Acoustic environments affected by impulsive disturbances are often governed by heavy-tailed α -stable behaviour, under which conventional active noise control (ANC) algorithms based on squared-error adaptation may exhibit unstable convergence and undesirable transient responses. Although a variety of robust ANC techniques have been proposed to improve impulsive-noise tolerance, most still rely on continuous cancellation-oriented adaptation in the presence of large outliers. This paper introduces Active Noise Blunting (ANB), an event-driven framework that moderates impulsive acoustic disturbances through bounded waveform reshaping rather than strict residual cancellation. During detected impulsive intervals, the proposed method applies nonlinear soft clipping together with finite impulse response (FIR) smoothing to generate a perceptually moderated target waveform. Anti-noise is then constructed using a deterministic difference-to-target mechanism, yielding a controlled residual response instead of aggressive cancellation-driven transients. An operator-level analysis is used to establish bounded residual behaviour of the proposed framework under α-stable excitation without relying on recursive coefficient adaptation, step-size tuning, or second-order statistical assumptions. Simulation results obtained under different impulsive-noise conditions show consistent moderation of impulsive peaks, attenuation of high-frequency transient content in the 2–3.5 kHz region, improvement in output SNR, and noticeable reduction in crest factor relative to conventional FxLMS-type ANC methods. The results suggest that perceptual waveform reshaping may provide a practical alternative for impulsive-noise mitigation in acoustic environments where maintaining stable adaptive cancellation becomes difficult under heavy-tailed excitation.
In this paper, we address a practical gap that appears frequently in signal processing; estimator comparison when parameters vary over time and no single θ dominates the operational regime. We question the blanket use of Mean Square Error (MSE) and the Cramér-Rao Lower Bound (CRLB) in classical (non Bayesian) estimation, especially in settings where the parameter of interest ( θ ) does not repeat during operation. We propose an Integrated MSE (IMSE): the average -over a parameter range- of the MSE, motivated as a Bayes risk with a uniform prior, and use this to compare estimators across a range of θ rather than at a fixed θ . We also argue that when estimating a fixed θ , classical MSE does not capture convergence speed, and thus we advocate considering a finite sample confidence requirement as a complementary performance metric.
Mutual coupling between antenna array elements causes radiation pattern distortion and performance degradation, leading to undesired signal suppression and a significant reduction in the signal-to-interference-plus-noise ratio (SINR). Adaptive array signal processing typically mitigates these distortions. However, it depends on a previous calibration of the array. Such calibration requires previous knowledge of the in-situ or realistic array manifold vector (AMV), whose estimation can be achieved through various techniques extensively explored by the antenna community over the past decades. These techniques often rely on pre-measurements or electromagnetic simulations of the array and are typically design-specific. In practical scenarios where the realistic AMV is unknown, adaptive beamforming algorithms tend to converge to distorted solutions, which deviate substantially from the performance of a calibrated system, resulting in significant SINR loss. This work addresses this challenge by proposing an algorithm that mitigates signal distortion effects without requiring prior knowledge of the realistic AMV. Also, the algorithm does not rely on a specific structure for the mutual coupling matrix (MCM), such as Toeplitz. The main idea is to combine a robust design with a technique that enforces sidelobe suppression while keeping the distortionless constraint, using additional linear constraints to the adaptive filter. As case studies, the MCMs of three uniform linear array antennas are estimated using full-wave 3D electromagnetic simulations. The results demonstrate the effectiveness of the proposed approach in significantly improving beamforming performance and SINR in scenarios where the realistic array manifold vector is unknown.
This paper considers the problem of finite-time sliding mode control for networked T-S fuzzy switched systems under a novel dynamic memory event-triggered mechanism (DMETM). The asynchronous control method is adopted where the asynchronous phenomenon is between the system mode and controller mode. The novel DMETM is introduced to alleviate the burden of data transmission in the communication channel, which not only transmited the current packets but also the past utilized packets, it can be adjusted according to the error of the system state. The asynchronous event-triggered finite-time sliding mode control of the closed-loop system are proved. Finally, a numerical example and a single-link rigid robot system model are presented to demonstrate the effectiveness of the proposed approach.
In this paper, an adaptive sliding mode control (SMC) scheme with fractional integral design is proposed for Euler-Lagrange system (ELS) subject to disturbances. The adaptive predefined-time fractional integral SMC (APISMC) scheme is developed based on a fractional-order integral nonsingular terminal sliding mode, incorporating a predefined-time stability framework to guarantee rapid and predictable convergence. An adaptive scheme is employed to adjust the control gain online in response to system uncertainties and external disturbances without requiring prior knowledge of their bounds. Lyapunov-based analysis is conducted to prove the predefined-time convergence of the tracking errors and stability of the closed-loop system. Numerical simulations on a two-link manipulator demonstrate the effectiveness of the proposed control strategy in achieving accurate trajectory tracking and strong robustness to disturbances.
Accurate detection of RR intervals in ECG signals is essential for reliable cardiac health assessment and early arrhythmia monitoring. The classical methods, specifically the Pan–Tompkins algorithm, suffer from noise sensitivity, increased false-positive detections, and difficulty in identifying low-amplitude R-peaks, particularly in real-time and long-duration ECG analysis. In order to overcome the classical limitations and by utilizing the advantage of quantum-enhanced signal analysis, we propose an AI-inspired quantum hybrid model that integrates Quantum Principal Component Analysis and Grover’s Algorithm for the detection of RR intervals. The proposed model is based on mapping the pipelined stages of classical Pan-Tompkins through quantum processing stages as amplitude encoding of ECG segments, QPCA-based noise reduction and feature compression, and based on amplitude amplification, the Grover's search performs structured and efficient R-peak detection, followed by RR intervals computed classically at the last. Our experimental results on a Synthetic Dataset and the MIT-BIH Arrhythmia Database have been used to validate that the proposed model, and for the real dataset a model achieved an accuracy of 98.8
In image denoising, neural network models based on the UNet architecture have become the mainstream due to their excellent performance and strong reconstruction capabilities. However, their heavy reliance on paired data and high computational costs limit their deployment and application on embedded devices. This paper adopts the Blind2Unblind (B2U) self-supervised training framework and proposes a lightweight U-Net-based denoising network with improved computational efficiency. This network is based on the UNet architecture, enhancing multi-scale denoising capabilities through multiple convolutional blocks and branch feature fusion, and introducing global attention by combining depthwise separable convolution, short skip connections, and GCBlock. To achieve lightweight, the softmax in GCBlock is replaced with L^-1ReLU to reduce computational costs. Experiments are conducted on infrared synthetic noise, SIDD, and synthetic datasets with Poisson/Gaussian noise. The results show that the proposed network significantly reduces model complexity while maintaining competitive denoising performance. Additional comparisons with recent representative denoising models further demonstrate that the proposed method offers a favorable trade-off between denoising quality and computational complexity, making it a denoising model with high computational efficiency and the capability of deploying on edge devices.
A feature fusion network integrating a convolutional neural network (CNN) and a Swin Transformer is proposed to enhance representation learning for acoustic scene classification (ASC). Log-mel filter bank energies extracted from audio recordings are first fed into both the CNN branch and the Swin Transformer branch. A feature fusion block is then introduced to serve as a bidirectional bridge between the two branches, enabling effective information exchange. Specifically, the global contextual representations learned by the Swin Transformer branch are conveyed to the CNN branch, while the local features extracted by the CNN branch are progressively transmitted to the patch embedding module to enhance the local representation capability of the Swin Transformer branch. Through this mechanism, the model achieves effective fusion of local and global information and promotes complementary learning between the two branches. Finally, the fused features are aggregated and passed through a feed-forward network (FFN) layer to perform ASC. Experimental results demonstrate that the proposed method achieves accuracies of 85.8
A compact tunable band-pass filter designed for EEG, ECG, and EMG signal processing is introduced, featuring low power consumption and employing class-AB amplifiers along with high-value programmable active pseudo-resistors. These components are biased using the quasi-floating-gate technique with capacitive gate-voltage averaging, thereby reducing distortion. The filter, implemented in a three-stage design using 0.5 μ m CMOS technology and simulated with HSPICE, covers a tunable frequency range from 0.5 Hz to 3 kHz, operates on a bias current of 25 nA, and achieves a total harmonic distortion (THD) of 0.56
Motion estimation (ME) plays a very significant role in any standard video codec. It is one of the most compute and bandwidth intensive operation in modern vide codes. The complexity of motion estimation algorithms often causes memory bandwidth bottlenecks. This paper introduces a co-deign of search algorithm, dataflow and hardware architecture to alleviate memory bottleneck issues. We present a compact inter-level systolic architecture for half-pixel motion estimation that preserves the low-complexity nature of iterative search while recovering much of the data reuse needed for high-throughput implementation. The proposed design combines a 9-point iterative search strategy with a 3 × 3 array of processing elements, queue-assisted data alignment, and a broadcast-oriented data-flow that supplies 2 × 2 reference pixels per cycle. This organization enables efficient execution of large, small, and fractional search stages using the same processing array, while significantly reducing external memory traffic and on-chip bandwidth pressure relative to straightforward iterative implementations. The architecture requires only nine processing elements and a structured current-frame buffer organization to sustain high throughput. Synthesized in UMC 180 nm technology, the design occupies about 49.5 K gates and, at 145 MHz, supports real-time processing of 1080p video at 30 frames/s. Experimental evaluation shows that the proposed search strategy achieves an average PSNR degradation of only 0.44 dB relative to full search with half-pixel refinement, demonstrating that memory-bandwidth savings and compact hardware can be achieved with limited loss in coding performance.
Among artificial neural network (ANN) architectures, the radial basis function neural network is particularly suitable for control applications due to its good generalization capability and simple structure, which reduces computational burden during learning. This paper presents an adaptive filter-based backstepping controller for nonlinear singularly perturbed systems using RBFNNs. High-gain filters are introduced to alleviate the complexity of conventional backstepping designs, while a high-gain disturbance observer (HGDO) is employed to estimate and mitigate the effects of unknown system states. The proposed control scheme efficiently addresses both fast and slow system dynamics, ensuring stability and robustness against perturbations. The effectiveness of the method is validated on an autonomous differential equation system and a robotic manipulator, demonstrating improved tracking accuracy compared to existing approaches.
Speech Emotion Recognition (SER) plays a major role in enhancing human–computer interaction and affective computing but still unimodal methods often struggle with poor generalization and robustness in noisy and cross-domain scenarios in SER. To address this limitation, this paper introduces a multimodal SER system that combines both acoustic and linguistic modalities for richer emotional representations. The acoustic feature extraction captures the prosodic, spectral, and deep acoustic representations yielded by the Multi-Layer Emotion Fusion Transformer (MLEFT). In contrast, the linguistic features are extracted from a Generative Pre-trained Transformer-2(GPT-2) model based on Whisper-transcribed transcripts. The features are then combined using a new cross-attention mechanism, followed by a Bi-directional Gated Recurrent Unit (BiGRU) and multi-head attention layer to capture temporal dependencies and rich cross-modal interactions. Empirical evaluation was conducted across a wide range of datasets, including RAVDESS, SAVEE, TESS, IEMOCAP, MELD, and in-house English and Hindi corpora. Experimental results show that the holistic fusion of prosodic, spectral, and deep acoustic features with contextual linguistic embeddings significantly enhances the recognition performance. Notably, the present cross-attention fusion approach, coupled with BiGRU, multi-head attention, and FCN classification pipeline, repeatedly produced state-of-the-art weighted and unweighted accuracies. The results confirm the excellence of the framework’s performance and its potential for stable, multilingual SER applications in complex real-world settings.
This study investigates the bumpless transfer control problem under actuator saturation for a class of switched positive linear systems for the first time. A unified definition of bumpless transfer performance under actuator saturation is put forward to suppress the control bump at switching instant, whether the actuator remains saturated or unsaturated. By virtue of convex hull technique and multiple linear co-positive Lyapunov functions method, a sufficient condition in terms of linear programming problem is formulated to obtain the positivity, exponential stability, and bumpless transfer performance under actuator saturated. The specific form of the controller is cleverly designed by applying the dual system theory. Additionally, an algorithm is presented to maximize the estimation of the attraction domain and optimize the bumpless transfer performance level. Ultimately, two simulation examples are provided to demonstrate that the proposed method can achieve satisfactory bumpless transfer performance, reduce the occurrence time of actuator saturation, and expand the estimation of the attraction domain.
Wearable devices are essential for the long-term monitoring of cardiovascular diseases (CVDs). However, the diagnosis of CVDs typically relies on large-scale neural networks, which, due to their significant hardware overhead and high energy consumption, pose challenges for deployment on wearable devices. This work proposes a low-power solution for wearable heart sound diagnosis using a lightweight convolutional neural network (CNN) design method, based on algorithm-hardware co-design. Quantization-aware training (QAT), along with weight/activation quantization and operator fusion, simplifies computations with minimal accuracy loss. A hybrid dataflow and dynamic multiplier-accumulator (HDF-DMAC) architecture is presented, combining weight stationary (WS) and output stationary (OS) dataflows with dynamic MAC reuse. This architecture reduces cache size, minimizes memory accesses, and enhances logic utilization. Field-programmable gate array (FPGA) verification shows that the proposed method decreases the model’s computational workload, reducing floating-point operations (FLOPs) from 0.594 M to 0.147 M and compressing weight storage from 80.01 KB to 20.18 KB, while maintaining a diagnostic accuracy of 93.74
The speech signal can be affected by the background noise and acoustic echo, which make the speech comprehension hard in the case of the conference room. In order to solve this problem, a novel real-time echo canceller in the conference room environment is proposed. The proposed method is designed with help of Neuron pruning Approximate Multiplier based Spiking Neural Network (NPAM-SNN). The audio recorded is initially divided into frames, and it is filtered, and it is processed assisted with the aid of the SNN which is developed through Neuron Pruning approximate multiplier. The input signal is sorted into interference sources and target using this network. The signal of the interference is then operated by Approximate Least Mean square(LMS) Filter that assists in operating with a tunable LMS step size and a defined Finite Impulse Response Filter(FIR) length. Finally, the signal that is wanted and the signal that is unwanted are added to form a better output, which contributes to the better state of being able to hear the speech and be understood in noisy environments. The proposed method achieves an average SNR of 7.00 dB, where 3.67 dB for C-SAR and 2.37 dB for RASP-AIOT, resulting in relative improvements of 33.45
Power electronics circuits are used to shape desired voltage or current signals. Widely utilized in power electronics, a single-phase active harmonic filter circuit plays a crucial role in improving power efficiency. There are various types of memristor emulators in the literature. However, only a few of them are power electronics-based or make use of switching to obtain memristive behavior. In this research, a new kind of memristor emulator is demonstrated using a single-phase active harmonic filter. To date, no active harmonic filter-based memristor emulators have been reported in the literature. This circuit utilizes a MOSFET-based synchronous rectifier placed between the source and a load resistor. The semiconductor switches of the synchronous rectifier are controlled to achieve the intended memristive waveforms by a microcontroller. The circuit that was designed is analyzed through both simulations and experiments. Experimental and simulation results are presented for the HP and Biolek models to demonstrate the emulator performance and model adaptability. It is shown that the suggested design operates as a memristor emulator, and it mimics the memristor models. It has good performance and can easily simulate different types of memristor models by modifying its software.
Image moments have been widely applied in watermarking techniques nowadays. However, most existing image moment-based watermarking algorithms still have some limitations such as insufficient robustness, high computational complexity, and small amount of watermark information. This paper proposes a zero-watermarking scheme based on the all phase discrete cosine biorthogonal transform (APDCBT) and quaternion fractional-order Jacobi-Fourier moments (QFJFMs) to protect color images. Firstly, the proposed scheme subdivides the original color image’s R, G, and B components into non-overlapping blocks, and APDCBT is performed on each block. Then, the DC coefficients of all blocks after APDCBT are combined into a stable color low-frequency segment, the mixed low-order moment feature (MLMF) of variable-parameter QFJFMs for the segment is computed and used to construct a binary feature image with the help of multiple different chaotic systems and the semi-tensor product. Finally, the binary feature image is XORed with the encrypted watermark to generate the zero watermark. Experimental results show that the proposed scheme provides outstanding robustness against common image processing attacks and geometric attacks. In the comparative test, the average BER of the proposed algorithm for multiple attacks is only 0.0071, showing more excellent robustness. At the same time, the zero watermark of size 64 × 64 is generated or detected just need about 1 s, which fully shows the efficiency of the proposed algorithm.
The steering vector optimization and interference-plus-noise covariance matrix (INCM) estimation are two pivotal factors that significantly influence the performance of adaptive beamforming systems. Standard adaptive beamformers suffer from performance degradation due to covariance matrix mismatches and steering vector (SV) errors. To address these issues while mitigating the high computational cost of convex optimization, this study proposes an improved covariance matrix reconstruction method and an innovative approach for steering vector optimization of the desired signal. Initially, this method adopts a threshold comparison of the Capon spatial power spectrum to ascertain whether a specific direction is associated with interference or noise, and subsequently applies the appropriate interference or noise reconstruction algorithm. Thereafter, it employs the projection of the steering vector onto the noise subspace, in conjunction with an iterative search strategy that maximizes the output signal-to-interference-plus-noise ratio (SINR), to achieve a precise estimation of the steering vector for the desired signal. The appropriate weight vector is then generated within the computational framework. Experimental results demonstrate that the proposed method achieves better overall performance under multiple conditions. Notably, by circumventing computationally expensive quadratic constrained quadratic programming (QCQP), the proposed algorithm reduces execution time by approximately 87
The IEEE 802.11ax wireless communication standard supports multiple frequency bands and high data throughput, which increases the susceptibility of receiver front-end circuits to interference and nonlinear distortion. Therefore, the low-noise amplifier (LNA) in the RF receiver must maintain a balanced trade-off among gain, noise figure, and linearity while operating over a wide bandwidth. In this work, an inductorless wideband LNA based on a single-to-double-ended Common-Gate–Common-Source (CG–CS) architecture is proposed to achieve compact implementation and improved RF performance. To enhance the linearity of the amplifier, a modified Complementary Derivative Superposition (CDS) technique is incorporated, which mitigates third-order nonlinear distortion by compensating the nonlinear transconductance components of the MOS devices. The inductorless topology also improves circuit integration and reduces chip area compared with conventional inductively matched LNAs. The proposed LNA is designed using UMC 65-nm CMOS technology, and post-layout simulations are performed using the Cadence Virtuoso environment. Simulation results demonstrate that the proposed design achieves a maximum gain of 17.19 dB at 2.4 GHz with an operating frequency range from 1.7 to 6.6 GHz. The amplifier exhibits an IIP3 of 2.81 dBm, indicating improved linearity due to the CDS-assisted architecture. The input return loss is maintained below − 10 dB across the operating band, while the minimum noise figure is 3.24 dB. The proposed LNA consumes 12.65 mW from a 1.2 V supply, making it suitable for compact and power-efficient RF receiver front-end applications such as Wi-Fi 6/IEEE 802.11ax systems.
Epileptic seizure prediction remains a major challenge in neuroscience, with significant implications for patient safety and quality of life. This study introduces a novel hybrid deep learning architecture that combines a three-dimensional Convolutional Neural Network (3D-CNN) with a Transformer encoder to predict and classify epileptic seizures from electroencephalogram (EEG) signals transformed via Short-Time Fourier Transform (STFT). The 3D-CNN component efficiently extracts local spatio-spectro-temporal features from multichannel EEG data, while the Transformer captures long-range temporal dependencies essential for modeling preictal brain dynamics. The proposed model was trained and evaluated on the CHB-MIT scalp EEG dataset and compared to a standalone Transformer-based approach. Experimental results demonstrate that the hybrid 3D-CNN–Transformer significantly outperforms the baseline, achieving an accuracy of 98.4
