IET Signal Processing最新文献

Understanding natural phenomena relies heavily on learning criteria aimed at minimizing model errors and enhancing fidelity to real-world conditions. This study investigates a linear combination of three criteria: minimum mean square error (MMSE), minimum error entropy (MEE), and maximum correntropy criterion (MCC), using a gradient descent algorithm. We explore three data scenarios: noise-free, Gaussian noise, and non-Gaussian noise. Incorporating feedback, we extend our analysis to encompass additional noise models such as Laplacian noise and Poisson noise. Our findings reveal that leveraging the MEE criterion in isolation effectively identifies the model’s target vector, showcasing its resilience across diverse noise conditions.

In the application of automotive radar, vehicle turning is a critical scenario. The introduction of rotational angular velocity causes Doppler shifts in forward-facing speed radars, leading to defocusing of multichannel echo data. This paper attempts to propose a method for compensating and correcting the rotational angular velocity to focus the energy across channels by compensating for the speed in each channel. Considering the rotational angles of the vehicle-mounted platform, the real incident angle of an unknown target must be accounted for since the main beam covers a relatively large range. Therefore, we include the target’s angle of arrival, conducting a search within the main beam range during detection. When the true angle of arrival of the target is identified, detection performance reaches its optimal level. After, employing the rotational speed search method proposed in this paper, under a four-channel rotating radar platform, the signal-to-noise ratio (SNR) for target detection is enhanced by approximately 15 dB, which aligns with the theoretical gain from coherent accumulation of SNR derived in the subsequent sections of the paper. Furthermore, obtaining the exact value of the platform’s rotational speed may not always be easy. Hence, we incorporate the rotational speed into the search range. After, performing a two-dimensional search, the peak of the detection performance graph corresponds to the true rotational speed of the vehicle-mounted platform and the angle within the main beam range where the target is located. Conversely, knowing the prior information of any dimension in the two-dimensional search can expedite and improve the detection performance of the other dimension’s parameters.

Weak coherent light interference heterodyne detection is the theoretical basis for fiber optic gyroscopes, optical coherence tomography, and optical time-domain reflectometers. Classical statistical optics provides the signal model for weak coherent light interference. However, this theory does not describe signal acquisition and nonpolarization, which are significant in the analysis of heterodyne detection frequency, coherent length, and polarization mode dispersion (PMD). Consequently, it has difficulty solving signal processing problems related to coherent frequency and length analysis. This article proposed a time-domain signal analysis method. The approach can describe the practical signal acquisition and the polarized direction interference and accurately obtain coherent frequency and length on weak coherent light interference heterodyne detection signals by integrating the interference signals of monochromatic light within the linewidth of weak coherent light. We obtained the final mode of the signals using MATLAB. We established an experimental system to validate the practical value of the approach in signal processing. The average deviation between the experimental and theoretical coherent frequency and length is 120.6 Hz/0.48% and −0.0072 μm/−0.06%, respectively. Compared with existing theory, the proposed method is advantageous for describing detector acquisition and has practical value in heterodyne detection analysis. The proposed method can be widely applied to the systems based on weak coherent interference.

Blind spot network (BSN) has gained increasing attention with its state-of-the-art performance in self-supervised image denoising. However, most existing BSN models are based on an unrealistic assumption of noise independence and use isotropic mask convolutions, which can lead to the loss of structural details in the denoised image. To address these limitations, we consider the spatially correlated noise and introduce directional adaptive downsampling and mask convolutions to the wavelet domain, resulting in a novel self-supervised denoising method called wavelet-adaptive BSN (WA-BSN). Specifically, we design the direction-adaptive pixel-shuffle downsamplings (PDs) and apply them to the wavelet decomposition subbands, where the spatial-correlated noise is eliminated and the inherent structure is well preserved in the wavelet domain. Then, based on the geometric direction of the wavelet subimages, we propose four shape-adaptive mask convolutions of a smaller size for each wavelet subband in WA-BSN. This enables adaptive pixel prediction within a structural neighborhood for each subband with reduced training time. Finally, total variation (TV) is added to the loss function to further preserve the edges. The results on public real-world datasets demonstrate that our method significantly outperforms existing self-supervised denoising methods and achieves great efficiency.

With the rapid advancement of the Internet of Things (IoTs) and 6G technologies, traditional terrestrial networks are becoming less capable of supporting demanding computational tasks. This limitation stems from their restricted coverage and poor adaptability to changing environmental conditions. Low earth orbit (LEO) satellite networks offer global coverage. However, existing mobile edge computing (MEC) frameworks struggle with unstable links, high decision complexity, and limited real-time performance. To overcome these challenges, this paper proposes a LEO satellite-enhanced MEC off-loading architecture based on improved multiagent deep reinforcement learning (MADRL). By integrating ground terminals, LEO satellite edge servers, cloud servers into a three-tier collaborative system, and introducing an independent Q-value mechanism, the proposed method jointly optimizes task off-loading and resource allocation in dynamic environments. This design reduces algorithm complexity and enhances decision flexibility. Experimental results show that the proposed method outperforms baseline approaches in end-to-end latency, energy efficiency, and convergence speed, while maintaining robust performance under varying satellite densities and user workloads. These results demonstrate the potential of the proposed approach for efficient task off-loading in dynamic 6G scenarios.

The development of smart transportation has raised the demand for perception and detection of vehicle targets on the road, and compared to traditional methods, such as video cameras, millimeter wave radar applications are becoming increasingly widespread. This article mainly focuses on the research of phased array millimeter wave radar, introduces the principles of beamforming and scanning, designs microstrip array antennas, and develops radar hardware RF boards using a four-chip cascade approach. Further elaborating on the data acquisition process of phased array millimeter wave radar in detecting vehicle targets, based on the obtained point cloud data, combined with the target data and point cloud characteristics under phased array millimeter wave radar, a target point cloud clustering method using the concept of region growing is proposed. Finally, through actual testing and comparison with other clustering algorithms, the superiority of this method in clustering accuracy and processing time was verified. This method can effectively solve the problem of two targets easily converging into one target when they are close, further improving the detection and tracking performance of phased array millimeter wave radar for vehicle targets.

In recent years, with the development of artificial neural networks, efficiently training models for synthetic aperture radar (SAR) image classification tasks has garnered significant attention from researchers. Particularly when dealing with datasets containing a large number of redundant samples, the selection of training samples becomes crucial for efficient model training. To address this, this paper proposes a SAR image quality evaluation-based training sample selection method, which integrates multiple indicators. First, a comprehensive SAR image quality evaluation index system is established, and then a SAR image quality evaluation model is constructed by combining representative quality evaluation metrics to guide sample selection. Experimental results demonstrate that the proposed method exhibits strong generalization capabilities on two datasets, MSTAR and OpenSarShip, effectively selecting efficient training samples.

Epilepsy is a neurological disorder characterized by frequent seizures and abnormal brain activity. It is typically diagnosed by examining electroencephalogram (EEG) recordings from epilepsy patients. Early detection and careful monitoring of children with epilepsy are crucial to preventing damaging spikes before the onset of the first seizure. Traditionally, this condition is examined manually by medical experts, a time-consuming process, especially during prolonged recordings. Therefore, an automated method for diagnosing focal (abnormal) EEG signals is essential. This study proposes an efficient model to classify and provide insights into focal and nonfocal stages. The model is based on an Inception ResNet v2 architecture pooled with a Deep Adagrad (Adaptive Gradient Descent Algorithm) Long Short-Term Memory (LSTM) network. EEG signal features are extracted using the Inception and ResNet layers, and significant features are then trained with a deep convolutional neural network (CNN) integrated with an Adagrad-optimized LSTM layer to classify focal and nonfocal EEG signals. The results demonstrate that the model achieves an impressive 99.76% accuracy in automatically detecting epilepsy abnormalities.

Real-time respiratory monitoring when sleeping is crucial for sleep apnea, chronic obstructive pulmonary disease, sleep quality assessment, and other issues related to the tracking of human health status. With the advantages of easy deployment, no wearing burden, and low privacy disclosure, recent years have witnessed a growing interest in device-free respiration monitoring leveraging radio-frequency (RF) sensing. This paper proposes a denoising diffusion probabilistic model (DDPM)-based human respiration monitoring method using an ultra-wideband (UWB) radar, where the localization calculation of the target based on the respiration-motion energy ratio, maximum ratio combining (MRC), and principal component analysis (PCA) are included for data enhancement. Moreover, a real-time sleep respiration monitoring system has been designed and implemented, which is composed of a civilian UWB radar development board, a Raspberry Pi 3B, and a PC, and extensive experiments have been carried out to validate our proposed method. Compared to the commercial respiratory tapes, our method shows that the respiratory rate estimation accuracy and the cosine similarity of respiratory waveforms can reach up to 94% and 87.9%, respectively, rendering it can be considered a viable solution for contact-free respiration monitoring for health.

In-bed posture classification plays a crucial role in health monitoring. However, existing research on classification involves a limited range of in-bed postures. Meanwhile, in classification tasks, Kolmogorov–Arnold networks (KANs), as an emerging neural network architecture, have research gaps in two areas: training strategies and architecture design. In our research, we propose Tanh-KAN, an efficient variant of KAN for in-bed posture classification. First, we validate that disabling the spline scaler not only preserves classification accuracy on the PoPu, Pmat, and SPN datasets, but also contributes to a reduction in model parameters and an increase in throughput. Second, we simplified the cubic B-spline basis functions in the original KAN using a Tanh-kernel. Compared to the original KAN, the accuracy remained stable, while the parameters were reduced by approximately 9% and the backpropagation and inference speeds increased by 42.3% and 53.9%, respectively. Experimental results further demonstrate that Tanh-KAN not only reduces model complexity and accelerates computation but also maintains high accuracy, achieving 99.6% on PoPu, 98.5% on Pmat, and 61.5% on SPN, matching the original KAN’s performance.