Sequential recommendation systems have become essential for personalized services in e-commerce and content platforms. While recent research has extended these systems with multi-modal features, existing approaches face three major challenges. First, they inadequately model fine-grained temporal interval distributions, failing to discriminate between high-frequency short intervals and low-frequency long intervals. Second, uniform fusion in the time domain leads to semantic misalignment across modalities because it ignores their inherent differences in the frequency domain. Third, rigid fusion strategies without self-supervised constraints lead to limited representation quality and semantic drift from pre-trained embeddings. To address these issues, we propose ATHWE, an A daptive T emporal Expert Routing with H ierarchical W avelet E nhancement framework. ATHWE employs exponential saturation time mapping to generate temporally adaptive embeddings. These embeddings guide a sparse mixture of experts to model multi-scale user behavior dynamics. A hierarchical wavelet decomposition with band-specific gating selectively fuses complementary frequency components across modalities. Furthermore, contrastive learning and cluster-preserving objectives preserve semantic information during multi-modal fusion. Extensive experiments on multiple datasets validate the effectiveness of our framework. Our code is available at https://github.com/lulusiyuyu/ATHWE .

Physics Guided In-line Holographic Reconstruction via Complex-Valued Multi-Scale Wavelet Decomposition and Cross-Frequency Gated Fusion

Abstract To enhance the safety of deep-buried underground tunnel construction, it is crucial to monitor microseismic signals generated by rock micro-fractures. However, these signals are often contaminated by significant noise interference, which significantly hinders accurate signal interpretation. To address this challenge, this paper proposes an improved denoising method, Matching Pursuit Variational Mode Decomposition (MP-VMD) is aimed at effectively suppressing noise in microseismic signals. This paper proposes a MP-VMD algorithm that effectively eliminates noise in microseismic signals. Firstly, the MP algorithm is employed to sparse decompose the micro-seismic signal containing noise. Secondly, the Variational Mode Decomposition (VMD) algorithm is utilized to adaptively decompose the preliminary de-noised signal to obtain a series of modal components. Finally, the MP-VMD algorithm is combined with the correlation coefficient method to select the effective modal components to reconstruct the final denoised microseismic signal. The MP-VMD algorithm is applied to the denoising in laboratory simulation and to the denoising micro seismic signals from an underground engineering tunnel, and compare with wavelet decomposition, VMD, Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD) and Local Mean Decomposition (LMD) algorithm. The experimental results show that the MP-VMD algorithm can effectively remove the noise from microseismic signals. MP-VMD achieves a 15.2% higher signal noise ratio compared to wavelet decomposition while exhibiting a 23.7% reduction in root mean square error relative to conventional VMD.

Multi-step forecasting of chlorophyll-a concentration in coastal waters through Wavelet Dense Attention Transformer model.

MDWD-KAN: Multilevel discrete wavelet decomposition with Kolmogorov–Arnold network for fall detection and activity recognition using wearable sensors

Multi-scale target-aware representation learning for fundus image enhancement.

Environmental semantic clustering-guided multimodal fusion for enhanced interpretability in methane concentration prediction.

Assessing methodological variability in wastewater surveillance: A wavelet decomposition approach.

Multimodal object detection utilizing RGB and infrared (IR) imagery has become a critical research area for unmanned aerial vehicle (UAV) surveillance applications, providing reliable perception under various lighting and environmental conditions. Nevertheless, current methods encounter three primary challenges: (1) insufficient utilization of frequency-domain properties in heterogeneous modalities, (2) restricted adaptability in crossmodal feature integration across different environmental scenarios, and (3) inadequate modeling of fine-grained spatial relationships for accurate object localization. To overcome these limitations, we introduce MFE-DETR, a novel Multimodal Feature-Enhanced Detection Transformer that achieves superior RGB-IR fusion through three complementary innovations. First, we present the Dual-Modality Enhancement Module (DMEM) with two specialized processing streams: the Haar wavelet decomposition stream (HWD-Stream) that conducts multi-resolution frequency-domain analysis to independently enhance low-frequency structural components and high-frequency textural information, and the Attention-guided Kolmogorov–Arnold Refinement Stream (AKR-Stream) that employs learnable spline-parameterized activation functions for adaptive nonlinear feature refinement. Second, we enhance the Cross-scale Channel Feature Fusion module by integrating an Adaptive Feature Fusion Module (AFAM) with complementary gating mechanisms that dynamically adjust modality contributions according to spatial informativeness. Third, we introduce the Bilinear Attention-Enhanced Detection Module (BADM) that models second-order feature interactions through factorized bilinear pooling, facilitating fine-grained crossmodal correlation analysis. Extensive experiments on the DroneVehicle benchmark show that MFE-DETR attains 78.6% mAP50 and 57.8% mAP50:95, outperforming state-of-the-art approaches by 5.3% and 3.7%, respectively. Additional evaluations on the VisDrone dataset further confirm the excellent generalization performance of our method, especially for small object detection with 18.6% APS, achieving a 1.5% improvement over existing techniques. Comprehensive ablation studies and visualizations offer detailed insights into the effectiveness of each proposed component.

The efficiency of Unmanned Aerial Vehicles (UAVs) in precision agriculture significantly depends on their design and functional adaptability. This paper presents a focused study on the lateral-directional dynamics of the HyVprop UAV, featuring a distinctive V-tail inverted configuration, tailored for advanced crop imaging tasks. Building prior wavelet transform-based methodologies that addressed sensor quality issues, this research adapts these techniques to the unique aerodynamic characteristics and challenges posed by the HyVprop’s design. The research uses multi-level wavelet decomposition to reduce sensor noise and synchronize lateral-directional signals accurately. This process is essential for precise system identification, which plays a key role in the UAV’s navigation and operational effectiveness within agricultural settings. Enhanced by the Output Error Method, which is refined to improve signal correlation specifically for the V-tail configuration, this approach is tested with simulated sensor data. The findings demonstrate significant improvements in signal quality and correlation coefficients, establishing a comprehensive framework for UAV system identification that enhances reliability and precision in crop monitoring. The paper not only confirms the methodology’s effectiveness but also highlights the specific advantages of adapting system identification techniques to unconventional UAV designs like the V-tail inverted HyVprop.

ABSTRACT In industrial equipment monitoring, how to accurately assess the fault severity of rolling bearings remains a challenging problem. Existing research mainly focuses on fault classification under Gaussian noise, but there are few studies on fault severity classification in complex noise environments. Although the Deep Residual Shrinkage Network (DRSN) can enhance feature selection through the Shrinkage mechanism, it is difficult for traditional convolution to ensure the accuracy and noise robustness at the same time when dealing with signals polluted by complex noise. In order to solve this problem, this paper proposes a Wavelet-based Residual Shrinkage Network (WRSN), which incorporates a wavelet decomposition convolution (WDConv) module. This module can decompose the STFT spectrogram into low-frequency smooth components and high-frequency detailed components, so as to extract fault features more accurately than traditional convolution. The experimental results on the HUSTbearing dataset show that WRSN successfully strikes a balance between diagnostic accuracy, noise resilience and computational efficiency.

Early detection can help slow down Parkinson’s Disease (PD), which is a progressive neurological condition characterized by immense impairment of motor and cognitive functions. EEG is useful as a non-invasive diagnostic approach because PD alters brain activity. However, the methodologies developed so far using EEG suffer from such issues as interference due to noise, high computational cost, and poor accuracy. This paper presents an efficient Memory-Efficient Hexagonal Vision Convolutional Neural Network optimized with the Snake Optimizer, known as MEHVCNNet[Formula: see text]SO, for accurate recognition of PD. This framework first removes noise from the raw EEG signals using a Wavelet Decomposition Threshold Anisotropic Filter (WDTAF), followed by DenseNet-201 that extracts robust features. The classified features obtained are then fed into MEHVCNNet, which is optimized by the use of the Snake Optimizer. This optimization in the proposed MEHVCNNet[Formula: see text]SO framework enhances accuracy while reducing computational complexity. The experimental results on three benchmark EEG datasets, San Diego, UNM, and Iowa, indicated an outstanding performance that outperformed current state-of-the-art techniques, with 99.2% accuracy, 98.4% recall, and 0.1% error rate. Offering great potential in clinical applications, these findings demonstrated the proposed MEHVCNNet[Formula: see text]SO framework as a reliable and efficient means of early identification of PD.

ABSTRACT Synthetic Aperture Radar (SAR) image change detection faces the dual challenge of speckle noise interference and complex structural changes. Most of the traditional methods are based on a single difference image (DI) or shallow network, which leads to difficulties in effectively suppressing speckle noise and extracting features. In this paper, we propose an end-to-end framework that fuses multi-operator difference images, wavelet decomposition reconstruction, and Squeeze-Excitation and Pyramid Pooling Residual Network (SEPP-ResNet). First, we apply a weighted fusion strategy to generate a weighted fusion difference image ( WFDI ). Secondly, we use the discrete wavelet transform to suppress the speckle noise in the WFDI while enhancing the edge texture information. Finally, we improve the Residual Network (ResNet) by 1) introducing the Squeeze-and-Excitation (SE) attention mechanism to dynamically adjust the channel features and enhance the discriminative features; 2) applying the Pyramid Pooling Module (PPM) for multi-scale contextual feature extraction, which captures the global information while preserving the local detail information. Extensive experiments on four real SAR datasets (Bern, Ottawa, Sulzberger, and Mexico) show that our method achieves outstanding performance. It attains Percentage of Correct Classification (PCC) values of 99.70 % , 98.72 % , 98.81 % , and 98.58 % , and Kappa Coefficients (KC) of 87.81 % , 95.22 % , 96.16 % , and 92.07 % on the respective datasets, outperforming several state-of-the-art methods.

WaveNet-SF: A hybrid network for retinal disease detection based on wavelet transform in spatial-frequency domain.

Adaptive wavelet decomposition and event-aware high-frequency modeling network for multivariate time series forecasting

Abstract This study provides a fully automated time-frequency analysis framework for recognizing radar waveforms It aims to overcome the noise sensitivity and the σ trade-off that are typical of the classical Choi–Williams distribution (CWD). Employing wavelet thresholding denoising (db8, global VisuShrink), continuous wavelet decomposition (CWD) with a fixed σ = 1.0, and first-order synchrosqueezing, the proposed methodology yields clear representations that suppress cross-terms, even when faced with significant non-Gaussian interference. Extensive evaluation of fourteen standard radar waveforms, including LFM, NLFM, Frank, Costas, P1–P4, T1–T4, and others, across various noise environments (AWGN, K-distributed colored noise, real X-band sea clutter, and α -stable impulsive noise) shows consistent superiority compared to STFT, WVD, standard CWD, FSST, and SST2. This is demonstrated by a signal-to-noise ratio (SNR) improvement of 7.2–10.2 dB and Rényi entropy values below 4.0, ranging from −15 dB to +15 dB. Synchrosqueezing shows that performance is almost unchanged whether σ is between 0.1 and 10. Using linear SVM classification on HOG features, we achieved an average accuracy of 92.6% (72.8% at −15 dB), which is 3.4 to 8.6 percentage points better than FSST/SST2. The proposed method is designed for immediate use, needing no manual adjustments.

Abstract Micro‐fluidized beds significantly enhance mass and heat transfer efficiency in gas–liquid–solid catalytic reaction systems due to their high specific surface area characteristics. However, scale effects often induce bubble coalescence and promote slugging tendencies. To address these limitations, this study utilizes a microporous distributor with apertures smaller than the particle size to compartmentalize conventional miniaturized fluidized beds, accordingly constructing a novel miniaturized confined fluidized bed. Furthermore, by employing a triple analysis framework integrating power spectral density, wavelet decomposition, and K ‐means clustering, the bubble dynamics within the confined gas–liquid–solid micro‐fluidized bed are quantitatively characterized. Based on the extracted bubble dynamics characteristics, thresholds for classing bubble motion states in the confined micro‐fluidized bed are summarized. Additionally, K ‐means clustering is utilized to objectively analyze geometric and operational parameters on bubble dynamics features, enabling the partitioning of operating gas‐velocity regimes corresponding to different bubble motion states without subjective human influence.

Abstract In recent years, hyperspectral anomaly detection utilizing spatial–spectral features has garnered considerable attention. However, a prevalent limitation among existing methods lies in their uniform processing of test points and their surrounding neighborhoods, thereby insufficiently emphasizing the test point itself. To address this, we propose a joint central attention network (JCAN) for anomaly target extraction in hyperspectral images (HSIs). First, JCAN employs an optimal clustering framework (OCF) for dimensionality reduction and spectral redundancy elimination while preserving the original dataset’s structure. The network comprises two complementary components: a central attention network (CAN) focusing on the test point, and another CAN dedicated to the background. Second, within each CAN, the dimensionally-reduced target tensor and dictionary tensor undergo single-level 3D wavelet decomposition, splitting them into eight sub-images. A 3D wavelet reconstruction rule then replaces high-frequency information with low-frequency components, effectively removing redundancy across both spatial and spectral domains. This process yields refined target and dictionary tensors—one component emphasizing the test point and the other the background. The spatial relationships captured within these two parts reflect the similarity between the emphasized point and its neighborhood. Finally, the integration of these two components enhances the separability between target and background, thereby improving detection accuracy. To evaluate JCAN’s performance, we conducted quantitative and qualitative analyses on four HSIs. Results demonstrate that JCAN outperforms seven state-of-the-art comparison methods in overall detection performance while maintaining computational efficiency.

In recent years, there has been a growing interest in the artificial intelligence (AI)-based analysis of electroencephalography (EEG) signals. This surge has made the potential of EEG more evident, both in monitoring cognitive states and in the early diagnosis of neurological disorders. This review systematically evaluates the academic literature from the past decade focusing on the processing of EEG signals through machine learning (ML), deep learning (DL), and other alternative techniques. The study compares personalized ML models (e.g., SVM, Random Forest) with wavelet decomposition–based optimized approaches and further analyzes the performance of Hilbert transform–based Convolutional Neural Network (CNN) architectures, label-free autoencoder frameworks, and multi-architecture DL systems in contemporary brain–computer interface (BCI) applications. In addition, incremental learning models based on multimodal data fusion are reviewed in the context of diagnosing disorders such as Alzheimer’s disease and epilepsy. The findings indicate that EEG–AI integration holds substantial potential for both research and clinical applications.

Abstract. This research establishes a new technique to effectively forecast climate variables, specifically Sea Surface Temperature (SS T) patterns for the Antalya region of southeast Turkey. The technique combines wavelet decomposition methods with advanced machine learning techniques to consider the many complexities that climate time series data adds to the task of forecasting. The separate wavelet components allowed us to decompose an intricate, nonstationary climate dataset into many of its temporal components that include high-frequency noise (d1), intermediate scale variables (d2), and long-term temporal trends (d3). Obviously, the disentanglement of different types of temporal variation improved the extraction of feature classes and ultimately made whatever machine learning modelling more accurate and reliable. With the one-way ANN, we examined the performance of machine learning models with wavelet pre-processing and without and reported an empirically significant reduction in error when the pipeline integrated these steps. We also demonstrated how remote sensing makes our vast area, expanding temporally and spatially, suitable for a broad range of geospatial applications. The results will provide guidance in the areas of regional climate research, emergency preparedness, and for making agricultural decisions, while showing how complementary approaches to satellite observations, utilizing signal processing techniques and machine learning can collectively contribute to improved environmental data monitoring and prediction. This research is spatially focused within the established bounds of a particular climate region and provides a detailed account of the machine learning methods used for recognition’s sake. The Wavelet decompositions (Hybrid) decreased the error percentage with a range of 10%-30% in different seasons.