Search Results (113)

Reliable diagnosis of incipient transformer faults is essential for preventing catastrophic failures and enabling predictive asset management in power systems. Although dissolved gas analysis (DGA) is the most established diagnostic tool for assessing transformer internal condition, fault discrimination remains difficult when gas features are highly correlated, redundant, and only partially separable across fault classes. This study presents a PCA-enhanced artificial neural network (ANN) framework for multiclass transformer fault diagnosis using DGA data. The method is developed on 595 samples classified into six IEC 60599 fault categories and uses a 15-feature representation comprising raw gas concentrations, total hydrocarbon content, and engineered gas-ratio descriptors. To identify an evidence-based diagnostic representation, principal component analysis (PCA) was evaluated across all dimensionalities from k = 1 to 15 before ANN training. The proposed model was benchmarked against alternative feature sets and conventional classifiers, including Gaussian Naïve Bayes, k-nearest neighbours, support vector machines, and ANN without PCA. The best-performing configuration was obtained at k = 13, yielding a test accuracy of 68.1%, compared with 63.9% for ANN without PCA, 56.3% for raw-gas-only ANN, and 33.6% for the IEC three-ratio feature configuration. In addition to improving diagnostic performance, the PCA stage revealed interpretable component structures associated with dominant gas and ratio patterns underlying fault separation. The results indicate that PCA-based feature extraction improves ANN generalization by reducing redundancy and multicollinearity in DGA-derived variables, and provides a practical, lightweight, and interpretable framework for transformer fault diagnosis. Full article

Water quality monitoring is a fundamental element of sustainable aquaculture management, as changes in parameters of physicochemical and biological properties directly affect the health, growth performance, and productivity of the aquaculture systems. Although traditional machine learning (ML) methods have demonstrated effectiveness in water quality classification, their performance often depends on large amounts of labeled data, which can be challenging and expensive to collect in real-world aquaculture environments. This study explores a few-shot learning (FSL) framework for data-efficient water quality classification under limited supervision to address this limitation. Several FSL models, including prototypical networks (ProtoNet), Siamese Networks, and Matching Networks were developed and evaluated in a comparative experimental framework against the traditional machine learning classifiers logistic regression, random forest, support vector machine and extreme gradient boosting. Low-data learning scenarios were simulated using a structured episodic evaluation approach. Experimental results demonstrate FSL techniques outperform traditional machine learning methods across all evaluated scenarios. Among the tested methods, ProtoNet achieved the highest performance, attaining an accuracy of 94.46% and an ROC-AUC score of 98.65%, indicating superior discriminative capability and robustness. Siamese Networks also demonstrated competitive performance under highly constrained data conditions. Furthermore, latent-space visualization, confusion matrix analysis, paired t-test statistical analysis, and ablation studies confirmed that episodic meta-learning enables the learning of highly discriminative latent representations with strong generalization capability under limited labeled data conditions. The findings highlight that FSL provides a robust and scalable framework for intelligent water quality classification in aquaculture systems, particularly in scenarios where labeled data are scarce, offering significant potential for sustainable aquaculture monitoring applications. Full article

Tea bud detection is a fundamental prerequisite for accurate tea yield estimation and intelligent mechanical harvesting. However, existing detection methods face several critical challenges, including ineffective extraction of multi-scale features, weak feature saliency for small tea bud targets, and the prevalent imaging issue in which the central regions of tea images are in focus while peripheral areas suffer from defocus blur. These factors collectively result in a high rate of missed detections, severely limiting detection accuracy and subsequent application performance. To overcome these technical bottlenecks, this paper proposes a novel tea bud detection framework, termed REB-Tea, which integrates image clarity optimization with multi-scale feature enhancement. First, the Restormer image restoration network is employed to improve overall image clarity and enhance the discriminative representation of tea bud features. Subsequently, a bidirectional feature pyramid network (BiFPN) structure and an efficient multi-scale attention (EMA) mechanism are incorporated into the neck of the YOLOv5 model to strengthen multi-scale feature fusion and guide the network to focus on fine-grained tea bud features across different scales, thereby improving detection performance for small and densely distributed targets. Experimental results based on 10-fold cross-validation demonstrate that the proposed REB-Tea model achieves an average mAP₅₀ of 95.5% on the Longjing 43 tea test set, representing a 9.9 percentage point improvement over the baseline YOLOv5 model, and Welch’s independent two-sample t-test verifies that this accuracy increment is highly statistically significant. Moreover, the model exhibits reliable detection performance across different tea varieties, including Cuifeng and Fuding White Tea. Specifically, the mAP₅₀ reaches 88.3% on Cuifeng, which shares similar appearance characteristics with Longjing, and 78.1% on Fuding White Tea, which has noticeably different appearance characteristics from Longjing. These results confirm the effectiveness of the REB-Tea framework in addressing challenges such as out-of-focus blurring, weak feature saliency, and multi-scale feature extraction. Overall, the proposed approach significantly enhances tea bud detection accuracy in natural environments and provides robust technical support for intelligent tea harvesting applications. Full article

Enterprise financial risk prediction is a fundamental task in financial risk management, yet its performance is often hindered by severe class imbalance, cross-enterprise heterogeneity, and the limited availability of labeled risky samples. These challenges are particularly pronounced in few-shot settings, where conventional machine learning and deep classification models tend to suffer from unstable representation learning, feature collapse, and weak decision boundaries. To address this issue, this study proposes a hierarchical metric learning framework for few-shot enterprise financial risk prediction on imbalanced tabular data. The framework integrates a state-space feature embedding network, an Adaptive Spectral Decomposition and Multi-Scale State Embedding module, and a Hierarchical Metric Manifold Alignment mechanism to enhance risk-sensitive representation learning, preserve geometric consistency across embedding levels, and improve prototype-based discrimination in the metric space. Experiments are conducted on three public datasets, namely American Bankruptcy, Corporate Financial Risk Assessment, and Enterprise Financial Network, under a unified 2-way 20-shot setting. The proposed method consistently achieves the best overall performance across Precision, Recall, Accuracy, F1-score, and AUC, with AUC values of 0.9526, 0.9687, and 0.9716 on the three datasets, respectively. Ablation studies and visual analyses further show that the proposed framework improves intra-class compactness, inter-class separability, and classification robustness under highly imbalanced conditions. These findings indicate that the proposed method provides an effective and robust machine learning solution for enterprise financial risk prediction and early warning in data-scarce financial scenarios. Full article

Landslides constitute a globally pervasive and highly destructive natural hazard. Although artificial intelligence (AI)-driven landslide susceptibility mapping has emerged as an effective tool for delineating high-risk zones, its predictive performance is frequently constrained by inherent data noise and insufficient characterization of landslide triggering factors, restricting the credibility of the mapping results. In this study, to remedy this limitation, we adopt fractal analysis to extract latent inherent information from topographic features. Specifically, the box-counting method and multifractal analysis are applied to excavate the intrinsic nonlinear characteristics embedded in eight topographic factors, and an improved K-means algorithm is utilized to perform feature selection and construct a dedicated fractal feature dataset, which is fed to advanced AI models. Our results indicate that the information dimension ($D_1$) of the slope gradient, the correlation dimension ($D_2$) of aspect, land relief, the $D_2$ of roughness, the $D_2$ of plan curvature, the multifractal spectrum width ($∆\mathsf{\alpha }$) of profile curvature, the $D_2$ of elevation, and the surface cutting depth were the most effective features, demonstrating superior performance in capturing landslide targets. Comparative performance evaluations reveal that AI models trained on fractal features demonstrate substantially superior predictive capabilities compared to AI models trained on raw features. This superiority is consistently evidenced across key evaluation metrics, including overall accuracy, kappa coefficient, F1-score, and predictive efficiency, demonstrating that the integration of fractal characteristics significantly augments model robustness and predictive efficacy. To mitigate the ‘black-box’ problem of AI modeling, Shapley additive explanations were employed to quantify individual feature contributions and elucidate the underlying predictive mechanisms. Our findings indicate that the integration of fractal analysis yields highly discriminative and robust feature representations, thereby expanding the representational capacity of the models and improving predictive accuracy. Furthermore, a joint assessment of spatial uncertainty and susceptibility maps demonstrates that these models exhibit low predictive variance and high spatial stability when delineating high-susceptibility zones. Notably, models utilizing fractal-derived features achieve superior spatial capture efficiency. The resultant topographic features characterized by fractal representation and selected via the improved K-means algorithm can significantly improve the predictive performance of trained AI models in landslide susceptibility mapping tasks, offering a scientific and viable technical approach for future landslide prediction and prevention. Full article

Hyperspectral images (HSIs) and Light Detection and Ranging (LiDAR) data offer complementary spectral and spatial information and are extensively applied to land cover classification. Nevertheless, current fusion–classification approaches frequently suffer from cross-modal feature entanglement and insufficient exploitation of LiDAR physical priors, particularly the Digital Surface Model (DSM), which limits the interpretability of learned features and restricts classification accuracy. To address these issues, this study presents a Physics-Guided Adaptive Decoupling and Collaborative Enhancement Network (ADCE-Net) that embeds explicit geometric guidance into multimodal feature learning. In ADCE-Net, the DSM serves as an explicit geometric conditioning signal to guide feature decoupling, decomposing input representations into modality-shared semantic features (SSF) and modality-specific discriminative features (MSF), thereby mitigating cross-modal interference at an early stage. Based on this decomposition, an adaptive collaborative enhancement mechanism is designed using bidirectional cross-attention and dynamic gating to achieve context-aware mutual refinement between SSF and MSF, facilitating more effective utilization of cross-modal complementary information. Furthermore, a multi-level collaborative classification architecture is constructed to integrate multi-scale contextual representations, enhancing spatial consistency and boundary delineation. Extensive experiments on three benchmark datasets—Trento, Houston 2013, and Muufl Gulfport—demonstrate that ADCE-Net achieves overall accuracies of 99.69%, 97.37%, and 94.90%, respectively, surpassing multiple representative methods including support vector machines, 3D convolutional neural networks, transformer-based models, and recurrent neural networks. Noticeable improvements are also achieved for minority classes and classes with highly similar spectral signatures. The DSM-driven physics guidance boosts both classification performance and feature interpretability, providing a reliable and explainable paradigm for multimodal remote sensing classification. Full article

Synthetic Aperture Radar (SAR) has broad application prospects in target recognition; however, intrinsic multiplicative speckle noise, geometric distortions, and the complex coupling of multimodal features often limit the comprehensiveness of representations and the efficiency of fusion, thereby restricting recognition accuracy. To address these limitations, this paper proposes a SAR image target recognition method based on multidimensional feature fusion. The proposed method first achieves noise suppression and contrast enhancement through an optimized preprocessing layer. Subsequently, a dual-branch hierarchical feature extraction network synchronously captures low-dimensional physical prior features driven by domain knowledge and highly discriminative deep convolutional features, ensuring a balance between physical interpretability and high-capacity representation. Finally, a variance-adaptive weighted fusion layer dynamically balances the contribution of different feature streams, mitigating information redundancy and feature conflict. Quantitative experiments on the MSTAR and public CETC38-SAR datasets demonstrate that under various pre-trained backbones, the proposed framework improves precision, recall, and F1-score by 5%–15% compared with baseline methods. Ablation studies and evaluations under extended operating conditions further validate the robustness, computational efficiency, and structural validity of the decoupled architecture. Full article

The performance of face recognition is highly dependent on detection accuracy and robustness. However, most face recognition methods assume that face regions are perfectly aligned and rely exclusively on visual features within bounding boxes. This assumption is often violated in complex environments, including low resolution, poor illumination, and partial occlusion. To overcome these limitations, this paper proposes a collaborative face detection and recognition framework enhanced by meta-auxiliary learning. The proposed method exploits shared representations by an improved feature extraction module, and introduces gender classification and age estimation as auxiliary tasks to construct a meta-composite loss. Furthermore, facial geometric structure are acquired by extracting facial landmark coordinates from detected face regions. This multi-task architecture imposes implicit constraints on shared features, while simultaneously boosting feature discriminability and system robustness. By fusing facial geometric structures with semantic meta-information, facial identity representations are effectively enriched, leading to substantial performance gains in complex scenarios. Extensive experiments on the SCFace and LRD benchmark datasets validate the effectiveness of the proposed method. On SCFace, our method achieves 77.21% Genuine Acceptance Rate (GAR) at a False Acceptance Rate (FAR) of 1%, surpassing ArcFace, Octuplet Loss, and HDFGD by 1–2 percentage points. On LRD200 and LRD100 under an extreme low-resolution condition of 15 pixels, the recognition accuracies of the proposed method reach 67.98% and 60.98%, respectively—on par with state-of-the-art methods and markedly superior to conventional baselines. These results confirm that meta-auxiliary learning effectively compensates for insufficient visual information by leveraging semantic cues. Non-parametric statistical tests (Wilcoxon signed-rank test) confirm that the improvements achieved by the proposed method are statistically significant (p < 0.05). Full article

Fine-grained aircraft recognition in optical remote sensing imagery remains highly challenging under cloud occlusion, as visibility degradation causes structural information loss and weakens discriminative representation learning. To address this issue, we propose DGR-MAE, a teacher–student masked image modeling framework based on posterior semantic correction for robust representation learning under incomplete observations. Unlike existing semantic-guided masking methods that modify token visibility during input construction, DGR-MAE preserves high-ratio stochastic masking in the student branch and introduces semantic correction after visibility degradation through teacher-guided differential reconstruction. Specifically, a semantic-aware teacher branch estimates patch-level importance to partition masked regions into semantic-critical and non-critical subsets, enabling region-dependent reconstruction prioritization. A collaborative feature refinement mechanism is further incorporated to enhance contextual consistency and structural reasoning during pretraining. To support controlled evaluation, we construct the ASRAir benchmark with hierarchical cloud occlusion levels. Experimental results show that DGR-MAE achieves 74.28% Top-1 accuracy on ASRAir-Occ and achieves the best Top-1 performance while maintaining competitive Top-5 accuracy compared with representative self-supervised baselines. In particular, it demonstrates substantially improved robustness under moderate-to-severe cloud occlusion, validating the effectiveness of posterior semantic correction for remote sensing representation learning under visibility degradation. Full article

Robotic visual perception and terrain recognition are critical for autonomous locomotion and adaptive control in complex environments. However, existing models often extract weak features, confuse classes, and deliver unstable recognition. Most prior studies use end-to-end convolutional networks or single-stream feature extraction, which limits the balance between fine-grained visual representation and adaptive discrimination of confusing samples. To solve this problem, this paper proposes a vision model that blends attention mechanisms with a confusion augmentation strategy. Using an improved ResNet50 backbone, we add a local feature sharpening module and a channel–spatial attention module to strengthen edge texture and global context representation. We also design a confusion augmentation strategy based on the similarity of hard samples. It generates mixed samples through cross-perturbation in feature space, thereby improving the discrimination of highly similar terrains. Experiments show that our model achieves an accuracy of over 98.19% on various terrains, including cement, asphalt, sand, and snow. t-SNE visualization and Grad-CAM analysis demonstrate clear class separability and good interpretability, confirming the effectiveness and robustness of the approach for robotic terrain recognition. Full article

To address the problem of fault-related structures and noise disturbances in rolling bearing vibration signals being highly coupled in the original one-dimensional signal domain under severe noise conditions, in this study, we propose a Phase-space adaptive Expert Denoising Network (PaEDNet), a robust fault diagnosis framework that integrates representation construction, adaptive restoration, and condition discrimination. Unlike existing methods that mainly enhance network modelling directly in the original signal domain, the proposed framework first constructs a spatially organised two-dimensional similarity representation through phase-space reconstruction, which further unfolds fault-related dynamic structures from temporal entanglement and provides a more suitable preliminary representation domain for subsequent restoration. On this basis, a CoPaMoE-augmented adaptive denoising module is introduced into the representation domain to improve structural restoration capability under heterogeneous noise and different local patterns. DenseNet is then employed for fault classification, thereby forming an integrated fault diagnosis framework combining representation reconstruction, noise restoration, and condition discrimination. The resulting pipeline performs end-to-end diagnosis from raw vibration signals to fault labels at inference, while training is conducted in a stage-wise manner. Experimental results derived using the two public datasets, CWRU and PU, show that the proposed method consistently outperforms multiple comparative models under different signal-to-noise ratio conditions and maintains stronger robustness in low-SNR scenarios. Under the −6 dB condition, PaEDNet achieves classification accuracies of 93.98% and 90.45% on the two datasets, respectively. Further ablation studies and expert-routing analysis demonstrate that the combination of structured representation construction and adaptive expert restoration jointly enables the improved performance of the model. In this study, we provide a new modelling perspective for the fault diagnosis of vibration signals in complex noisy environments. Full article

Speech-based non-invasive screening offers a cost-effective and scalable approach for the early detection of Alzheimer’s disease (AD). However, the clinical utility of deep learning models remains severely constrained by the scarcity of labeled speech data in low-resource languages, necessitating cross-lingual transfer learning. Conventional domain adaptation paradigms typically assume semantically consistent feature domains and focus heavily on aligning marginal distributions; however, they suffer catastrophic performance degradation when applied to cross-lingual pathologic speech. By analyzing disease-associated representation vectors within a self-supervised HuBERT space, we uncover a systematic mechanism driving this failure, a phenomenon we term cross-lingual polarity flip, where the direction of disease-relative-to-control feature offsets fundamentally reverses between languages. While prior multilingual studies have largely discarded such dimensional inconsistencies as ungeneralizable noise, a 500-round Monte Carlo stability analysis demonstrates that these flips occur in a highly stable, structural manner across 18.3% of top discriminative dimensions. Leveraging this insight, we introduce Monte Carlo Polarity Flip Calibration (MC-PFC), a few-shot framework designed to explicitly rectify flip orientations. Requiring only five labeled support samples per class from the target domain, MC-PFC robustly estimates direction flips via a separability-weighted ensemble voting mechanism. Evaluated on a strictly held-out Chinese blind test set, MC-PFC achieves an area under the receiver operating characteristic curve (AUC) of 0.871, recovering 99.5% of the performance achieved by a full in-domain trained upper bound (AUC = 0.875). Ablation experiments confirm that direction calibration yields a substantial +0.361 AUC gain, vastly outperforming standard distribution alignment (+0.081). This work establishes a data-efficient paradigm for cross-lingual medical analysis, shifting the clinical AI focus from discarding cross-lingual discrepancies to actively modeling and calibrating them. Full article

Biologically related time-series data characterize the dynamic evolution of biological systems, including genetic inheritance, disease diagnosis, and the biological microenvironment. However, accurate prediction of these data remains challenging due to their pronounced time-varying, non-stationary, and noisy characteristics. Existing approaches often fail to capture latent shifts of biologically related time series, limiting both predictive performance and time-varying pattern recognition capability. Thus, in this study, we first propose a time-varying neural network (TVNN) model that combines frequency-domain information with Koopman theory. TVNN-model Koopman transition matrices are used to model global dynamics and local time-varying behaviors for pattern extraction. Secondly, a time-varying pattern recognition large language model (TVPRLLM) is introduced to recognize and classify the extracted time-varying patterns, enabling the identification of potential pattern categories. Thirdly, we have developed a biology-related time-series predictive platform that can offer visualization, data analysis, and predictive services. Experimental results demonstrate that the TVNN model outperforms existing mainstream methods in predicting biology-related time-varying time series, and that it achieves competitive forecasting performance, though its behavior depends strongly on the design of the frequency-domain decomposition. Additional robustness analyses reveal that the choice of Fourier masking strategy can materially affect both RMSE and long-horizon stability. We further show that Koopman-derived time-varying representations are highly discriminative for dynamic state recognition. Full article

Skin cancer image data often exhibit long-tailed distributions due to the inherent challenges in data collection and annotation. Specifically, a few predominant classes dominate a dataset of interest, while minority classes, referred to as tail classes, are underrepresented with only limited numbers of samples. Such imbalance is highly likely to adversely affect the performance of deep learning models. To address this issue, previous methods employ mixup techniques to synthesize tail-class images, thereby attempting to balance the training data. However, traditional mixup methods typically do not specifically pay attention to specific regions of interest, blending two images with indistinction between objects of interest and background. Such disregard for important semantic features may result in synthetic samples with broken or distorted diagnostic features. In this work, we introduce a novel framework, the Tail-synthetic Learning for Boosting Long-tailed Skin Cancer Image Classification (TailBoost) framework. Our approach generates a new tail-class image by combining a tail-class image with a head-class image under the guidance of their corresponding saliency maps. This strategy, namely SPMix, preserves and enhances the discriminative features of the tail-class image with minimum interference from the head-class image. We further refine the learned representations by incorporating supervised contrastive learning with class-center rebalance. Extensive experiments on the ISIC2018, ISIC2019, and PAD-UFES-20 datasets demonstrate that TailBoost outperforms existing state-of-the-art long-tailed learning methods. Full article

Point cloud scene flow estimation aims to capture the three-dimensional motion of each point in a sequence of point clouds. Although progress has occurred in this field, existing methods often face significant challenges. In particular, two key issues persist: the absence of corresponding local information from the source point cloud to the target, preventing correct feature matching, and the presence of highly similar adjacent structures in target regions, which leads to ambiguous correspondences due to indistinguishable point features. To address these problems, this paper introduces a novel self-supervised method for point cloud scene flow estimation. Theoretically, we establish a new framework that integrates discriminative feature learning with probabilistic flow refinement. A new network architecture, SetConv++, is designed to learn more discriminative point feature representations, enhancing differentiation in similar structures. Additionally, a refinement module uses the random walk algorithm to adjust initial flow estimates. This approach reconstructs low-confidence flows with high-confidence surrounding ones, reducing missing correspondence issues. Crucially, a new flow smoothing loss term ensures local consistency while suppressing error propagation—a fundamental limitation in existing methods. Through comprehensive experiments on the KITTI Scene Flow dataset, our method demonstrates superior performance. It significantly outperforms existing self-supervised approaches across multiple standard evaluation metrics. Specifically, on the KITTI Scene Flow dataset, our method reduces the Endpoint Error (EPE) by 13.6% (from 0.0411 to 0.0355) and improves Accuracy Strict (AS) by 2.43 percentage points (from 92.68% to 95.11%) compared to baseline self-supervised approaches, while also reducing the outlier rate (Out) by 1.5 percentage points. This advancement not only provides a robust theoretical framework for handling ambiguous correspondences but also enables more reliable and efficient downstream applications—such as autonomous driving perception systems requiring real-time motion accuracy in complex scenes. Full article