Search Results (3,144)

Innovation-oriented productive forces are increasingly concentrated in cities, but the multiscale mechanisms through which the built environment shapes these forces remain insufficiently understood. This study develops a spatial analytical framework linking firm-level new quality productive forces (NQPF) to fine-grained urban spatial structures. Using 89 A-share listed firms in the Xiamen–Zhangzhou–Quanzhou (XZQ) urban agglomeration, we first construct an entropy-weighted NQPF index from eleven financial indicators related to R&D human capital, advanced capital stock, intangible assets, and operational efficiency. Kernel density estimation is then used to transform discrete firm-level NQPF values into a continuous 600 m × 600 m grid surface as the dependent variable. On the explanatory side, 27 built-environment variables are organized into an integrated indicator system covering urban form, natural conditions, jobs–housing structure, and service infrastructures. We combine cross-validated recursive feature elimination (RFE-CV) with multiscale geographically weighted regression (MGWR) to construct two model specifications: a 7-variable parsimonious subset and a 14-variable highest-performing subset. This dual-subset design allows us to distinguish core structural drivers from more context-dependent spatial mechanisms. The results reveal three mechanisms. First, ecological adaptation reflects the scale-dependent enabling and constraining effects of infrastructure and natural-foundation variables. Second, structural coordination shows that mature cores may experience crowding-related suppression when functional and institutional resources become spatially mismatched. Third, boundary activation indicates that transport, public-service, and leisure-related facilities can activate peripheral and cross-jurisdictional interface zones when supported by network connectivity and institutional coordination. By coupling variable-specific bandwidths with local coefficients, this study advances the analysis of spatial heterogeneity and provides evidence for differentiated, innovation-oriented urban regeneration. Full article

Accurate detection and characterization of surface defects in concrete dams is vital for ensuring safe operation. To address the limitations of existing research focused solely on cracks and the challenges traditional convolutional networks face in adapting to deformation and multiscale features, this study introduces DCN-YOLO, a deformable convolution-augmented framework for the simultaneous detection and classification of multiple defect types from UAV-acquired imagery. The model outputs bounding box localizations and categorical labels. Based on YOLOv12, this proposed model integrates DCNv4 deformable convolutions with the C3k2 module. By leveraging adaptive sampling offsets and dynamic modulation, the proposed model enhances geometric modeling for irregular defects, improving the detection of small and medium defects while achieving an acceptable trade-off in inference efficiency. To address multiple defect coexistence, we adopt Binary Cross-Entropy (BCE) loss to decouple classification and localization, improving training stability in multi-label scenarios. A Multi-defects dataset was created using UAV images, and performance was validated on the CrackSeg public dataset. The proposed model achieved 77.4% ± 0.2% overall precision under complex conditions, exceeding the YOLOv12l baseline by 7.1% and improving mAP_50-95 by 4.2%. It demonstrated competitive performance in detecting cracks, aggregate exposure, and construction joints, thereby providing a potentially robust and efficient approach for intelligent inspection of concrete dam surface defects. Full article

Precision livestock farming requires monitoring approaches that extend beyond static activity thresholds to enable dynamic, animal-level decision support. A time-aware machine learning framework was developed to transform RFID-derived goat activity records into an interpretable behavioral monitoring system. Time-stamped activity data were processed into temporal features such as lagged activity, rolling mean, activity change, and elapsed-time variables to capture short-term behavioral history. The framework integrated latent-state discovery, fuzzy-uncertainty analysis, transition modeling, supervised classification, short-horizon forecasting, and dashboard-based alert visualization within a predictive dashboard-based monitoring framework. Four latent behavioral clusters were identified, with the dominant cluster representing a stable low-activity baseline and accounting for 77.88% of observations. Boundary-zone analysis indicated that 7.30% of observations were in transitional regions, while fuzzy clustering classified 21.42% as uncertain or mixed-state points, suggesting gradual shifts in activity. Transition analysis revealed greater persistence in baseline states and lower persistence in high-activity/non-baseline states, which exhibited the highest volatility and entropy. Using nested time-blocked validation, Random Forest predicted future high-activity/non-baseline onset with AUC values of 0.869 and 0.841 for 5 and 10 min horizons, respectively. These results demonstrate that activity instability can be detected and forecasted over short horizons, supporting behavior-based monitoring. However, external biological validation is still required before implementation as a health- or disease-detection system. Full article

Long-tailed data are very common in industrial scenarios because equipment failures occur with a low probability, resulting in far fewer faulty samples than normal ones. However, when facing long-tailed data distributions, existing deep learning methods suffer from a significant degradation in performance and exhibit high bias. To overcome this limitation, this paper proposes a network that combines balanced adaptive logit-compensated cross-entropy loss with quadratic convolution (BALQNet) to improve diagnostic performance under long-tailed data conditions. The proposed method mainly consists of a balanced adaptive logit-compensated cross-entropy loss (BAL) and a quadratic convolution backbone. By jointly incorporating logit compensation, label smoothing, and class reweighting, BAL enhances the optimization of minority-class samples, thereby improving the classifier’s ability to distinguish different categories without introducing additional architectural complexity. Meanwhile, quadratic convolution further improves the effectiveness of feature representation learning. Finally, experiments are conducted on self-built bearing, gear, and motor datasets. The results show that BALQNet maintains strong diagnostic performance when handling long-tailed data. In addition, the ablation results provide further evidence for the effectiveness of the proposed approach. Full article

Internet of Things (IoT) devices are increasingly targeted by rapidly evolving malware, yet collaborative detection remains challenged by privacy leakage, noisy and imbalanced training data, and weak integrity guarantees when sharing model updates. This paper presents Mal-Fedchain, a secure and privacy-preserving framework for image-based IoT malware detection and prevention that couples federated learning with blockchain and honeypot-assisted behavioral monitoring, targeting Linux-capable IoT gateway devices. Portable Executable (PE) binaries are transformed into grayscale images using a corrected fixed-width byte-mapping pipeline stabilized by an information-maximizing GAN (IMGAN). A bi-level preprocessing pipeline applies two-sided weighted sparse representation (T-WSR) denoising—designed to selectively suppress zero-padding artifacts, high-entropy packed regions, and sparse opcode noise while preserving discriminative section-boundary texture—followed by geometric augmentation to mitigate class imbalance. Malware detection and family attribution are performed using a residual capsule-based network (RBCN) that fuses discriminative visual representations with PE-header features via concatenation, improving robustness against polymorphism and obfuscation. A formal threat model governs three adversary classes: a semi-honest aggregation server, a bounded fraction of malicious clients (up to 30%), and a passive eavesdropper. To enable collaboration without exposing raw data, clients train locally and share only MemCbar-encrypted updates; a permissioned Hyperledger Fabric blockchain ledger records hashed updates and security events to provide integrity, traceability, and tamper resistance. A file-system-integrated honeypot captures evasive behaviors and logs auditable evidence to strengthen prevention. Experiments on the Malimg dataset across five ablation configurations demonstrate that the corrected RBCN pipeline achieves $93.52\%$ accuracy, $92.40\%$ precision, $93.52\%$ recall, $92.52\%$ F-measure, MCC of $0.9245$, and AUC of $0.9976$ in its centralized configuration, and $65.62\%$ accuracy with AUC of $0.9840$ in the full federated configuration with five clients and eight communication rounds, substantially outperforming all baselines across all reported metrics. Full article

Accurate footprint image segmentation is challenging in forensic applications because fine anatomical structures, weak boundaries, and background interference can degrade segmentation performance. This study presents a task-oriented generative adversarial network (GAN)-based framework for forensic footprint image segmentation. Channel Prior Convolutional Attention (CPCA) modules are integrated into the decoder stages of the generator to recalibrate fused encoder–decoder features and preserve fine details in the toe, arch, and heel regions. In addition, a dual-branch discriminator processes image–mask pairs at the original and downsampled scales, providing complementary constraints on local boundary details and global footprint morphology. The framework is trained with a least-squares adversarial loss and a binary cross-entropy (BCE)–Dice segmentation loss. Experiments on the self-collected aFoot_2025 dataset show that the proposed framework achieves an IoU of 0.9448 and a Dice coefficient of 0.9713, outperforming the evaluated baseline and attention-based alternatives. Under the evaluated synthetic Gaussian-noise settings, the proposed method retained relatively stable segmentation performance. Furthermore, an exploratory footprint-based height-prediction analysis showed modestly lower prediction errors than the baseline GAN. These findings indicate that, under the controlled acquisition conditions of the aFoot_2025 dataset, CPCA-based feature calibration and dual-scale discrimination may improve segmentation-mask quality and provide a possible benefit for subsequent anthropometric analysis. Full article

This paper proposes a Cross-Representation Attention-Based Neural Network with fuzzy image enhancement for joint classification and segmentation of chest X-ray and kidney images. First, each input image is transformed into three complementary representations using histogram spread, fuzzy entropy, and fuzzy standard deviation-based enhancement. These representations emphasize different intensity distributions, informative regions, and local structural variations. A Cross-Representation Attention Fusion module then models multidirectional relationships among the enhanced representations and adaptively integrates their complementary features into a unified feature space. The fused features are processed by a shared encoder with task-specific classification and segmentation heads. The framework is evaluated for clinically relevant chest X-ray abnormalities, including pneumonia, pneumothorax, pleural effusion, and lung opacity, and for kidney-image classes comprising normal, tumor/renal cell carcinoma, and cystic renal mass cases. Experimental results show that the proposed method outperforms conventional and recent baseline models in both classification and segmentation. Ablation studies confirm that the fuzzy enhancement branches, cross-representation attention, and joint multi-task learning each contribute to the overall performance. Statistical and qualitative analyses further demonstrate the stability of the results and the model’s ability to localize relevant lesion regions. The proposed framework provides an effective and interpretable approach to unified medical image classification and segmentation while maintaining a reasonable balance between predictive performance and computational cost. Full article

Background: Metastatic liver tumors (MLT) and parasitic liver cysts (PLC) are common liver conditions that often exhibit similar imaging characteristics, making accurate diagnosis challenging using imaging alone. This overlap can result in diagnostic errors and delayed treatment, particularly in resource-limited settings or when invasive procedures such as biopsies are not feasible due to risk or unavailability. This study aimed to develop a reliable and transparent machine learning approach to distinguish MLT from PLC using radiomic features derived from computed tomography (CT). Methods: We propose an explainable radiomics-based machine learning framework for the non-invasive, accurate, and interpretable discrimination of MLT and PLC, designed to assist radiologists in reducing diagnostic ambiguity and expediting patient management. This retrospective study included 30 adult patients, comprising 15 with liver metastases and 15 with pathologic hepatic cysts. Radiomic features were extracted from pre-treatment CT scans using PyRadiomics. Feature selection was performed using three complementary methods: Mutual Information, Lasso regression, and LightGBM importance ranking. HistGradientBoosting classifiers were then trained on each selected feature set. Results: Model performance was evaluated using 5-fold cross-validation and assessed with ROC AUC, accuracy, precision, recall, and F1-score. SHAP analysis was applied to interpret the models and identify key radiomic biomarkers. Statistical comparisons were performed using DeLong’s test for AUCs, McNemar’s test for classification agreement, and paired t-tests for metrics such as accuracy and F1-score. The Mutual Information-based model achieved the highest mean AUC (0.9717 ± 0.0267), significantly outperforming the other models (p < 0.035). Key features contributing to classification included texture entropy, interquartile range, and gray level non-uniformity. Conclusion: We developed a robust and interpretable machine learning framework for differentiating metastatic liver tumors from parasitic liver cysts using CT-derived radiomic features. The integration of Mutual Information feature selection, ensemble learning, and SHAP explainability ensured high diagnostic accuracy, strong calibration, and transparency. The proposed framework demonstrates substantial clinical relevance and holds promise for real-world implementation. Full article

In the industrial strip steel production process, complex background interference and varying defect scales make existing semantic segmentation models prone to scale confusion and noise propagation issues during the feature representation stage, thereby limiting segmentation accuracy and stability. To address these problems, this paper proposes a steel surface defect segmentation network based on dynamic receptive field modulation and semantic-guided adaptive modulated fusion, termed DAMF-Net. First, in the encoding stage, a kernel selection fusion attention module (KSFA) is designed, which constructs multi-scale receptive field responses through cascaded depthwise convolution, and combines spatial average prior with channel-level competitive weights to perform adaptive gating modulation on bottleneck features, thereby enhancing the model’s ability to represent different scale defect structures; second, in the decoding stage, a task-oriented Adaptive Modulated Fusion Module (AMFM), inspired by modulation-based feature fusion, is introduced to adaptively fuse shallow detail features and deep semantic features through branch-wise competitive weighting for each channel, thereby reducing the propagation of shallow background noise during decoding; additionally, a hybrid optimization objective combining binary cross-entropy and Dice loss is constructed to enhance the model’s learning ability for small-scale and low-contrast defects. Experimental results on the Severstal and ESDIs-SOD datasets show that the proposed method improves the mDice/Dice metrics by 2.38% and 4.40% respectively compared to the U-Net baseline, and exhibits lower prediction errors and more stable segmentation performance under complex background conditions. Meanwhile, DAMF-Net achieves an inference speed of 55.8 FPS while ensuring improved accuracy, demonstrating a good balance between precision and efficiency. This method provides an effective solution for high-precision segmentation of steel surface defects in complex industrial scenarios. Full article

One of the major challenges in securing medical image communication systems is the secure and efficient management of cryptographic key material. In this paper, we propose a multi-layer image encryption algorithm that addresses image security while reducing per-image key-storage and transmission overhead under a pre-shared protected-generator model. The proposed algorithm integrates a Generative Adversarial Network, a Piecewise Linear Chaotic Map, DNA complement operations, and bit-level zigzag permutation. A distinguishing feature of the proposed algorithm is that the key image is generated from an image-specific 100-dimensional noise vector, which serves exclusively as the input to the trained generator, while the chaotic parameters and diffusion materials are derived from the generated key image. In this approach, under the assumption of a pre-shared protected generator, transmitting only the image-specific 100-dimensional noise vector that bears no structural relationship to the key image reduces per-image key storage and transmission overhead. Comprehensive numerical evaluations were performed on eleven images, comprising both standard test images and medical images, to assess the security and robustness of the proposed algorithm. The experimental results demonstrate entropy values exceeding 7.996 bits, along with NPCR and UACI values of 99.60% and 33.46%, respectively. Adjacent pixel correlations are reduced to near-zero levels across all tested images. The proposed algorithm exhibits strong robustness against common attacks, including up to 75% cropping and 50% salt-and-pepper noise. The proposed algorithm achieves competitive performance compared with several existing encryption methods. Successful decryption requires the correct image-specific noise vector and the original trained generator. Full article

Reliable assessment of rockburst tendency is essential for maintaining the stability and operational safety of deep underground excavations. However, the complex coupling among stress conditions, lithological characteristics, and structural features of rock masses introduces significant uncertainty into rockburst prediction. Conventional evaluation approaches relying on individual indices frequently produce inconsistent classifications and are often insufficient to represent actual rockburst behavior. To address this issue, a hybrid evaluation framework integrating unascertained measure theory, cloud-based uncertainty analysis, and a game-theoretic weighting strategy was developed in this study. Four representative parameters, including the strain energy storage index (W_et), geostress index (S), rock quality designation (RQD), and rock mass integrity factor (K_v), were adopted to characterize the energy-storage capability, stress environment, and structural condition of the surrounding rock mass. The conventional unascertained measure approach was further enhanced using the normal cloud model to describe the uncertain mapping relationship between quantitative measurements and qualitative rockburst classifications. In addition, a combination weighting scheme incorporating AHP, entropy weight (EW), and CRITIC methods was established to improve the stability and rationality of index weighting. The developed framework was subsequently applied to a deep phosphate mine in China. The calculated comprehensive weights of the four evaluation parameters were 0.1982, 0.3446, 0.2173, and 0.2399, respectively, demonstrating that the stress-related parameter has the greatest influence on rockburst evaluation. The results indicate that the investigated rock masses generally exhibit moderate-to-strong rockburst tendency. The shallow and moderately deep zones exhibited relatively high rockburst potential, while the ultra-deep dolomite formations mainly showed a moderate tendency due to the development of joints and fractures, which weakened the integrity of the deep rock mass. The proposed framework provides an effective and practical approach for preliminary hazard assessment, rockburst risk zoning, and prevention strategy design in deep mining engineering. Full article

Liquefied natural gas (LNG) submersible centrifugal pumps are critical components in marine energy transportation systems, and fault-induced degradation may threaten operational safety, transfer reliability, and maintenance efficiency. However, condition monitoring and fault diagnosis often rely on heterogeneous multi-source data, where redundant information, unequal channel sensitivity, and inter-signal coupling may obscure discriminative fault features. To address this challenge, this paper proposes a hierarchical multi-source condition monitoring and fault diagnosis framework for LNG submersible centrifugal pumps by integrating an Entropy-Weighted Sensor Selection Method (EWSSM) with a hybrid convolutional neural network (CNN)–Transformer model. Functional information is used for front-end abnormality screening, while selected response signals are used for fault category recognition. EWSSM evaluates channel contribution and suppresses redundant inputs to construct a compact fault-sensitive input space. The CNN–Transformer model combines local feature extraction with global dependency modeling to identify complex fault patterns. A laboratory-scale fault simulation platform was established, and vibration, acoustic, internal pressure-pulsation-related response information, and operating parameter data were collected under ten operating states. Experimental results show that the proposed framework achieves effective abnormality screening and accurate fault diagnosis, with an average classification accuracy of 98.73% over repeated experiments. Covariance-difference analysis further provides interpretable evidence for condition assessment by revealing fault-related multi-source response redistribution. The proposed framework provides an effective, intelligent monitoring and diagnosis solution for LNG submersible centrifugal pumps and supports reliability-oriented operation and maintenance of marine energy transportation equipment. Full article

Traditional manual grading of fresh chili peppers suffers from inconsistent quality control and low efficiency. To meet the demand for accurate fruit shape recognition during the post-harvest stage, this study proposes an intelligent recognition method based on an improved DenseNet-121 network. This approach facilitates the application of machine vision in agricultural sorting equipment. DenseNet-121 serves as the backbone network. The Convolutional Block Attention Module (CBAM) is introduced to enhance feature focus on fruit shapes. A regularization strategy (Dropout = 0.3, weight decay = 1 × 10⁻⁴) and a cross-entropy loss function with label smoothing (LS = 0.1) are integrated to optimize decision boundaries. These configurations prevent the model from overfitting to hard training labels and yield a robust classification architecture. Experimental results demonstrate that the proposed model achieves a precision of 90.09%, a recall of 89.60%, an F1-score (the harmonic mean of precision and recall) of 89.53%, and an overall accuracy of 89.74%. The model contains 7.09 M parameters and requires a single-frame inference time of 7.35 ms. Comprehensive evaluations indicate that the proposed model achieves an optimal balance among environmental noise robustness, prediction accuracy, and computational efficiency. Consequently, by maintaining high fine-grained classification accuracy alongside a low memory footprint and rapid inference speed, the model demonstrates strong potential for real-time deployment on resource-constrained edge devices within actual agricultural optical sorting equipment. Full article

Brain tumor classification from magnetic resonance imaging (MRI) remains challenging in settings where only image-level labels are available and tumor classes exhibit overlapping visual characteristics. In this study, we consider the publicly available Brain Tumor MRI Dataset from Kaggle, a four-class dataset composed of 2D MRI slices belonging to the categories glioma, meningioma, pituitary tumor, and no tumor. Accordingly, the proposed framework is formulated as a slice-based multiclass classification approach rather than a volumetric 3D analysis pipeline. We propose a lightweight and interpretable framework that integrates handcrafted multiscale MRI descriptors, an artificial neural network (ANN), Bee Colony Optimization (BCO)-based neural architecture search, and fuzzy softmax confidence modeling. Each MRI slice is represented by a compact 9-dimensional feature vector derived from intensity, local entropy, and gradient magnitude computed globally and over non-overlapping spatial blocks. The ANN design problem is formulated as a discrete–continuous optimization task, where BCO is employed to optimize network architecture and training hyperparameters by maximizing validation macro-$F_1$. To quantify predictive reliability, the softmax outputs are interpreted as fuzzy class memberships and further analyzed using maximum membership, normalized entropy, decision margin, and ambiguity measures, enabling confidence-aware reliability assessment. These fuzzy confidence descriptors enable confidence-threshold-based selective classification and rejection of low-confidence predictions. Across repeated runs, the optimized BCO-ANN achieved a mean test accuracy of $0.781\pm 0.009$, mean macro-$F_1$ of $0.775\pm 0.010$, mean Brier score of $0.319\pm 0.012$, and mean Expected Calibration Error (ECE) of $0.0273\pm 0.0080$, compared with $0.748\pm 0.011$, $0.738\pm 0.013$, $0.352\pm 0.010$, and $0.0446\pm 0.0071$ for the baseline ANN, respectively. Under confidence-threshold-based rejection, selective macro-$F_1$ increased to $0.820\pm 0.009$ at $\tau =0.55$ and to $0.874\pm 0.020$ at $\tau =0.85$, with the expected reduction in coverage. These results indicate that the proposed framework provides a transparent and reproducible approach for optimization-aware and confidence-aware multiclass brain tumor MRI classification in a lightweight handcrafted feature setting. Full article

While plug-in hybrid electric vehicles (PHEVs) can significantly reduce fuel consumption and emissions, their real-world benefits strongly depend on heterogeneous driver usage patterns. Understanding these usage patterns is therefore essential for optimizing energy management and electrification policies. This study presents a data-driven framework for identifying and characterizing PHEV driving behavior using two primary indicators: the Utility Factor (UF), which quantifies the proportion of electric-mode driving, and the annual vehicle kilometers traveled (VKT), which measures driving intensity. A Gaussian Mixture Model (GMM) is employed in a transformed feature space characterized by the logit of UF and the logarithm of VKT to capture nonlinear relationships and diverse usage patterns. The Bayesian Information Criterion (BIC) is used to find the optimal number of behavioral clusters. To assess the robustness of the clustering, we perform a bootstrap stability analysis and compare it with k-means and a density-based clustering method (DBSCAN). Based on an analysis of real-world PHEVs, three distinct usage patterns are identified. The dominant cluster (96.6%) exhibits moderate electric usage (UF = 0.38) with an annual mileage of 18,617 km and fuel consumption of 3.89 L/100 km, whilst two smaller clusters represent near-full electric operation (2.3%, UF about 1.0, negligible fuel) and low-mileage users (1.1%, approximately 2444 km/year). The clustering demonstrates high assignment confidence (posterior entropy $<{10}^{-4}$) and moderate stability, with a mean Adjusted Rand Index (ARI) of 0.448. The findings indicate that the proposed probabilistic clustering framework offers a comprehensible and statistically robust method for identifying diverse PHEV usage patterns. These insights can support adaptive energy management strategies, effective planning of charging infrastructure, and evidence-based policies to maximize the real-world electrification benefits of PHEVs. Full article