ABSTRACT. Prescription Recommendation is critical for improving treatment efficacy and enabling precision medicine. Current methods are limited by redundant features, insufficient fusion of hierarchical cross-modal herb information, inadequate modeling of high-order dependencies in heterogeneous networks, etc. We learn individualized patient representations, effectively fuses multi-modal herb information, and captures high-order dependencies in patient–syndrome–herb networks, enabling accurate herb recommendation and dosage prediction. We evaluated CMGRL in real-world clinical scenarios. A series of experiments was conducted, and the results demonstrate that our model consistently outperforms baselines across multiple evaluation metrics. In conclusion, CMGRL establishes an effective and promising paradigm for prescription recommendation.

ABSTRACT. To address the challenges of frequent ID switching and object occlusion in complex multi-object tracking (MOT) scenarios, this paper proposes a high-precision tracking framework based on YOLOv11 and an improved BoTSORT. First, a re-parameterizable feature extraction network, RepOSNet, is developed. By integrating structural re-parameterization techniques into the OSNet baseline, the network achieves a significant boost in inference speed while maintaining high-fidelity re-identification performance. Second, EIoU is employed instead of traditional IoU in the data association stage to achieve more accurate matching of overlapping targets and faster convergence by incorporating the aspect ratio and center distance of bounding boxes. Third, YOLOv11 is utilized as the core detector to provide high-quality candidate boxes and confidence scores for the tracking pipeline. Experimental results demonstrate that the RepOSNet improves mAP and Rank-1 accuracy by 6.33% and 5.33%, respectively, compared to the original backbone. On the MOT20 dataset, the proposed YOLOv11+RE-BoTSORT method achieves state-of-the-art performance with a MOTA of 72.4%, an IDF1 of 71.8%, and a reduction of IDs to 890, proving its superior robustness in dense environments.

ABSTRACT. Predictive performance in complex systems is often constrained by heterogeneous entities and their dynamic interdependencies. Existing learning-based methods typically fall short in modeling multi-level dependencies and adaptive interactions. We propose DGCNet—a dynamic graph network that integrates multi-level similarity aggregation for predictive tasks. The model constructs graph topologies by fusing multiple similarity measures, eliminating the need for predefined static structures. Hierarchical feature propagation is achieved through a dynamic graph convolution module that captures both fine-grained local patterns and broader contextual relationships. To further enhance representation learning, a dual-view representation fusion mechanism enables bidirectional information exchange between dual views, ensuring complementary features are jointly optimized. Evaluated on real-world dataset, DGCNet consistently outperforms strong baselines across standard metrics. Component-wise analyses further confirm that each architectural choice contributes to improved accuracy, robustness, and generalization.

ABSTRACT. Traditional personality assessment methods have inherent limitations, including social desirability bias, insufficient situational dependence, and restricted large-scale application. To address these issues, this study proposed a lightweight Big Five personality assessment method based on large language model (LLM) agent interaction. Five types of high-trait agents (high Openness, high Conscientiousness, high Extraversion, high Agreeableness, and high Neuroticism) were constructed on the DeepSeek platform using prompt engineering. An unconstrained natural text interaction scenario was designed to collect multi-dimensional behavioral data from users. A dual-logic assessment system was established, consisting of Direct Assessment (DA) and Questionnaire-based Assessment (QA). A total of 48 college students participated in the experiment. The Chinese version of the BFI-44 scale was used as the criterion to validate the method. The results showed a highly significant positive correlation between the proposed method’s assessments and the criterion values. The overall Pearson correlation coefficient was r = 0.71 (p < 0.01). Specifically, DA achieved optimal adaptability for the Extraversion dimension (mean absolute error, MAE = 0.58; r = 0.82), whereas QA performed better for the Conscientiousness dimension (MAE = 0.61; r = 0.79). The average user experience score was 4.35 on a 5-point scale, indicating good system acceptance. This method does not require complex model fine-tuning or system development, realizing a lightweight implementation solution relying on existing LLM platforms. It enables efficient and low-cost personality assessment and provides a practical supplementary approach for psychological assessment and human-computer interaction.

ABSTRACT. Cross-session EEG brainprint recognition is limited by session drift. Label-free Riemannian aligners such as EA and RPA avoid target-session identity labels but are usually defined on a single broadband covariance representation. We study whether alignment itself should be frequency-aware within a low-complexity geometric pipeline. Band-RPA decomposes each trial into alpha and beta bands, applies label-free RPA independently in each band, and fuses the resulting tangent-space features. Under a controlled comparison with Raw, EA, RPA, RC, and their band-wise variants under a shared TS+SVM backend, Band-RPA reaches 97.27% on BCI Competition IV 2a and consistently improves broadband RPA on OpenBMI from 10 to 54 subjects. Compared with deep reference models, it preserves a lightweight label-free geometric deployment model, requiring only band-wise target-session summary statistics rather than a gradient-trained backbone. The results support band-resolved alignment as a useful reformulation of RPA for cross-session EEG biometrics.

ABSTRACT. Human pose estimation (HPE) needs both global dependency modeling and pose-aware reasoning. Existing Transformer-Mamba HPE models still suffer from weak structural priors in keypoint-token initialization and lack explicit pose-aware guidance before recursive state propagation stages. To address these issues, we propose KTMPose, a keypoint-prior guided Transformer-Mamba network with a unified global modeling--pose enhancement--recursive propagation pipeline. Within this unified design, KPGCN serves as the keypoint-token initializer by injecting skeletal topology, and KPGN acts as the pose-enhancement bridge by applying keypoint-guided channel modulation before bidirectional Mamba recursion. On COCO, KTMPose outperforms peer methods and shows a better accuracy-efficiency trade-off against benchmark models. Ablations show that pose-aware tokenization and conditional interaction, rather than deeper stacks or more scans, drives consistent gains. This study provides new insights for hybrid HPE design.

ABSTRACT. This paper presents AttnLight, a real-time single-channel EEG-based brain-machine interface system for adaptive desk lamp control. The system enables intuitive and non-contact lighting interaction by mapping the user’s attention state to brightness regulation commands. A NeuroSky TGAM module is used to acquire raw EEG signals at 512 Hz, and wavelet packet transform is employed to decompose the signals into eight frequency bands for feature extraction. Based on the extracted features, the system estimates the user’s attention level and relaxation level, and uses the attention level as the primary control variable. The control signal is transmitted through MQTT to an ESP32-based lighting unit, where it is mapped to PWM output for adaptive LED brightness adjustment. In addition, a web-based client-server software platform built on Tomcat, Python, and MySQL supports real-time visualization of EEG signals, band-energy variations, and attention and relaxation curves, while also enabling data export. The prototype provides a feasible framework for EEG-driven adaptive lighting in assistive smart desk environments.

ABSTRACT. To address the challenges posed by semantic heterogeneity among multi-source monitoring data and the extreme scarcity of attack samples in modern power systems, this paper proposes a semantic fusion and imbalanced anomaly detection framework, termed SF-IAD. The framework integrates a CAE-GAN-based semantic fusion module with an enhanced deep reinforcement learning detector to improve anomaly recognition under highly skewed class distributions. Specifically, the CAE-GAN module is employed to extract unified latent representations from heterogeneous sources, including network traffic, physical measurements, and system logs, thereby reducing cross-modal discrepancies and strengthening the discriminative capability of fused features. On this basis, an improved deep Q-network is developed for anomaly detection, in which an asymmetric reward mechanism is introduced to increase sensitivity to rare attack events, while prioritized experience replay is adopted to enhance the utilization efficiency of informative minority-class samples during training. Experiments conducted on the HAI dataset under an extreme class imbalance setting demonstrate that the proposed method achieves superior performance over several baseline models in terms of recall, F1-score, and AUC, while maintaining a relatively low false positive rate. The results verify the effectiveness and robustness of SF-IAD for anomaly detection in complex power system environments.

ABSTRACT. Neural painting refers to the process of converting a real image into an artistic picture via a neural network. In this paper, we propose Strokeformer, a novel neural painting method that is capable of generating high-quality painting images. Standard stroke optimization methods can generate a sequence of strokes for this task, but they often fail to capture the fine texture details of real images. For this purpose, we design a new feature extraction module namely Effective Residual Squeeze Excitation (ERSE). In this new module, we propose a lightweight attention mechanism for feature assignment, where a residual structure is used to extract feature information from the canvas at different levels and thus obtain channel-wise enhanced features. This approach not only enables the network to learn stroke features more efficiently but also enhances the detail in the generated paintings. However, while neural painting approaches are effective in producing strokes, their ability to reproduce natural colors is limited. We have designed a new pixel loss function by synchronizing the pixel and color information of the generated images. Finally, we design a stroke filtering algorithm to reject redundant, tiny strokes, thus mitigating the image quality degradation caused by lattice artifacts. Experiments show that our Strokeformer generates better quality images, compared to the other state-of-the-art methods. Compared with classical methods, our approach has improved SSIM and PSNR metrics by 15.25% and 24.78%, respectively.

ABSTRACT. A millimeter-wave (mmWave)-enhanced integrated sensing and communication (ISAC) framework is examined to enhance both communication rates and sensing capabilities for internet of things (IoT) systems. This framework considers scenarios involving multiple target angles, various configurations of weighting coefficients, and an analysis of the tradeoff between communication and sensing. A dual-function base station (BS) supports multiple users using mmWave technology, while simultaneously utilizing the communication signal for diverse target detection purposes. The objective is to optimize the weighted sum of the sensing power and communication rate by formulating a beamforming design problem. A tailored efficient gradient ascent framework is provided to address this issue. The numerical findings indicate that the proposed framework achieves improved sensing-communication tradeoff performance compared to the conventional ISAC system under different user configurations (U = 3 and U = 6).

ABSTRACT. With the advancement of the "Dual Carbon" strategy and the accelerated construction of new power systems, intelligent operation and maintenance (O\&M) based on unmanned aerial vehicles (UAVs), inspection robots, and multi-source visual sensors has become a critical means to ensure the safe and stable operation of the power grid. However, the massive, heterogeneous, and severely long-tailed distribution of multi-modal power image data poses significant challenges to equipment recognition and defect diagnosis in complex scenarios. This paper systematically reviews the evolutionary trajectory and frontier progress of object detection technologies for image data within power systems. First, it provides a profound analysis of the data characteristics of multi-source imaging modalities and elucidates fundamental pain points, such as extreme scale variations, morphological diversity, and intensive background interference. Second, following a brief review of the limitations inherent in traditional feature engineering models, the paper focuses on the evolution of two-stage networks represented by Faster R-CNN, one-stage networks centered on YOLO, and Transformer architectures that transcend the constraints of local receptive fields in the context of power vision. Furthermore, it summarizes domain-specific optimization strategies tailored for small objects, rotated bounding boxes, and lightweight deployment. On this basis, a performance comparison of mainstream object detection algorithms is conducted, analyzing their performance metrics on the COCO dataset. Finally, addressing the bottleneck of insufficient generalization capabilities of current deep learning models under long-tail and adverse operating conditions, the paper prospectively explores core future breakthrough directions, including Generative AI-based few-shot learning, fine-tuning and reconstruction of Vision Foundation Models (VFM) for the power industry, and edge-cloud-terminal collaboration. This work aims to provide a theoretical reference for the transition toward highly reliable cognitive intelligence in power systems.

ABSTRACT. Reliable motor imagery electroencephalography decoding remains challenging due to low signal quality and severe inter-subject variability. To tackle these issues, this study introduces the Multi-Scale Dilated Frequency-Gated Network. The architecture integrates a dilated multi-scale spatio-temporal convolution module to capture diverse local patterns, alongside a Global Gated Temporal Encoder that models extensive temporal dynamics while dynamically suppressing noise. Additionally, a Frequency Aware Positional Encoding mechanism explicitly embeds physiological priors into the sequence. Comprehensive tests on the BCIC-IV-2a and BCIC-IV-2b datasets validate its superiority over existing baselines, achieving peak subject-dependent accuracies of 83.80% and 88.72% respectively. Consequently, the proposed network delivers remarkable decoding precision and adaptability, presenting a robust framework for advanced brain-computer interfaces.

ABSTRACT. Abstract. Emotion has multidimensional characteristics. EEG signals can directly reflect neural electrical activity and serve as an important pathway for emotion recognition. Single-modal methods struggle to fully exploit cross-modal mapping relationships, and insufficient information fusion limits their recognition performance. Multimodal fusion offers the advantages of information complementarity and robustness. Therefore, this paper conducts research on multimodal emotion recognition. For EEG features, we propose a multi-scale spatial EEG feature extraction method, achieving an accuracy of 80.25% on the SEED-IV dataset. For eye movement features, we optimize the original 33-dimensional features to 29 dimensions using one-way analysis of variance (ANOVA). For feature fusion, we design a cross-attention gating fusion model, which achieves a four-class classification accuracy of 86.25%, significantly outperforming single-modal approaches and other state-of-the-art methods.

ABSTRACT. Sonar image segmentation often suffers from insufficient exploitation of local spatial information and sensitivity to outliers. To address these challenges, we propose a robust level set segmentation algorithm driven by the local probability of a Student’s t-distribution mixture model based on Markov Random Field. First, local posterior probabilities from the Markov Random Field -Student’s t-distribution Mixture Model are introduced to fully exploit local spatial features and drive the evolution of the level set, enhancing the algorithm’s robustness to outliers and intensity inhomogeneity. Second, a new geometric prior is further constructed by integrating Markov’s local spatial texture information with the spatial information of the level set function. The spatial details required for segmentation are enriched by projecting Euclidean spatial information into a discrete probability space through Bayesian estimation. Third, the Markov Random Filed -Student’s t-distribution Mixture Mode modifies the evolution process of the level set in the Local Binary Fitting model, and the results of the level set evolution are then reflected back into the Markov Random Filed -Student’s t-distribution Mixture Model, enabling the two models to converge rapidly and achieve accurate segmentation. Finally, to prevent overlap and omission in segmented regions, a cyclic initialization method for the level set function is designed. Experimental results show that the algorithm achieves a segmentation accuracy index consistently above 0.9, confirming its effectiveness for precise sonar image segmentation.

ABSTRACT. With the rapid advancement of the Industrial Internet, smart manufacturing, and cyber-physical systems, multivariate time series anomaly detection has become a cornerstone of safe and reliable operation in complex systems. Context-aware anomaly detection requires not only identifying numerical deviations but also characterizing contextual consistency shaped by temporal dependencies, cross-variable couplings, and operational semantics. This survey reviews multivariate time series contextual anomaly detection from the perspective of evolving context-modeling capabilities. It first examines traditional deep learning methods based on prediction and reconstruction, then summarizes Transformer-based methods for global temporal modeling and structural correlation learning and finally reviews large language model-based approaches that inject semantic knowledge through tokenization, prompt engineering, and multimodal alignment. The survey also discusses major open challenges, including deployment cost, reliability assessment, hallucination mitigation, and evaluation methodology. Overall, the field is moving from numerical context modeling toward semantically enriched context understanding, opening new opportunities for anomaly detection and intelligent diagnosis in industrial environments.

ABSTRACT. Unsupervised anomaly detection aims to identify rare samples that significantly deviate from normal patterns without relying on labeled data. Although existing methods based on statistics, distance, density, clustering, graphs, and deep learning have achieved considerable progress, many of them still adopt a fine-grained sample-level processing paradigm. This often leads to high computational cost, sensitivity to noise, unstable neighborhood relationships, and limited ability to capture multi-granularity data structures. Granular-ball computing, as an emerging multi-granularity representation and computation method, provides a new perspective for addressing these challenges. By adaptively representing the data space with a set of hyperballs, granular-ball computing transforms anomaly detection from point-level modeling to granule-level modeling, thereby improving efficiency, robustness, and interpretability. This paper reviews recent advances in granular-ball computing-driven unsupervised anomaly detection. First, the basic concepts of unsupervised anomaly detection and granular-ball computing are introduced. Then, existing methods are summarized from several perspectives, including local density-based methods, fuzzy rough set and information fusion-based methods, random walk and graph-based methods, and methods for complex data types. Finally, the limitations of current studies are discussed, and future research directions are highlighted, including high-quality granular-ball generation, adaptation to high-dimensional and mixed-type data, integration with deep learning, extension to graph and multimodal anomaly detection, and theoretical interpretability analysis.

ABSTRACT. Anomaly detection plays a critical role in identifying rare faults, attacks, frauds, and abnormal behaviors in distributed systems, such as industrial Internet of Things, power systems, cybersecurity, financial risk control, and smart devices. In practical scenarios, anomaly detection is often constrained by two factors simultaneously: anomalous samples are usually scarce and highly imbalanced, while raw data may contain sensitive user, device, or operational information that cannot be centrally collected or directly shared. These joint constraints make conventional centralized anomaly detection methods insufficient. On the one hand, privacy-preserving mechanisms may introduce noise, communication overhead, and encryption-related computational costs. On the other hand, imbalanced learning requires models to maintain high sensitivity to weak, rare, and evolving anomalous patterns.This paper reviews anomaly detection methods under the joint constraints of privacy protection and sample imbalance. We first summarize the evolution of anomaly detection techniques, ranging from traditional statistical methods and shallow machine learning approaches to deep learning-based reconstruction, prediction, and representation learning methods. We then analyze privacypreserving techniques, including federated learning, differential privacy, secure aggregation, and homomorphic encryption, and discuss their impacts on anomaly sensitivity, minority-class learning, and distributed optimization. Furthermore, we review imbalance-aware strategies, including loss reweighting, generative augmentation, and few-shot adaptation, with emphasis on their roles in improving the detection of rare anomalies. Finally, we discuss representative applications in modern power systems and identify future research directions, including privacy-aware generative augmentation, anomaly-sensitive federated optimization, reliable synthetic anomaly generation, and lightweight privacy-preserving deployment.

ABSTRACT. Abstract: Modern power systems are evolving toward smart grids and the energy internet. While the high-precision data exchange among vast numbers of smart devices facilitates condition monitoring and optimized dispatch, it also raises serious data privacy and security concerns. Drawing on recent domestic and international research findings, this paper provides a systematic review of current developments in the field of privacy protection in power systems, systematically outlining three key categories of privacy protection technologies and their applications in power systems. This paper classifies privacy-preserving technologies in power systems into three categories: First, cryptography-based strong protection methods, with a focus on privacy-preserving technologies such as AES, RSA, and homomorphic encryption, which are used to achieve secure transmission and aggregation of data in the ciphertext domain; second, lightweight anonymization methods based on data perturbation, with a focus on differential privacy techniques, which satisfy the real-time processing requirements of high-frequency data by introducing statistical noise; Third, distributed methods based on federated computing, which eliminate the risk of raw data leakage at the source, with a focus on frameworks such as lateral, vertical, and federated transfer learning. Finally, to address the shortcomings of current privacy protection methods, this paper outlines future directions for technological research.

ABSTRACT. Human-machine shared driving requires timely intervention that improves vehicle behavior without unnecessarily overriding the nominal driver. This paper presents a short-horizon optimization-based shared-driving strategy in a closed-loop driving simulator. The framework integrates the Assetto Corsa driving simulator, real-time telemetry acquisition, hot-start synchronization of a high-fidelity vehicle model, forward prediction, optimization-based command generation, and feedback intervention. At each intervention instant, candidate commands are evaluated over a finite horizon using a cost function considering trajectory tracking, forward progress, control smoothness, boundary proximity, and drift-related risk. To examine the operating range of the strategy, the Assetto Corsa AI driver is used as the nominal driver, and AIpush is varied to generate different aggressiveness levels. Results show that the strategy improves tracking under low-to-moderate aggressiveness, especially at AIpush=80, 90, 93, and 94, but becomes marginal or harmful under excessive aggressiveness. Segment-level results show more consistent benefits in connecting segments than in corners.

ABSTRACT. Epigenetic clocks are valuable tools for assessing biological age, but present models face challenges such as low prediction accuracy for small samples, inadequate representation of higher-order dependencies, etc. In this study, we propose an adaptive weighted feature selection and gradient optimization strategy-based biological age prediction model, namely AWFS-GO. Firstly, the model employs an adaptive weighted feature selection mechanism that integrates five evaluation methods to obtain a comprehensive importance score. It adaptively determines the screening ratio and weighting thresholds based on the cumulative contribution of the feature importance distribution, thereby enabling adaptive adjustment of feature selection. Second, we propose a contrastive learning-based data augmentation strategy to expand the training dataset and mitigate the overfitting in small-sample scenarios. Then, an enhanced Transformer encoder is used to capture the long-range dependencies among methylation sites, and attention pooling is applied to compress the high-dimensional features. After training the model, a reinforcement learning framework based on the PPO (Proximal Policy Optimization) algorithm is proposed to fine-tune the model. This framework achieves gradient-based policy optimization by adaptively adjusting the learning rate by observing state information, such as the gradient norm. Finally, we ensemble multiple base models into a weighted prediction system to improve predictive capabilities. Experiments on real-world methylation datasets demonstrate that AWFS-GO outperforms existing baseline models on most metrics. In summary, AWFS-GO provides an effective solution for epigenetic clock prediction.

ABSTRACT. Multimodal survival prediction has shown great potential for tumor prognosis assessment. However, existing methods are limited by suboptimal multimodal fusion performance, as well as the high computa tional and memory overhead of self-attention for ultra-long whole-slide image (WSI) sequences. This study proposes SS-MEF, a state space based multimodule fusion framework for lung adenocarcinoma survival prediction. The framework combines pathology-specific representation learning with state space sequence modeling to jointly analyze histopatho logical and genomic data. Pathology-specific pretraining is used to obtain histopathological features. The state space module is then applied to model ultra-long WSI sequences together with genomic features under linear computational complexity, supporting efficient multimodal sur vival prediction. Experimental results on both the TCGA-LUAD and CPTAC-LUAD datasets demonstrate that SS-MEF achieves accurate and effective survival prediction across all the evaluation metrics. Compared with multimodal baselines, SS-MEF improves the C-index by 2.3%–10.2% on TCGA-LUAD and by 2.8%–8.6% on CPTAC-LUAD. It also achieves the lowest Log-rank p-value among all baselines, suggesting improved separation between the predicted high-risk and low-risk patient groups. Meanwhile, SS-MEF reduces GPU memory consumption by 26.8%–63.8% and inference time by 34.0%–55.6%. Overall, SS-MEF provides an effective framework for multimodal survival prediction.

ABSTRACT. Accurate modeling of mine backfill pipeline networks is critical for industrial system reliability. Conventional pathfinding methods frequently struggle with complex three-dimensional terrains and strict fluid constraints. This paper pro-poses an intelligent simulation framework for the multi-objective topological optimization of these networks. Driven by a digital twin model, the framework integrates Prüfer sequence encoding for acyclic topological modeling and a Pa-reto-based evolutionary engine for simulation optimization. Validated through computational experiments using real-world topographical data from Inner Mongolia, this approach significantly outperforms conventional baselines such as minimum spanning trees and standard genetic algorithms. It effectively bal-ances economic efficiency (minimizing capital and operational costs) and op-erational safety (limiting hydraulic gradient risks). The proposed strategy de-livers a robust, highly adaptable decision-support tool for large-scale industrial engineering in complex environments.

ABSTRACT. Neural machine translation has been widely used in medical text processing and cross-lingual knowledge services, but Traditional Chinese Medicine (TCM) translation remains challenging due to specialized terms, culture-loaded expressions, and theory-dependent concepts. General-purpose translation models may produce term mistranslation, terminology omission, semantic shifts, and unstable outputs. To improve terminology preservation and semantic consistency, this paper proposes an entity-enhanced neural machine translation framework based on NLLB and LoRA. The framework constructs entity-enhanced inputs from TCM terminology resources and applies LoRA-based domain adaptation. Four settings are evaluated with automatic metrics and further examined through multi-seed testing, bootstrap significance testing, and manual error analysis. The multi-seed average results show that the proposed model achieves BLEU, chrF, TER, TA, and TCR scores of 31.29, 52.61, 55.42, 60.78, and 77.18, respectively. Manual analysis of seed-42 outputs further shows that entity-enhanced input helps reduce term omission, semantic shifts, incomplete generation, and repetitive outputs. These results indicate that terminology-aware input combined with parameter-efficient adaptation is useful for TCM machine translation.

ABSTRACT. Seals in complex government documents often exhibit characteristics such as small size, occlusion, blurring, and text overlap, leading to insufficient robustness of traditional OCR methods. Existing approaches that employ layout detection combined with OCR, such as the mainstream PP-DocLayout in government service scenarios, face typical problems including high miss detection rates, low detection accuracy, and weak anti-interference capabilities, lacking hierarchical visual parsing mechanisms specifically designed for seal regions. To address these challenges, this paper proposes a hierarchical seal detection and recognition framework for complex government documents. A dedicated seal detection model YOLOv11_Seal is trained using a specific dataset of attachment review documents from provincial science and technology plan project review affairs, the detection boxes are utilized as structured segmentation prompts to guide segmentation, and after fine segmentation by SAM, the complete seal regions are fed into PP-OCRv4 for seal text detection and recognition, thereby constituting the "coarse detection-fine segmentation-detailed recognition" cascade collaborative mechanism. Furthermore, ablation experiments are conducted to verify the effectiveness and generalization capability of the framework on an independent test set of real samples. Compared with layout detection models, the seal detection success rate of sample images is improved from 54.76% to 86.31%, and the single seal detection recall rate is increased from 58.26% to 93.46%. The trained YOLOv11_Seal model achieves an mAP@0.5 of 97.2% and an F1-Score of 97.0%. Experimental results demonstrate that the framework significantly enhances the detection capability of seal regions under complex backgrounds and can effectively adapt to the actual visual parsing requirements of complex government documents. However, as the coverage of seal target detection expands, the seal text recognition accuracy under the ablation experiment definition standards has been affected to a certain extent, which will be addressed in future research.

ABSTRACT. Abstract. With the increasing demand for intelligent and flexible operations in industrial manufacturing, sample transportation has become an important component of smart factory construction. In practical factory environments, samples are frequently transferred among production stations, inspection areas, and management areas. Conventional manual transportation is affected by personnel flow, long-distance transfer, route complexity, and environmental variation, which may result in high labor intensity, unstable efficiency, and limited task continuity. To address these problems, this paper proposes an autonomous sample transportation system based on a quadruped robot. A Go2 Edu quadruped robot is used as the mobile execution platform, integrating LiDAR, an onboard computing unit, a communication module, and a sample-carrying device. A layered architecture is developed, including the robot execution layer, perception and computation layer, communication and interaction layer, and remote monitoring layer. Based on this architecture, an autonomous navigation framework and a transportation task interaction framework are designed for multi-point sample transfer. Field tests in a real factory environment demonstrate that the system can perform point-cloud mapping, ICP-based initial localization, multi-point path loading, autonomous navigation, dynamic obstacle avoidance, and button-triggered task execution. The results indicate that the proposed system provides a feasible engineering solution for autonomous sample transportation in complex factory environments.

ABSTRACT. Degradative anonymization is widely used in face privacy protection because it reduces identity recognizability; however, it often degrades the natural visual quality of facial images. This makes post-anonymization face reconstruction an important problem. Existing PULSE-based reconstruction methods usually adopt globally uniform error constraints, which are insufficient for emphasizing key facial regions such as the eyes, nose, mouth, and facial contour, thus limiting local structural recovery. To address this limitation, we propose a region-aware adaptive reconstruction method based on PULSE. Our method introduces a prior soft mask for facial regions and integrates it with an errordriven adaptive weighting mechanism to enforce stronger constraints on important local areas. We further design two variants to study different regional priors: a center-guided scheme and a landmark-guided scheme. The latter builds the soft mask from dense landmark heatmaps generated by MediaPipe Face Mesh and automatically falls back to a center prior when landmark detection fails. Experiments on CelebA-HQ show that the proposed method effectively improves reconstruction quality in facial regions. The center-guided scheme yields stable improvements, while the landmark-guided scheme exhibits stronger ability to enhance fine local structural details. These results indicate that the proposed method can improve local face reconstruction quality in privacy-preserving scenarios with only minor changes to the original PULSE framework.

ABSTRACT. This paper proposes an enhanced Faster R-CNN method integrating two attention mechanisms: the Squeeze-and-Excitation module and the Convolutional Block Attention Module. The SE module adaptively recalibrates channel-wise feature responses by explicitly modeling interdependencies between channels, effectively suppressing irrelevant information while emphasizing informative features. The CBAM module extends this idea by jointly attending to both channel and spatial dimensions, enabling the model to focus on "what" is important and "where" it is located. To fully leverage their complementary strengths, we further propose a sequential CBAM+SE combination, where the input feature map first passes through CBAM for joint channel-spatial recalibration, followed by SE for refined channel-wise weighting. Extensive ablation studies are conducted on the KAIST multispectral pedestrian detection dataset using 15 test images. Experimental results demonstrate that the SE module reduces the mean feature activation by 50.0%, while CBAM achieves a 75.1% reduction. Notably, the proposed CBAM+SE combined module further reduces activation by 87.5%, outperforming either module used alone. Qualitative visualizations confirm that the attention-enhanced models produce sparser and more focused feature maps, concentrating on object-relevant regions while effectively suppressing background noise. These results validate that channel and spatial attention are complementary, and their integration significantly improves feature representation for object detection tasks.

ABSTRACT. In recent years, State Space Models (SSMs), especially Mamba, have attracted increasing attention in visual tasks due to their favorable balance between computational efficiency and modeling performance. However, most existing visual Mamba methods flatten images into 1D sequences using predefined scan orders. This may place semantically related regions far apart in the sequence. As a result, their interactions may be weakened by dependency decay in long-sequence modeling. In addition, such sequence construction may disrupt the original spatial structure of images, making it difficult to preserve the continuity between neighboring regions. To address these issues, we propose SGI-Mamba, a Semantic Group Interaction visual Mamba network, for image classification. SGI-Mamba first designs the SPSA module to reconstruct visual sequences from both cluster and spatial perspectives. Specifically, to alleviate dependency decay in long-sequence modeling, SPSA assigns related tokens to the same cluster group according to a common principle: those tokens with similar feature embeddings and close spatial locations are more likely to be assigned to the same group. This reduces the modeling distance between them. Then, SPSA reorders tokens within each semantic group according to local spatial continuity. Considering that variable-length of cluster groups are not beneficial to highly efficient parallel computation, these cluster groups are further divided into fixed-length subgroups for modeling. However, independent subgroup-wise modeling restricts information flow among different subgroups. To address this limitation, SGI-Mamba introduces a global subgroup interaction mechanism to promote information exchange across subgroups. Together, these components construct a cluster- and spatially aware visual sequence that supports more effective long-range dependency modeling. Experiments on the ImageNet-1K image classification task show that SGI-Mamba achieves 76.9% Top-1 accuracy with only 8.01M parameters and 1.28G FLOPs, demonstrating the effectiveness and efficiency of the proposed method. The source code has been publicly released.

ABSTRACT. Diffusion models have achieved remarkable success for image generation, but their multi-step denoising leads to slow generation speeds. Consistency models overcome this limitation by learning a direct noise-to-data mapping, but they often suffer from low generation quality and high training costs. ACT-Diffusion uses a discriminator to directly minimize the Jensen-Shannon (JS) divergence between generated and real distributions, thereby enhancing generation quality and convergence. However, when the generated and real distributions have non-overlapping supports, the JS-divergence becomes nearly constant and provide weak or even vanishing gradients, which slows the alignment of the generated distribution toward the real one. To address this issue, we propose Wasserstein-Distance(WD) Guided Consistency Training(WGCT), which incorporates a WD-based discriminator into consistency training to replace the JS-divergence-based discriminator. Since the Wasserstein distance can provide smooth and meaningful gradients even when two distributions do not overlap, WGCT offers more effective distribution-level guidance and is less sensitive to the initial degree of overlap between the generated and real distributions, thus accelerating convergence and reducing training costs. We further observe that fixed alternating generator-discriminator updates can destabilize the later training stage, leading to training collapse and degraded generation quality. To alleviate this problem, WGCT increases the generator update frequency in the late training stage, which improves training stability and generation quality. The experimental results demonstrate that our method achieves better FID scores on CIFAR10 and ImageNet datasets, but only requiring approximately 60% training steps compared to ACT-Diffusion.

ABSTRACT. This paper proposes SFE-CMAF, a multimodal classification model for social network image-text posts, to address the limited recognition accuracy caused by unimodal methods insufficient semantic correlation mining and traditional fixed-weight fusion’s poor adaptability to diverse social content. The model builds an SFE feature extraction structure with ResNet50 and BERT-Chinese as backbones to extract highly discriminative core features, and adopts a CMAF cross-modal attention strategy to realize bidirectional interaction and adaptive dynamic fusion of image and text features. Experiments on a self-built 3500-sample Chinese social dataset of 5 categories show that the model achieves a weighted F1-score of 0.8755, 21.87, 8.91 and 2.02 percentage points higher than text-only, image-only and weighted fusion baselines respectively, with 100% F1-score for 3 categories and 88.5% recall rate for online rumors, presenting superior classification performance and good practical application value.

ABSTRACT. This study presents a low-cost, intent-controlled direct command sys tem based on a non-invasive brain-computer interface. The system uses an Emotiv EPOCX 14-channel EEG headset to acquire motor imagery signals, which are classified into three motor imagery tasks corresponding to control commands: left hand movement, right hand movement, and both feet. To sup press transient misclassifications caused by ocular and muscle artifacts, we in troduced a majority voting smoothing mechanism with a queue length of five, which reduces servo jitter by approximately 90%. The host PC performs real time sliding-window inference and sends single-byte commands via UART to an ESP32 microcontroller, which generates 50 Hz PWM signals to drive an SG90 servo. Offline experiments on the BCI Competition IV-2a dataset yield a three-class classification accuracy of 75.0% and a cross-entropy loss lower than 0.15. The end-to-end system latency was measured to be less than 0.65 s. The proposed system provides a practical, stable, and affordable solution for brain controlled motor actuation.

ABSTRACT. Cooperative control of multiple legged robots poses a tightly coupled learning problem in which local dynamic stability and team-level coordination must be optimized simultaneously under contact-rich interactions. This paper investigates a cooperative transport task in which two homogeneous ANYmal-C quadrupeds jointly regulate the linear and angular motion of a rigid shared bar inside the NVIDIA Isaac Lab simulator. We formulate the task as a decentralized partially observable Markov decision process (Dec-POMDP) and solve it with a multi-agent reinforcement learning pipeline built on the Heterogeneous-Agent Reinforcement Learning (HARL) framework and the Heterogeneous-Agent Proximal Policy Optimization (HAPPO) algorithm, following the centralized training with decentralized execution (CTDE) paradigm. The key design choice is an object-centric reward that couples both agents to the same tracking objective, inducing cooperative behavior without explicit inter-agent communication during execution. The methodology is grounded in the official environment specification: a 48-dimensional local observation, a 12-dimensional joint-target action, startup randomization of contact and inertial parameters, a shared object-centric tracking reward, and failure-aware termination conditions for robot falls and object collapse. We conduct a controlled ablation over three reward configurations (R0, R1, R2) using up to 5.0 M environment steps per run. Under the tuned configuration (R2), the task success rate improves from 62.7% to 71.4%, the coordination score increases from 0.53 to 0.66, the collision rate decreases from 18.6% to 15.7%, and the yaw tracking RMSE decreases from 0.43 to 0.35 relative to the baseline. These results demonstrate that object-centric reward design combined with coordinationaware reward shaping yields measurable improvements in cooperative stability and cooperation quality for homogeneous quadruped teams

ABSTRACT. Abstract.To tackle drastic scale variation, frequent occlusion and hard-to-quantify congestion in complex traffic scenes, this paper presents a CBAM-enhanced YOLOv8s detector and a dual-dimensional normalized congestion evaluation method based on object quantity and area.CBAM is embedded into the shallow feature extraction of YOLOv8s backbone to boost feature representation of small targets and key regions. A composite congestion index is established for automatic four-level traffic classification and alarming.Experiments on a self-built traffic dataset show the improved model reaches 93.8% mAP@0.5,which achieves a 3.9% improvement over the original YOLOv8s.The mAP@0.5:0.95 metric reaches 81.0%, which is 8.7% higher than the original YOLOv8s. A PyQt visualization platform is developed to realize image/video detection, target counting, CSV export and sound-light alarms for severe congestion, offering technical support for intelligent traffic monitoring.

ABSTRACT. Power-equipment visual systems require abnormal samples for detector and segmenter training, but real defects are rare, expensive to annotate, and often tied to specific components. Recent text-to-image and image-editing models provide powerful synthesis capabilities, yet smaller or open-domain generators often place defects on invalid regions or ignore physical constraints in power-equipment imagery, while stronger large generators make the cost of large-scale data construction difficult to control. To generate qualified abnormal samples under controllable cost, this paper studies controllable abnormal image generation for power equipment as a constrained visual-data construction problem. We propose GridDefectGen, which first trains student generators from teacher-produced normal/abnormal pairs, masks, and edge maps, then builds equipment-defect generation conditions from a detection knowledge base, and finally uses multi-path generation to output abnormal images together with synchronized masks, boxes, metadata, and class labels. Experiments on four public power-equipment datasets and one private field-image set show consistent gains in downstream detection/segmentation utility and controllability: GridDefectGen improves held-out abnormal F1 over real-only training by 16.6 points on CableInspect-AD and 18.7 points on the private set, while also achieving the best placement validity, region leakage, mask IoU, and box IoU among the evaluated generators.

ABSTRACT. Document Visual Question Answering (Document VQA) answers natural-language questions over a document image, where the evidence may sit in headings, tables, charts, or handwritten notes. The task underpins practical applications such as contract review, financial-report analysis, and infographic understanding. On a layout-dense page, the answer hides in just a few OCR tokens, and locating them efficiently is a key bottleneck for Document VQA. Methods built for natural images estimate saliency on a patch grid and average all attention heads, which does not fit documents whose evidence is discrete OCR tokens.To address this, we propose QPEH (Question-Peak Expert Heads). We find that the attention peaks of question tokens point to the OCR-token region holding the answer more reliably than the [CLS] or pooled signals used before, and that only a few heads carry this ability. QPEH selects a shared pool of these expert heads with three complementary metrics, Normalized Precision, Recall, and MRR, rather than averaging all heads. At inference, the selected heads produce a token-level saliency map that is added to the upper layers as an attention bias, steering the model toward the answer without extra parameters or any change to the pretrained weights. With LayoutLMv3-large as the backbone, QPEH improves consistently over the baseline on DocVQA, InfoVQA, and DUDE (extractive), raising ANLS from 80.24 to 81.39, 40.16 to 40.76, and 53.53 to 54.08, showing that the added saliency region improves the model's document understanding.

ABSTRACT. This paper proposes an improved ship target detection algorithm based on YOLOv11-OBB for remote sensing imagery, aiming to address the performance degradation caused by large-scale variations, arbitrary orientations, and complex background interferences. By incorporating a P2 high-resolution detection layer into the feature pyramid to construct a four-scale P2–P5 architecture and embedding the CBAM attention module via residual connections for adaptive feature refinement, the proposed method effectively enhances the detection of small-scale ships while suppressing false positives. Evaluated on the HRSC2016 dataset, the improved model achieves 94.3% mAP50 and 75.0% mAP50–95, outperforming the baseline by 2.61 and 2.32 percentage points, respectively, while maintaining a compact model size of 2.66M parameters and a high inference speed of 98 FPS.

ABSTRACT. Stanford Type A aortic dissection is a life-threatening cardiovascular emergency with high mortality, where accurate identification of the entry tear location is essential for surgical planning and risk stratification. However, in CTA images, entry tears are extremely small, poorly defined, and highly variable in shape, making manual interpretation difficult and operator-dependent.To address these challenges, we propose SG-MA3DNet, a segmentation-guided multi-scale attention 3D network for entry tear location classification. The method first employs a 3D segmentation framework based on nnU-Net, enhanced with a multi-scale depth-wise convolution module and a residual linear attention feature fusion module to better capture aortic anatomy, true false lumen boundaries, and potential tear-related structures.In the classification stage, segmentation outputs are used to construct multi-channel 3D ROI patches, integrating CTA intensity, aortic masks, false lumen cues, and weak tear localization priors. These enriched inputs are then fed into a 3D CNN to classify entry tear locations into four categories . Experiments on the Image TAAD dataset with case-level 5-fold cross- validation show that the proposed method achieves an accuracy of 83.75% and a macro-F1 score of 84.82%. For segmentation, the model reaches a Dice Similarity Coefficient of 53.71% and an IoU of 36.73%. Results demonstrate that incorporating anatomical priors and weak localization cues from segmentation significantly improves classification performance. Compared with relying on pixel-level tear segmentation alone, modeling structured aortic anatomy and region-level relationships provides more robust and informative guidance for entry tear location classification in Stanford Type A aortic dissection.

ABSTRACT. Achieving high-quality Video Super-Resolution (VSR) under limited computational resources remains a challenging problem. Although bidirectional recurrent networks have shown strong capability in exploiting temporal information, most existing methods still process all frames with a fixed propagation, fusion, and refinement strategy. Such a static computation paradigm neglects the inherent variation in frame-level reconstruction difficulty, which may lead to redundant computation for simple frames and insufficient restoration capacity for challenging frames with complex motion or rich textures. To address this issue, we propose a Frame Difficulty-driven Bidirectional Adaptive Network (FDBANet). Specifically, we first design a Motion-Detail Decoupled Propagation Network (MDDPN), where a lightweight forward branch propagates motion cues and a deeper backward branch restores high-frequency details, thereby reducing the interference between motion alignment and texture reconstruction. Then, a Frame Difficulty Evaluation Block (FDEB) is introduced to estimate a compact difficulty index from each low-resolution frame. Guided by this index, an Adaptive Temporal Fusion Block (ATFB) dynamically adjusts multi-scale temporal aggregation weights, while a Dynamic Depth Refinement Block (DDRB) adaptively assigns different refinement depths according to frame complexity. In this way, FDBANet allocates computational resources in a frame-adaptive manner rather than uniformly processing all frames. Extensive experiments on multiple standard VSR benchmarks demonstrate that FDBANet achieves a favorable trade-off between reconstruction accuracy and model complexity, validating the effectiveness of difficulty-driven adaptive computation for efficient video super-resolution.

ABSTRACT. This paper investigates the optimal control problem under the trajectory tracking of the flexible-joint robot(FJR) system. The existing approach, combining continuous sliding mode control (CSMC) with generalized proportional integral observers (GPIOs), suffers f rom c umbersome p arameter t uning a nd i nadequate d isturbance rejection.To address these limitations, an adaptive parameter optimization framework using deep reinforcement learning (DRL) is proposed. Firstly, a CSMC-GPIOs control environment for the FJR system is established. Then, a real-time interaction interface module between DRL and the control system and a composite reward function are designed. The controller’s state quantities are mapped to a standardized state space that DRL could handle, and the control performance under new parameters is calculated in real time based on the state quantities. The DRL agent utilizes the deep deterministic policy gradient (DDPG) approach for training, and the adaptive tuning and optimization of the CSMC and GPIOs parameters are realized through the parameter modification mechanism.The results on the FJR system demonstrate that the optimized control method exhibits enhanced capability in suppressing both matched and mismatched disturbances.

ABSTRACT. Existing literature focuses broadly on the general digital twin framework or standalone teleoperation systems. This review addresses this critical gap by estab-lishing a research scope strictly focused on the algorithmic intersection of digital twin controller designs for teleoperation transparency. The objective is to evaluate how digital twins are embedded into bilateral control loops to isolate the operator from network latencies and break the stability transparency trade-off. To ensure objectivity, a systematic review protocol inspired by the PRISMA guidelines was implemented. A comprehensive screening of published records using specific targeted keywords and strict inclusion and exclusion criteria was conducted, and the results were synthesized. The literature was organized around a coherent framework categorized by the digital twin's architectural role, controller types, and formal stability proofs. A critical comparative analysis revealed clear domain-specific superiorities. For low-latency industrial tasks, replacing fixed observer gains with Long-Short Term Memory LSTM predictors reduced synchronization errors. For high-delay space and deep-sea teleoperation, faster-than-real-time (FTRT) predictive twins minimized tracking error. In safety-critical telerobotics, physics-informed neural networks (PINNs) and model reference adaptive systems successfully compensate for unmodelled interaction forces. The systematic review provides practitioners with an actionable cross-domain mapping tool. By framing communication latency as a model correction problem, it bridges theoretical stability proofs with practical deployment constraints. Finally, it defines a future roadmap to address remaining engineering bottlenecks in sub-millisecond edge computational loops, cloud-based cyber-physical security vectors, and human cognitive factors like cross-cultural trust.

ABSTRACT. Attention Deficit and Hyperactivity Disorder (ADHD) affects nearly 7% of children in China, posing significant obstacles to cognitive development. Traditional interventions such as medication have inherent limitations and potential side effects. This paper presents a non-invasive brain-computer interface (BCI) concentration training system that converts real-time EEG-based attention levels into speed control signals for a toy race car. The system uses a TGAM module to acquire prefrontal EEG signals, an HC-05 Bluetooth module for wireless transmission, and an Arduino UNO as the main controller, while an IRF520 MOSFET module regulates the driving voltage. The NeuroSky eSense algorithm extracts real-time concentration indices (0–100), which are mapped to PWM signals and converted to DC voltage to control the car. A Python-based program realizes real-time visualization of concentration and meditation curves. Experimental tests on ten adults under three mental states confirm that the system-derived concentration values align well with subjective self-rated scores. With a latency below 200 ms, the system supports real-time concentration monitoring and sustained concentration training, suitable for home-based ADHD intervention and parent-child interaction.

ABSTRACT. Production diffusion-model services exhibit heavy-tailed and time-varying latency, yet a globally calibrated upper bound can conceal severe failures for particular base-model, task, and LoRA-presence workloads. We study this problem on the public Alibaba dataset trace using 26,392 successfully completed requests. The 95th-percentile execution time rises from 62~s in the training block to 85~s in the held-out block; a nominal 90% gradient-boosted quantile consequently covers only 79.92% of test requests, and static split conformal prediction covers 86.25%. We introduce a deployment-oriented replay protocol in which a request updates a calibrator only at its completion time, and evaluate a lightweight delayed hierarchical adaptive conformal inference (our method) adapter. It maintains leaf-specific error feedback while backing sparse residual pools off to a task/LoRA parent and a global root. We do not claim a new group-conformal principle: GCACI, POGO, and multivalid prediction are direct predecessors. The contribution is a completion-aware service design and its controlled production-trace evaluation. At a 90% target, global ACI attains 90.05% overall coverage but undercovers the worst of 13 count-supported groups by 20.98 percentage points and emits finite bounds for 99.41% of test requests. Our method attains 90.59% overall coverage, limits worst-group undercoverage to 0.88 points, and remains finite throughout. A 1,000-replicate moving-block analysis of the fixed prediction trajectory gives an 11.11--31.09 point improvement interval over global ACI; their common-finite headroom difference interval includes zero. The held-out pattern persists at 80% and 95% targets. We separately show that leaf-calibration-cold groups remain unsolved and that latency-only claims are conditional on successful completion.

ABSTRACT. Single-view X-ray object detection is widely used in high-throughput security inspection because of its low deployment cost, but it lacks the structural complementarity available from paired dual-view scans. In this paper, we study how dual-view X-ray data can be used as privileged information during training while keeping the deployed detector strictly single-view. We propose G2D-Net, a regime-aware Learning Using Privileged Information (LUPI) framework that transfers conveyor-aligned cross-view structural priors from a dual-view teacher to a single-view student. Instead of treating long-tailed X-ray data as a homogeneous imbalance problem, G2D-Net softly associates categories with frequency-related regimes and applies specialized objectives: material-structure disentanglement for frequent confounded classes, adaptive manifold calibration for mid-frequency diverse classes, and prototype-guided recovery for rare threat categories. The method uses constrained cross-view correspondence, but it does not require explicit depth reconstruction or calibrated 3D recovery. During inference, the teacher branch, auxiliary view, and privileged losses are removed, so the student keeps the same architecture, parameter count, and theoretical FLOPs as the single-view detector. Experiments on the dual-view LIXray benchmark show a 9.1% absolute mAP gain in the T-Scarcity regime, and auxiliary single-view experiments on public X-ray datasets further support the robustness of the student-side design.

ABSTRACT. The capture of images in real-time for digital twin technology can be very com-plex. However, several efforts have been made to achieve an accurate representa-tion of physical images in the virtual space. In this work, a few applications of LiDAR scanning were reviewed. Based on the analysis of each approach, the most recent approach that adopts recent technologies, seems promising. This ap-proach is a combination of intelligent methods of data capture and image pro-cessing. In addition, commonly used image processing software was used mak-ing the process easily replicable and adaptable for other applications.

ABSTRACT. This paper investigates front-rear nozzle clearance regulation for a hydraulic collector head used in deep-sea polymetallic-nodule mining. Instead of regulating only the collector-center height, the actual front and rear nozzle clearances are selected as constrained outputs so that collector pitch, local seabed slope, and nozzle spacing are represented directly in the control objective. A control-oriented model is formulated for a collector head driven by a height-control hydraulic cylinder and a coupled angle-control cylinder group. An efficiency-oriented clearance domain and a wider safety domain are converted into a flexible prescribed-performance boundary. Filtered terrain-rate, actuator-saturation, and boundary-proximity indicators temporarily relax this boundary when the nominal efficiency requirement is difficult to realize. A filtered lumped-disturbance estimate and an input-saturation auxiliary system are incorporated into a barrier-Lyapunov-based controller. A Lyapunov analysis establishes uniform ultimate boundedness of the error-system states and non-contact of the flexible clearance boundaries under the stated feasible operating conditions. Simulation studies evaluate the method under terrain variation and severe input saturation.

ABSTRACT. Speech lock denotes caregiver utterances that restrict a care recipient's actions, intentions, or choices and may compromise autonomy. Such utterances often resemble ordinary requests or safety guidance, making keyword- or similarity-based detection unreliable. This study evaluates whether domain-informed prompting improves sentence-level speech-lock classification with GPT-4o mini. The proposed zero-shot prompt specifies an operational definition, positive criteria such as commands, prohibitions, coercion, and denial, exclusion criteria for supportive or autonomy-respecting utterances, an utterance-only decision rule, and a structured JSON response. On a balanced dataset of 100 Japanese caregiving utterances (50 Speech Lock and 50 Normal), GPT-Proposed was compared with a Doc2Vec baseline and GPT-Simple, which used the same model without domain-specific criteria. GPT-Proposed achieved an accuracy of 0.970, a precision of 1.000, a recall of 0.940, and an F1 score of 0.969, exceeding Doc2Vec (F1 = 0.809) and GPT-Simple (F1 = 0.701). It reduced false negatives from 12 and 23, respectively, to 3 while producing no false positives. These results indicate that explicit domain criteria can substantially improve zero-shot LLM classification. However, the small curated dataset, single-annotator labels, utterance-only inputs, and single-run evaluation limit conclusions about real-world deployment.

ABSTRACT. This study evaluates whether GPT-5.5 can remove synthetically inserted pseudo-tics from Japanese speech-like text while preserving the wording and intended meaning of the underlying utterance. Because large-scale collection of authentic conversational data from people with tic disorders is constrained by privacy and consent requirements, we constructed a controlled synthetic benchmark. Here, pseudo-tics denote artificial word or vocalization sequences that reproduce selected surface characteristics of phonic tics; they are not intended to model the full clinical variability of tic disorders. One hundred Japanese base texts were generated across ten everyday domains. Each base text was combined with three punctuation and insertion conditions and five pseudo-tic categories, yielding 1,500 instances. Each in-stance was submitted independently to GPT-5.5 through the Responses API with a prompt requiring minimal editing, removal of only unintended insertions, and natural punctuation. The primary metric was the normalized exact-match rate after NFKC normalization and removal of spaces and punctuation. Character-level deletion precision, recall, and F1 were computed to characterize non-exact outputs. GPT-5.5 exactly restored 1,443 of 1,500 instances (96.2%). Overall deletion precision, recall, and F1 were 0.9988, 0.9998, and 0.9993, respectively. Exact-match rates were 97.2% for punctuated inputs, 96.4% for unpunctuated inputs, and 95.0% for dense unpunctuated inputs. Most mismatches reflected over-deletion, orthographic changes, minor paraphrasing, or unnecessary additions rather than residual pseudo-tics. These findings indicate strong performance on this synthetic benchmark, but they do not establish clinical effectiveness; evaluation on consented real speech, automatic speech-recognition outputs, and human judgments remains necessary.

ABSTRACT. Acoustic monitoring provides a low-burden means of screening sleep-related sounds, yet many prior sleep-bruxism studies have relied on small or con-trolled datasets. This study evaluates interpretable temporal acoustic features on the large-scale DreamCatcher dataset for three-class classification of brux-ism, snoring, and breathing. Six feature configurations were derived from MFCCs, zero-crossing rate, root-mean-square (RMS) energy, spectral centroid, delta features, spectral bandwidth, roll-off, and spectral flux. The configurations isolate the effects of signed versus absolute delta means, sample standard deviations, dynamic ranges, and exponential transforms of selected ranges. Sequential minimal optimization (SMO) and random forest (RF) were compared under (i) an exploratory balanced stratified 10-fold protocol with 26,961 clips and (ii) the primary official participant-disjoint split with 17,796 training clips and 4,137 test clips. The absolute-delta configuration augmented with standard deviations achieved the best official-split result with SMO: 86.42% accuracy, 0.865 weighted F1, and 0.799 multiclass Matthews correlation coefficient (MCC). Relative to the absolute-delta mean baseline, the gains were 4.60 per-centage points, 0.045, and 0.067, respectively. Standard-deviation features were consistently strongest across both classifiers and protocols. Signed delta means alone lost information through temporal cancellation, whereas range-based variants were less effective. Most residual errors occurred between snoring and breathing. The study establishes a reproducible, lightweight benchmark for future audio models while making no claim of clinical diagnosis.