Feature fusion enables joint classification of surface electromyography (sEMG) and acceleration (ACC) signals, but classification supervision alone does not explicitly optimize statistical dependence between their learned representations. To address this gap, we propose a dual-branch gesture recognition method that combines supervised classification with mutual information (MI) regularization. A Mutual Information Neural Estimation (MINE)-based dependence regularizer encourages cross-modal dependence during training; the auxiliary projection and critic are removed at inference. The experiments included 11 participants performing 10 gestures under four postures across five repetitions, with within-subject testing and strict leave-one-subject-out evaluation. The method achieved an accuracy/Macro-F1 of 93.78%/93.74% within subjects, 37.02%/34.72% before cross-subject adaptation, and 85.61%/85.57% after adaptation using one target repetition of 40 gesture–posture trials (9 min 20 s of scheduled recording). Compared with the Basic sEMG–ACC fusion baseline using the same recognition architecture without MI regularization, accuracy increased by 2.52, 6.00, and 3.06 percentage points in the three settings, respectively; all matched comparisons remained significant after Holm correction. After one-repetition adaptation, MI also outperformed the evaluated local implementations of MSCNN-TL and an ACC-adapted Multistream CNN with fine-tuning. MI achieved higher mean performance than cross-temporal attention alone (CTA-only) and its combination with MI (CTA+MI), although the cross-subject differences were not significant after Holm correction. These results support MI regularization as an effective addition to classification supervision in the evaluated cross-subject sEMG–ACC framework. The current evaluation focuses on offline steady-action classification and does not cover continuous interaction scenarios such as gesture detection and transition handling.
NR-V2X sidelink broadcast is an important communication mode for delivering safety-related information to surrounding vehicles with low latency. However, in non-line-of-sight (NLOS) environments such as intersections, direct links are blocked, and the blind retransmission mechanism defined for NR-V2X provides limited reliability improvement. This paper extends our previously reported relay-assisted retransmission scheme, RRT + ACK, which combines relay selection using link-quality indicators aggregated per direction (AvgLQIs/Dir) with ACK-assisted transfer of retransmission responsibility to the selected relay. The extension is a dynamic retransmission control (RT-Ctrl) method based on the Channel Busy Ratio (CBR), intended to limit retransmission-induced congestion. Simulations using Scenargie in NLOS intersection environments, under the assumption of ideal error-free ACK delivery, show that the basic RRT + ACK scheme improves the evaluation packet delivery ratio within 200 m (EPDR200m) by approximately 8.6% compared with conventional blind retransmission. Under high vehicle density, the RT-Ctrl method with the best parameter setting improves EPDR200m from 0.783 to 0.867 and reduces the average CBR from 0.594 to 0.329 compared with RRT + ACK using a fixed retransmission count.
Low Earth orbit (LEO) maneuver analysis remains challenging because maneuver events are sparse and short in duration, while accurate analysis requires simultaneous maneuver detection, onset-time localization, and continuous three-dimensional velocity-increment (Δv) estimation from long multivariate tracking sequences. To address these challenges, an improved multi-task iSpikformer framework is proposed for the joint detection, onset-time localization, and three-dimensional velocity-increment estimation of low Earth orbit maneuvers from multivariate tracking time series. The model combines a local convolutional encoder, stacked spiking Transformer blocks, and task-specific output heads within a unified multi-task framework. Event-centered supervision, hard-negative optimization, and robust Δv regression are employed to improve sparse maneuver detection and continuous parameter estimation. For full-scene inference, predictions from overlapping windows are fused and converted into discrete maneuver events through boundary-aware event extraction and validation-based calibration. Unlike approaches that separately handle maneuver detection and parameter estimation or rely on conventional dense sequence modeling, the proposed framework jointly learns maneuver occurrence, onset location, and three-dimensional Δv from a shared temporal representation. Evaluation on a synthetic dataset of Starlink-like LEO trajectories showed that the proposed method achieved an F1-score of 0.9018, an onset-time MAE of 14.54 s, a component-wise Δv MAE of 0.0307 m/s, and a vector RMSE of 0.1302 m/s. Compared with representative baseline methods, the proposed method showed improved maneuver-detection performance and more accurate Δv estimation while maintaining comparable onset-localization accuracy. The inference-stride analysis further showed that inference time could be substantially reduced over a moderate stride range with limited changes in detection and estimation performance, whereas an excessively large stride reduced detection sensitivity. The proposed framework therefore provides a unified data-driven approach to event-level LEO maneuver analysis and demonstrates the applicability of spiking temporal modeling to joint maneuver detection and continuous orbital-parameter estimation.
Sensor failures reduce the spatial information available for greenhouse microclimate monitoring, but their practical consequences depend on whether agronomic risk decisions remain reliable. We evaluated the progressive loss of paired indoor temperature–relative humidity (T/RH) sensors in a Chinese solar greenhouse using whole-sensor holdout. Longitudinally and vertically balanced subsets and transparent estimators were frozen before a sealed final test of 28 dates, four calendar blocks and six stage-aware risks. Relative to a 20-input reference, a longitudinally and vertically balanced six-input network increased the temperature mean absolute error by 0.014 °C and the relative-humidity error by 0.027 percentage points; macro-recall changed by −0.001 and support-aware macro-F1 by −0.004, while median false-negative exposure remained unchanged. The results remained within the predefined limits across 26 agricultural-threshold variants. In an exploratory validation-only comparison, spatially concentrated six-input layouts had greater temperature error (+0.080 °C, 95% block-bootstrap interval +0.058 to +0.104 °C), relative-humidity error (+0.099 percentage points), and macro-F1 loss (−0.004) than longitudinally and vertically balanced layouts. Macro-F1 did not differ consistently between the longitudinally and vertically balanced layouts and the unconstrained random layouts. Longitudinal and vertical coverage therefore functions as a robustness constraint against concentrated input loss in this system, rather than evidence of universal six-sensor sufficiency or cross-greenhouse generalisation.
Fatigue during table tennis training can impair performance and increase the risk of injury. This study proposes a real-time fatigue recognition system that integrates an embedded edge-AI racket with a camera for synchronized multimodal data acquisition. On the racket side, a neural network was deployed on an edge-AI microcontroller unit (MCU) to enable on-device inference, while the test data were transmitted to a computer for multimodal feature fusion. Given the continuous nature of fatigue development, the data were divided into three stages: non-fatigue, transition, and fatigue to facilitate progressive evaluation. In the binary classification task, which distinguished between the two extreme states, the multimodal hybrid model achieved an accuracy of 99%. When the transition stage was incorporated into a three-class classification task, performance declined because of the ambiguous boundaries between fatigue states. However, the use of multimodal decision-fusion strategies, such as weighted soft voting, maintained an accuracy of 96%. Overall, the results demonstrate the feasibility of deep learning for real-time fatigue recognition during table tennis training and provide a practical monitoring tool to help players track their physiological states and reduce injury risk.
