compared to pixel-fidelity communication and other generative communication approaches, our method consistently achieves superior image reconstruction quality even under adverse channel conditions, such as extremely low signal-to-noise ratios (SNRs), and exhibits significant advantages in downstream tasks.

Existing end-to-end stereo matching networks face significant deployment challenges due to their reliance on large training datasets and the limited ability of synthetic data to represent real-world scenarios, leading to a severe domain shift. To address these challenges, this paper proposes a novel stereo matching network along with an efficient fine-tuning strategy. First, a lightweight modular stereo matching network is proposed, which incorporates domain knowledge and optimizes specific modules to enhance generalization. Second, a confidence estimation network is developed to generate occlusion masks, which filter erroneous self-supervised loss. Then, the Mann–Kendall (MK) trend detection method is used to evaluate the loss change trend of the most recent few frames in the online adaptive process to measure the model’s adaptation degree to the scene, and control the online operation mode of the model. Finally, online and offline experiments on multiple datasets demonstrate the competitive performance of our method.

For effective management of the ecological balance of natural resources, it is crucial to conduct real-time environmental monitoring in the context of mining activities. This research demonstrates five hybrid machine learning models—XGBC + neural networks (NNs), autoencoder + isolation forest, SVM + long short-term memory (LSTM), graph NN + random forest (GNN-RF), and transformer + CatBoost—for the integrated real-time ecosystem monitoring in mining regions framework. The GNN-RF model performed best with a training accuracy of 99.12% and a testing accuracy of 93.81%, surpassing the rest. Five optimizers were tested in optimizing the GNN-RF model: momentumized, adaptive, dual-averaged gradient (Madgrad), AdaHessian, layer-wise adaptive rate scaling (LARS), sharpness-aware minimization (SAM), and LION. The GNN-RF model with the Madgrad optimal setting achieved the best training record of 99.34% while maintaining 93.81% during tests, showcasing high generalization capabilities and surpassing other configurations. This Madgrad-optimized GNN-RF model configuration incorporates the advantages of graph NNs in spatial and relational data pattern recognition, along with RFs’ complex-feature interactions, enhancing performance for streaming monitoring tasks. GNN-RF with Madgrad becomes the proposed model for scalable and precision-centric detection of anomalies in mining regions to maintain the balance of all ecosystem elements within a rapidly changing ecological environment. This research highlights the promise that advanced optimization methods combined with a hybrid machine learning framework hold in mitigating environmental issues in developed industrial sectors, forming a basis for progress in sustainable mining techniques.