User authentication plays a key role in user modelling within computing systems, particularly for both normal control (NC) and Alzheimer’s disease (AD) users. Conventional authentication methods are vulnerable to external attacks and often rely on memory, which poses challenges for AD users experiencing cognitive decline. Electroencephalography (EEG) signals offer an alternative due to their subject-specific and difficult-to-replicate properties. However, differences in EEG patterns between NC and AD populations require intelligent authentication approaches that can generalise across heterogeneous user groups. This study proposes Biometric using microstates of EEG (Bio-MEEG), an intelligent EEG authentication framework that integrates EEG microstate analysis, one-dimensional convolutional neural networks (1D-CNNs), and ensemble learning based on echo state networks (ESNs) for user authentication under a closed-set, repeated-sample identification setting. The model is evaluated on 1,015 samples from 203 participants across multiple countries and EEG configurations. Results show that Bio-MEEG achieves stable authentication performance with minimal accuracy fluctuation across demographic subgroups and remains robust under feature-space adversarial perturbations. By supporting both NC and AD users without significant performance disparity, the proposed framework contributes toward more accessible, intelligent, and reliable EEG-based authentication systems.

Precipitation nowcasting is essential for weather warning and rapid-response decision-making, yet existing deep spatiotemporal models often struggle to emphasize meteorologically salient echo structures and to keep their predictions coherent across future frames. We address these issues with ConCast, an enhanced SimVP for radar-based precipitation nowcasting that couples convolutional block attention module (CBAM) refinement with temporal consistency regularization (TCR). The CBAM stage strengthens channel-wise and spatial feature selection so that the network attends to informative precipitation patterns, whereas an auxiliary cosine consistency term constrains the relationship between successive predictions during training. On the Shanghai-Radar and SEVIR datasets, ConCast improves forecasting quality over representative baselines and produces sharper and temporally steadier echo fields. Ablation studies show that the two components act on different aspects of the problem, with attention sharpening spatial discrimination and the consistency term improving sequence coherence, and that their gains do not overlap. Because ConCast adds only marginal overhead to SimVP, attention refinement combined with TCR is a practical option for nowcasting.

Rapidly capturing moving targets is critical for maintaining border security and public safety. However, traditional swarm-based target acquisition models often rely on the navigator’s observations and focus on fixed-formation capture methods. When target positions are uncertain and moving rapidly, deadlocks may occur due to target loss. To address this issue, this paper formulates a sliding-window pose graph optimization framework combined with trajectory prediction to estimate and predict the three-dimensional trajectory of agile targets in real time. By adopting a virtual center-of-mass extension strategy to optimize non-uniform circular capture formations, compact target enclosure and obstacle avoidance are achieved. This paper further integrates field-of-view awareness with yaw coordination by formulating a multi-constraint optimization problem. The proposed approach focuses on target-state fusion, prediction, and cooperative capture, demonstrating strong performance in multiple unmanned aerial vehicles (UAVs) cooperative tracking, target visibility maintenance, obstacle avoidance planning, and collision avoidance in dense environments.

A two-stage apple-harvesting robot was analyzed to improve motion coordination and harvesting efficiency in dwarf and high-density orchards. Screw theory was used to formulate the forward and inverse kinematics and spatial Jacobians of the positioning and picking mechanisms, while the Lagrangian and Newton–Euler methods were used to establish their dynamic models. Three predefined joint motion-time allocation schemes were compared under the same cycle constraint, and the link-length-proportional scheme was selected for subsequent analysis and experiments because it produced comparatively smooth actuator-load responses. Numerical comparison with an ADAMS rigid-body model showed close agreement under matched assumptions. Payload analysis indicated that the translational joint was most affected by increased fruit load. Planar boundary-point tracking tests produced maximum errors of 1.7 and 1.9 mm in the Y and Z directions, respectively, while the maximum joint errors were 1.7° for revolute joints and 1.6 mm for the prismatic joint. Orchard experiments achieved 287 successful picks among 371 targets, corresponding to a pooled success rate of 77.36% (95% Wilson CI: 72.83%-81.32%), with a mean picking time of 6.94 s per fruit. These results demonstrate the feasibility of the proposed modeling and motion-parameter allocation framework under the tested orchard conditions.

Gaze estimation has applications such as visual attention analysis and human-computer interaction. However, performance under free head movement conditions can be further enhanced. This study aims to improve the accuracy and robustness of appearance-based gaze estimation without head fixation. We develop a gaze-tracking system using a consumer depth camera and a deep-learning-based gaze estimation model [Gaze-Point-Net (GPN)]. A customized YOLOv5-based detector is developed to simultaneously localize the eyes and mouth, providing reliable facial landmarks for gaze estimation. The detected regions and depth image were used to calculate a vector of location and posture. In GPN, a double-channel convolutional neural network with a Squeeze-and-Excitation module extracts features from binocular images, and these features are concatenated with the location-and-posture vector to complete the multimodal gaze-position prediction task. Our GPN presents competitive performance in four experiments: (1) The error distribution in the sequential point test ranges from 4 to 10 degrees; (2) In the random point test, the calibrated and filtered GPN achieved a pixel error of 185.19 ± 116.57 pixels and an angular error of 4.52° ± 2.87°, demonstrating superior performance over counterpart methods; (3) In the trajectory tracking test, approximately 81.91% of gaze points were located within a 200-pixel tolerance radius of the target trajectory; (4) In the browsing test, the generated gaze heat and trajectory maps showed satisfactory results. The developed eye-tracking system, integrating a depth camera and deep learning models, demonstrates competitive performance and strong potential for several eye-tracking applications.

A key challenge in Internet of Vehicles (IoV) environments is accurately predicting collision risks, given the dynamic nature of traffic systems and the variety of available data sources. Existing approaches often rely on structured IoV data, while visual contextual information is not fully incorporated into the prediction process. This limitation motivates a unified framework that integrates environmental perception with sensor-derived information to enable more accurate collision-risk prediction. In this study, we propose a two-stage context-aware meta-ensemble framework that combines image-based perception and IoV-based predictive analysis for collision-risk prediction. The first stage uses multiple pre-trained convolutional neural networks to obtain rich visual representations from vehicle images through feature-level fusion. The second stage employs a hybrid stacked ensemble architecture that combines the probabilistic outputs of heterogeneous machine learning models with IoV and image-derived features, which a meta-learner then processes for final collision-risk prediction. The proposed feature-fusion model performs best in the perception stage, with an accuracy of 0.97. In the IoV-based prediction stage, the proposed stacked ensemble architecture outperforms all baseline models, achieving an accuracy of 0.9847. Robustness analysis shows that the perception module tolerates blur and rotation but remains sensitive to Gaussian noise. Cross-stage integration incorporates image-derived features into the IoV feature space, providing complementary visual context for collision-risk prediction and demonstrating its potential for intelligent transportation and cyber-physical IoV applications.

Soft robots, characterized by compliance, safety, and environmental adaptability, have shown great potential in healthcare, intelligent manufacturing, and human-machine interaction. Actuation technologies determine motion generation, workspace, and response capability, while sensing technologies provide the information required for perception, feedback, and control. Hence, the development of soft robots toward embodied intelligence raises the demand for both high-performance actuators and accurate multimodal sensors to enable robust control in complex environments. This review summarizes the major actuation strategies for soft robots, highlighting their working principles, advantages, and challenges. On the sensing side, flexible sensors based on different sensing mechanisms are discussed, with emphasis on their working principles, sensing characteristics, advantages, and limitations. Subsequently, recent advances in actuation–sensing integration are reviewed, including strategies for proprioception, exteroception, and multimodal information fusion, which collectively enable precise control and intelligent interaction. Current challenges in soft robots are also critically discussed, and future perspectives are provided for deeply coupled actuation–sensing systems. This review provides a comprehensive perspective on the development of actuation, sensing, and their integration, and is expected to serve as a useful reference for the design of next-generation intelligent soft robotic systems.

Anomaly detection is an important research topic in the Industrial Internet of Things (IIoT). In recent years, deep learning has been exploited to analyze complex IIoT data and build anomaly detection models. Due to the lack of abnormal samples and the difficulty of labeling industrial data, unsupervised deep learning has become the mainstream technique for IIoT anomaly detection, with autoencoders being the most representative approaches. However, the existing autoencoder-based IIoT anomaly detection models predominantly focus on a single data modality, since existing data fusion frameworks are mostly designed for supervised learning tasks, while it is infeasible to simultaneously reconstruct heterogeneous data modalities in a unified autoencoder. To address this limitation, this paper focuses on IoT sensing data and network traffic data and proposes AD-FIT, a novel autoencoder framework for IIoT anomaly detection via the fusion of IoT sensing and network traffic data in an unsupervised manner. Specifically, it creates multiple local autoencoders with different architectures to fit the two data modalities, and then fuses their reconstruction errors through a global autoencoder. We conducted extensive experiments based on a public IIoT dataset. Experimental results show that AD-FIT achieves the best overall anomaly detection performance among the evaluated baseline methods, with F1-score improvements ranging from 7.1% to 38.6%.

As wind turbines evolve toward larger capacities, fleet-level clustering, and operation under complex conditions, fault mechanisms in key drivetrain components show multi-physics coupling and complex evolution, creating a major bottleneck in condition monitoring: models are often constructible but hard to generalize. Although deep learning is effective for end-to-end feature extraction, industrial data challenges - such as limited samples, long-tailed distributions, and domain shifts - severely restrict model generalization and engineering applicability. To address these issues, this paper systematically reviews fault diagnosis and intelligent operation and maintenance (O&M) technologies for wind turbine drivetrains. It analyzes generative-model-based sample augmentation methods and their limitations in authenticity and generalization improvement; examines the roles of semi-supervised, contrastive, and cost-sensitive learning in handling imbalanced and unlabeled data; and discusses how multimodal fusion, federated learning, and domain adaptation help alleviate data silos and environmental drift. Finally, it summarizes major challenges, including cross-domain generalization, explainability, and edge deployment, and proposes a future adaptive O&M paradigm integrating physical mechanisms with data-driven methods across the full asset lifecycle.