Fusion Strategy Research Articles

Clustering scRNA-seq data with the cross-view collaborative information fusion strategy.

Single-cell RNA sequencing (scRNA-seq) technology has revolutionized biological research by enabling high-throughput, cellular-resolution gene expression profiling. A critical step in scRNA-seq data analysis is cell clustering, which supports downstream analyses. However, the high-dimensional and sparse nature of scRNA-seq data poses significant challenges to existing clustering methods. Furthermore, integrating gene expression information with potential cell structure data remains largely unexplored. Here, we present scCFIB, a novel information bottleneck (IB)-based clustering algorithm that leverages the power of IB for efficient processing of high-dimensional sparse data and incorporates a cross-view fusion strategy to achieve robust cell clustering. scCFIB constructs a multi-feature space by establishing two distinct views from the original features. We then formulate the cell clustering problem as a target loss function within the IB framework, employing a collaborative information fusion strategy. To further optimize scCFIB's performance, we introduce a novel sequential optimization approach through an iterative process. Benchmarking against established methods on diverse scRNA-seq datasets demonstrates that scCFIB achieves superior performance in scRNA-seq data clustering tasks. Availability: the source code is publicly available on GitHub: https://github.com/weixiaojiao/scCFIB.

DGCL: dual-graph neural networks contrastive learning for molecular property prediction.

In this paper, we propose DGCL, a dual-graph neural networks (GNNs)-based contrastive learning (CL) integrated with mixed molecular fingerprints (MFPs) for molecular property prediction. The DGCL-MFP method contains two stages. In the first pretraining stage, we utilize two different GNNs as encoders to construct CL, rather than using the method of generating enhanced graphs as before. Precisely, DGCL aggregates and enhances features of the same molecule by the Graph Isomorphism Network and the Graph Attention Network, with representations extracted from the same molecule serving as positive samples, and others marked as negative ones. In the downstream tasks training stage, features extracted from the two above pretrained graph networks and the meticulously selected MFPs are concated together to predict molecular properties. Our experiments show that DGCL enhances the performance of existing GNNs by achieving or surpassing the state-of-the-art self-supervised learning models on multiple benchmark datasets. Specifically, DGCL increases the average performance of classification tasks by 3.73$\%$ and improves the performance of regression task Lipo by 0.126. Through ablation studies, we validate the impact of network fusion strategies and MFPs on model performance. In addition, DGCL's predictive performance is further enhanced by weighting different molecular features based on the Extended Connectivity Fingerprint. The code and datasets of DGCL will be made publicly available.

Firefighting Robot Extinguishment Decision‐Making Based on Visual Guidance: A Novel Attention and Scale U‐Net Model and Genetic Algorithm

ABSTRACTAt present, rescue firefighting relies mainly on firefighting robots, and robots with perception and decision‐making functions are the key elements for achieving intelligent firefighting. However, traditional firefighting robots often lack the ability for autonomous perception and decision‐making when extinguishing multiple fire sources, leading to low rescue efficiency and increased risk for rescue personnel, especially when making firefighting decisions in extreme fire scenes, which poses a challenge. To effectively handle firefighting tasks and ensure operational efficiency, a robot firefighting decision‐making method based on drone visual guidance is proposed. First, we introduce a novel Attention and Scale U‐Net (ASUNet) model to accurately capture crucial target information, including fire location and size, in a fire scene. The ASUNet model adopts an effective multiscale fusion strategy and attention mechanism to enhance the model's performance. Subsequently, based on the results of the ASUNet model, through pixel segmentation clustering and a genetic optimization algorithm, we obtain the robot's firefighting decision results, thereby guiding the robot to carry out firefighting operations systematically. Finally, through numerical experiments, it is verified that the proposed ASUNet model is superior and effective, as the model can perceive important information in a fire scene and extract it well. The use of improved genetic optimization can further accelerate algorithm convergence. To our knowledge, this study is the first to use drone‐based monocular vision guidance for firefighting decision‐making, providing significant engineering value.

Recover then aggregate: unified cross-modal deep clustering with global structural information for single-cell data.

Single-cell cross-modal joint clustering has been extensively utilized to investigate the tumor microenvironment. Although numerous approaches have been suggested, accurate clustering remains the main challenge. First, the gene expression matrix frequently contains numerous missing values due to measurement limitations. The majority of existing clustering methods treat it as a typical multi-modal dataset without further processing. Few methods conduct recovery before clustering and do not sufficiently engage with the underlying research, leading to suboptimal outcomes. Additionally, the existing cross-modal information fusion strategy does not ensure consistency of representations across different modes, potentially leading to the integration of conflicting information, which could degrade performance. To address these challenges, we propose the 'Recover then Aggregate' strategy and introduce the Unified Cross-Modal Deep Clustering model. Specifically, we have developed a data augmentation technique based on neighborhood similarity, iteratively imposing rank constraints on the Laplacian matrix, thus updating the similarity matrix and recovering dropout events. Concurrently, we integrate cross-modal features and employ contrastive learning to align modality-specific representations with consistent ones, enhancing the effective integration of diverse modal information. Comprehensive experiments on five real-world multi-modal datasets have demonstrated this method's superior effectiveness in single-cell clustering tasks.

A Wide-Bandwidth Inexpensive Current Sensor Based on the Signal Fusion of Tunneling Magnetoresistance and a Current Transformer.

In technology and industrial production, many applications require wide-bandwidth current measurements. In this paper, a signal fusion scheme for a current sensor comprising tunneling magnetoresistance and a current transformer is proposed, achieving a flat frequency response in the DC to MHz range. The measurement principles in different cases of the scheme are introduced, and the total transfer function of the entire scheme is derived by analyzing each section separately. Furthermore, the feasibility and selected parameters of the scheme are verified through a systematic simulation utilizing the MATLAB software. Based on the proposed scheme, a group of principal prototypes are built to experimentally evaluate the bandwidth, amplitude and phase flatness, accuracy, sensitivity, and impulse response. The relative amplitude variation in the passband of the fusion sensor is less than 4%, and the estimated bandwidth of the fusion sensor is close to 17 MHz. The accuracy is better than 0.6%, even when measuring the current at 1 MHz, and the relative standard deviation is 5% when measuring the impulse signal. The sensors developed using this scheme, with a low financial cost, have advantages in many wide-bandwidth current measuring scenarios.

Multimodal emotion recognition based on a fusion of audiovisual information with temporal dynamics

AbstractIn the Human-Machine Interactions (HMI) landscape, understanding user emotions is pivotal for elevating user experiences. This paper explores Facial Expression Recognition (FER) within HMI, employing a distinctive multimodal approach that integrates visual and auditory information. Recognizing the dynamic nature of HMI, where situations evolve, this study emphasizes continuous emotion analysis. This work assesses various fusion strategies that involve the addition to the main network of different architectures, such as autoencoders (AE) or an Embracement module, to combine the information of multiple biometric cues. In addition to the multimodal approach, this paper introduces a new architecture that prioritizes temporal dynamics by incorporating Long Short-Term Memory (LSTM) networks. The final proposal, which integrates different multimodal approaches with the temporal focus capabilities of the LSTM architecture, was tested across three public datasets: RAVDESS, SAVEE, and CREMA-D. It showcased state-of-the-art accuracy of 88.11%, 86.75%, and 80.27%, respectively, and outperformed other existing approaches.

Efficient spectrum allocation by heterogeneous automated frequency coordination network within 6 GHz band

This work proposes a heterogeneous automated frequency coordination network (HAFCN) to enhance reliability and enable dynamic spectrum allocation for unlicensed Wi-Fi devices (UWDs) within the 6 GHz band. This process relies heavily on spectrum sensing within the HAFCN. However, the widespread implementation of spectrum sensing by multiple UWDs in an automated frequency coordination (AFC) network using conventional fusion schemes poses computational challenges at the AFC provider. In response to this challenge, we present a selective soft-information (SSI) fusion scheme for the proposed HAFCN. First, we present generic mathematical expressions of false-alarm and missed detection probabilities for the HAFCN using an SSI fusion scheme. Second, a generalized AFC SSI fusion problem (GASFP) is formulated to minimize the system’s total error probability. Further, to mitigate the AFC provider’s overhead in solving the GASFP, this work presents the swift-sensing problem, determining the minimum antennas at UWD required to achieve a desired total error probability. Finally, comparative numerical results demonstrate that the HAFCN with the SSI fusion scheme shows a significant performance improvement over conventional fusion schemes in terms of the total error probability.

Sensorless control in full speed domain of interior permanent magnet synchronous motor based on hybrid observer

Aiming at the problem that an Interior Permanent Magnet Synchronous Motor (IPMSM) cannot be smoothly started in zero-low speed range and smoothly transitioned to medium-high speed range by a single observer. This paper proposes a full-speed range control algorithm based on the fusion of pulsating high-frequency injection and back electromotive force (EMF) position error information. In the low-speed range or at start-up, a square-wave high-frequency signal is injected, and the obtained high-frequency current signal is processed to obtain the rotor position error information. The phase shift due to the introduction of a filter is reduced, which improves the control bandwidth and reduces the noise. To ensure smooth switching of the observer, the observer uses a dual second-order generalized integrator module to output the angular frequency in the low-speed range. A higher-order sliding mode observer based on an inverse EMF model obtains rotor position error information at high speeds. During switching, the rotor position information is processed by a fusion strategy, and the obtained hybrid information is fed into the system to improve the stability of the motor operation. A 0.2 kW IPMSM position sensorless vector control system verifies the algorithm’s accuracy.

Consensus SE(3)-Constrained Extended Kalman Filter for Close Proximity Orbital Relative Pose Estimation

In this paper, a recently proposed SE(3)-constrained extended Kalman filter (EKF) is extended to formulate a strategy for relative orbit estimation in a space-based sensor network. The resulting consensus SE(3)-constrained EKF utilizes space-based sensor fusion and is applied to the problem of spacecraft proximity operations and formation flying. The proposed filter allows for the state (i.e., pose and velocities) estimation of the deputy satellite while accounting for measurement error statistics using the rotation matrix to represent attitude. Via a comparison with a conventional filter in the literature, it is shown that the use of the proposed consensus SE(3)-constrained EKF can improve the convergence performance of the existing filter for satellite formation flying. Moreover, the benefits of faster convergence and consensus speed by using communication networks with more connections are illustrated to show the significance of the proposed sensor fusion strategy in spacecraft proximity operations.

Brain tumor segmentation by combining MultiEncoder UNet with wavelet fusion.

Accurate segmentation of brain tumors from multimodal magnetic resonance imaging (MRI) holds significant importance in clinical diagnosis and surgical intervention, while current deep learning methods cope with situations of multimodal MRI by an early fusion strategy that implicitly assumes that the modal relationships are linear, which tends to ignore the complementary information between modalities, negatively impacting the model's performance. Meanwhile, long-range relationships between voxels cannot be captured due to the localized character of the convolution procedure. Aiming at this problem, we propose a multimodal segmentation network based on a late fusion strategy that employs multiple encoders and a decoder for the segmentation of brain tumors. Each encoder is specialized for processing distinct modalities. Notably, our framework includes a feature fusion module based on a 3D discrete wavelet transform aimed at extracting complementary features among the encoders. Additionally, a 3D global context-aware module was introduced to capture the long-range dependencies of tumor voxels at a high level of features. The decoder combines fused and global features to enhance the network's segmentation performance. Our proposed model is experimented on the publicly available BraTS2018 and BraTS2021 datasets. The experimental results show competitiveness with state-of-the-art methods. The results demonstrate that our approach applies a novel concept for multimodal fusion within deep neural networks and delivers more accurate and promising brain tumor segmentation, with the potential to assist physicians in diagnosis.

Development and Validation of a One‐Dimensional Finite Difference Simulation Scheme for Polymer Laser Powder Bed Fusion with Application to the Effect of the Inter Layer Time

The division of the polymer laser powder bed fusion process polymers (PBF‐LB/P) into temporal regimes originating from laser motion, thermal diffusion, viscous flow, crystallization kinetics, and powder application is exploited by considering only the building direction. The reduction of dimensionality enables fast simulations which are used to investigate the influence of the inter layer time on the final part density. Numerical simulation results are validated by experiments using polyamide 12 (PA 12) with good agreement in terms of part density. It is shown that an inter layer time of 90 s leads to nearly dense PA 12 parts while a time of 45 s leads to less dense parts. The cooling effect of the applied powder is identified as a cause for insufficient densification of the previous layer. The simulation tool is quantitatively validated against experimental results for the surface temperature of PA 12 as function of scan speed and hatch distance, for the melt pool depth of polyamide 6 as function of scan speed and for the melt pool depth of polyetherketoneketone as function of laser power. The presented simulation method enables rapid process parameter adjustment for new polymer materials in the PBF‐LB/P process.

Optimized clustering-based fusion for skin lesion image classification: Leveraging marine predators algorithm

This manuscript presents a comprehensive approach to enhance the accuracy of skin lesion image classification based on the HAM10000 and BCN20000 datasets. Building on prior feature fusion models, this research introduces an optimized cluster-based fusion approach to address limitations observed in our previous methods. The study proposes two novel feature fusion strategies, KFS-MPA (using K-means) and DFS-MPA (using DBSCAN), for skin lesion classification. These approaches leverage optimized clustering-based deep feature fusion and the marine predator algorithm (MPA). Ten fused feature sets are evaluated using three classifiers on both datasets, and their performance is compared in terms of dimensionality reduction and accuracy improvement. The results consistently demonstrate that the DFS-MPA approach outperforms KFS-MPA and other compared fusion methods, achieving notable dimensionality reduction and the highest accuracy levels. ROC-AUC curves further support the superiority of DFS-MPA, highlighting its exceptional discriminative capabilities. Five-fold cross-validation tests and a comparison with the previously proposed feature fusion method (FOWFS-AJS) are performed, confirming the effectiveness of DFS-MPA in enhancing classification performance. The statistical validation based on the Friedman test and Bonferroni-Dunn test also supports DFS-MPA as a promising approach for skin lesion classification among the evaluated feature fusion methods. These findings emphasize the significance of optimized cluster-based deep feature fusion in skin lesion classification and establish DFS-MPA as the preferred choice for feature fusion in this study.

Enhancing alzheimer’s diagnosis through optimized brain lesion classification in MRI with attention-driven grid feature fusion

This paper explores cognitive interface technology, aiming to tackle current challenges and shed light on the prospects of brain-computer interfaces (BCIs). It provides a comprehensive examination of their transformative impact on medical technology and patient well-being. Specifically, this study contributes to addressing challenges in classifying brain lesion images arising from the complex nature of lesions and limitations of traditional deep learning approaches. It introduces advanced feature fusion models that leverage deep learning algorithms, including the African vulture optimization (AVO) algorithm. These models integrate informative features from multiple pre-trained networks and employ innovative fusion techniques, including the attention-driven grid feature fusion (ADGFF) model. The ADGFF model incorporates an attention mechanism based on the optimized weights obtained using AVO. The objective is to improve the overall accuracy by providing fine-grained control over different regions of interest in the input image through a grid-based technique. This grid-based technique divides the image into vertical and horizontal grids, simplifying the exemplar feature generation process without compromising performance. Experimental results demonstrate that the proposed feature fusion strategies consistently outperform individual pre-trained models in terms of accuracy, sensitivity, specificity, and F1-score. The optimized feature fusion strategies, particularly the GRU-ADGFF model, further enhance classification performance, outperforming CNN and RNN classifiers. The learning progress analysis shows convergence, indicating the effectiveness of the feature fusion strategies in capturing lesion patterns. AUC-ROC curves highlight the superior discriminatory capabilities of the ADGFF-AVO strategy. Five-fold cross-validation is employed to assess the performance of the proposed models, demonstrating their accuracy, and few other accuracy-based measures. The GRU-ADGFF model optimized with AVO consistently achieves high accuracy, sensitivity, and AUC values, demonstrating its effectiveness and generalization capability. The GRU-ADGFF model also outperforms the majority voting ensemble technique in terms of accuracy and discriminative ability. Additionally, execution time analysis reveals good scalability and resource utilization of the proposed models. The Friedman rank test confirms significant differences in classifier performance, with the GRU-ADGFF model emerging as the top-performing method across different feature fusion strategies and optimization algorithms.

CTAFFNet: CNN–Transformer Adaptive Feature Fusion Object Detection Algorithm for Complex Traffic Scenarios

As the core technology of an environmental perception system, object detection has received more and more attention and has become a hot research direction for intelligent driving vehicles. The CNN–Transformer hybrid model has poor generalization ability, making it difficult to meet the detection requirements for small objects in complex scenes. We propose a novel convolutional neural network (CNN)–Transformer Adaptive Feature Fusion Network (CTAFFNet) for object detection. First, we design a Local–Global Feature Fusion unit known as the Convolutional Transformation Adaptive Fusion Kernel (CTAFFK), which is integrated into CTAFFNet. The CTAFFK kernel utilizes two branches, namely CNN and Transformer, to extract local and global features from the image, and adaptively fuses the features from both branches. Subsequently, we develop an adaptive feature fusion strategy that combines local high-frequency and global low-frequency features to obtain comprehensive feature information. Finally, CTAFFNet employs an encoder–decoder structure to facilitate the flow of fused local–global information between different stages, ensuring the model’s generalization capabilities. Results from the experiment conducted on the large and challenging KITTI dataset demonstrate the effectiveness and efficiency of the proposed network. Compared with other mainstream networks, it achieves an average precision of 91.17%, particularly excelling in the detection of small objects at longer distances with a remarkable 70.18% accuracy, while also providing real-time detection capabilities.

An hetero-modal deep learning framework for medical image synthesis applied to contrast and non-contrast MRI

Some pathologies such as cancer and dementia require multiple imaging modalities to fully diagnose and assess the extent of the disease. Magnetic resonance imaging offers this kind of polyvalence, but examinations take time and can require contrast agent injection. The flexible synthesis of these imaging sequences based on the available ones for a given patient could help reduce scan times or circumvent the need for contrast agent injection. In this work, we propose a deep learning architecture that can perform the synthesis of all missing imaging sequences from any subset of available images. The network is trained adversarially, with the generator consisting of parallel 3D U-Net encoders and decoders that optimally combines their multi-resolution representations with a fusion operation learned by an attention network trained conjointly with the generator network. We compare our synthesis performance with 3D networks using other types of fusion and a comparable number of trainable parameters, such as the mean/variance fusion. In all synthesis scenarios except one, the synthesis performance of the network using attention-guided fusion was better than the other fusion schemes. We also inspect the encoded representations and the attention network outputs to gain insights into the synthesis process, and uncover desirable behaviors such as prioritization of specific modalities, flexible construction of the representation when important modalities are missing, and modalities being selected in regions where they carry sequence-specific information. This work suggests that a better construction of the latent representation space in hetero-modal networks can be achieved by using an attention network.

Four‐Membered Ring‐Embedded Cycloarene Enabling Anti‐Aromaticity and Ultra‐Narrowband Emission

AbstractThe synthesis of fully fused π‐conjugated cycloarenes embedded with nonbenzenoid aromatics is challenging. In this work, the first example of four‐membered ring‐embedded cycloarene (MF2) was designed and synthesized in single‐crystal form by macrocyclization and ring fusion strategies. For comparison, single bond‐linked chiral macrocycle MS2 without two fused four‐membered rings and its linear‐shaped polycyclic benzenoid monomer L1 were also synthesized. The pronounced anti‐aromaticity of four‐membered rings significantly adjusts the electronic structures and photophysical properties of cycloarene, resulting in an enhancement of the photoluminescence quantum yield (PLQY) from 10.66 % and 10.74 % for L1 and MS2, respectively, to 54.05 % for MF2, which is the highest PLQY among the reported cycloarenes. Notably, owing to the embedded anti‐aromatic four‐membered rings that reduce structural displacements, MF2 exhibits an ultra‐narrowband emission with a single‐digit full‐width at half‐maximum (FWHM) of only 7 nm (0.038 eV), which sets a new record among all reported organic narrowband luminescent molecules, and represents the first example of ultra‐narrowband emission in conventional polycyclic aromatic hydrocarbons (PAHs) devoid of heteroatoms.

Granular correlation-based label-specific feature augmentation for multi-label classification

Multi-label classification is an extension of single-label classification with generations of multi-output for unseen instances. Label correlation is an essential component in constructing multi-label classifiers. How to optimize the representation of label correlation while preserving the semantics of label-specific remains an uncertain issue. Instead of estimating label correlation by a holistic feature representation, we present an augmented label correlation model by generating multi-granularity label-specific features. Firstly, we devise a mixture distance measure to characterize the closeness of an instance by weighing the Pearson correlation coefficient with cosine similarity. Secondly, we explore the local label-specific relative discrimination by leveraging from both the instance-level and class-level correlation distribution within k nearest neighborhood. Finally, we conduct an information fusion strategy to comprehensively integrate the positive and the negative tendencies at the neighborhood level. Instances with salient positive tendency and compact neighborhood structure receive larger values while receiving smaller values with salient negative tendency and sparse neighborhood structure. With the concatenation of original features and augmented features, we examine the classification performance of the proposed granule correlation-based feature augmentation (GOFA) on well-established second-order multi-label classification methods. Extensive comparisons on thirteen benchmarks demonstrate the statistical superiority of GOFA over state-of-the-art multi-label classifications.

High performance RGB-Thermal Video Object Detection via hybrid fusion with progressive interaction and temporal-modal difference

RGB-Thermal Video Object Detection (RGBT VOD) is to localize and classify the predefined objects in visible and thermal spectrum videos. The key issue in RGBT VOD lies in integrating multi-modal information effectively to improve detection performance. Current multi-modal fusion methods predominantly employ middle fusion strategies, but the inherent modal difference directly influences the effect of multi-modal fusion. Although the early fusion strategy reduces the modality gap in the middle stage of the network, achieving in-depth feature interaction between different modalities remains challenging. In this work, we propose a novel hybrid fusion network called PTMNet, which effectively combines the early fusion strategy with the progressive interaction and the middle fusion strategy with the temporal-modal difference, for high performance RGBT VOD. In particular, we take each modality as a master modality to achieve an early fusion with other modalities as auxiliary information by progressive interaction. Such a design not only alleviates the modality gap but facilitates middle fusion. The temporal-modal difference models temporal information through spatial offsets and utilizes feature erasure between modalities to motivate the network to focus on shared objects in both modalities. The hybrid fusion can achieve high detection accuracy only using three input frames, which makes our PTMNet achieve a high inference speed. Experimental results show that our approach achieves state-of-the-art performance on the VT-VOD50 dataset and also operates at over 70 FPS. The code will be freely released at https://github.com/tzz-ahu for academic purposes.

Learning Autonomous Navigation in Unmapped and Unknown Environments.

Autonomous decision-making is a hallmark of intelligent mobile robots and an essential element of autonomous navigation. The challenge is to enable mobile robots to complete autonomous navigation tasks in environments with mapless or low-precision maps, relying solely on low-precision sensors. To address this, we have proposed an innovative autonomous navigation algorithm called PEEMEF-DARC. This algorithm consists of three parts: Double Actors Regularized Critics (DARC), a priority-based excellence experience data collection mechanism, and a multi-source experience fusion strategy mechanism. The algorithm is capable of performing autonomous navigation tasks in unmapped and unknown environments without maps or prior knowledge. This algorithm enables autonomous navigation in unmapped and unknown environments without the need for maps or prior knowledge. Our enhanced algorithm improves the agent's exploration capabilities and utilizes regularization to mitigate the overestimation of state-action values. Additionally, the priority-based excellence experience data collection module and the multi-source experience fusion strategy module significantly reduce training time. Experimental results demonstrate that the proposed method excels in navigating the unmapped and unknown, achieving effective navigation without relying on maps or precise localization.

Research on model transfer strategies based on the fusion of NIR-MIR spectral data

This study investigated the impact of different data fusion strategies on the performance of soluble solids content (SSC) prediction models based on near-infrared and mid-infrared spectroscopic techniques. In the data-level fusion approach, we applied standard normal variate and multiplicative scatter correction for pre-processing the NIR and MIR data. For the feature-level fusion, we utilized successive projections algorithm and competitive adaptive reweighted sampling to select informative wavelengths, and then applied direct orthogonal projection (DOP) for model transfer. The study employed a dataset of 150 honey samples to evaluate the impact of different data fusion strategies on model performance. To effectively evaluate model performance, we utilized the coefficient of R2 and RMSEP as evaluation metrics. By comparing the results of data-level fusion, feature-level fusion and single-spectrum model transfer, the results showed that spectral data fusion improved the model transfer performance compared to the single-spectrum approach, with feature-level fusion exhibiting the most significant advantages. The effective variable selection techniques in feature-level fusion successfully removed a substantial amount of interfering data and significantly reduced noise influence, thereby improving the model accuracy. Specifically, the use of feature-level fusion improved the predictive model’s R2 from 0.319 to 0.878 and reduced the RMSEP from 1.974 to 0.613°Brix, demonstrating the significant advantages of this approach in enhancing model transfer performance. The research findings provide important reference and theoretical support for future studies in the field of food quality assessment and other near-infrared spectroscopic data applications. This not only validates the effectiveness of the feature-level fusion approach, but also lays the foundation for establishing efficient and reliable predictive models.