Search Results (674)

With the rapid development of the Industrial Internet of Things, data-driven deep learning has achieved remarkable success in bearing fault diagnosis. However, traditional static models suffer from catastrophic forgetting when facing continuously emerging fault categories and limited edge storage. Existing class-incremental learning frameworks expose critical limitations when applied to 1D vibration signals on micro edge devices, including feature space oscillation, difficulty in anchoring lightweight classifiers, and prototype drift over long incremental cycles. To address these challenges, this paper proposes a novel end-to-end class-incremental fault diagnosis method based on lightweighting and feature calibration tailored for severe memory-constrained conditions. Specifically, a lightweight feature extraction mechanism based on an $L_2$ constraint is introduced to replace computationally expensive similarity distillation, effectively suppressing feature space oscillations and providing stable spatial coordinates for old knowledge. Moreover, a mandatory balanced center–margin hybrid replay (CAHM) strategy is designed to balance class representation while proportionally retaining class center prototypes and marginal hard examples, balancing the anchor accuracy of the Nearest Class Mean (NCM) classifier and the discriminability of the decision boundary. Furthermore, an ultra-low-cost linear prototype calibration module is constructed using a learnable affine transformation to actively redirect shifted old class prototypes with negligible inference latency. Extensive long-tail incremental experiments on the CWRU bearing dataset demonstrate that the proposed method forms a highly synergistic anti-forgetting closed loop. Under an extremely limited memory budget ($K=40$), the proposed framework achieves an outstanding final average accuracy of 98.92% after five incremental stages, significantly outperforming mainstream baselines such as iCaRL, PRIL, and SCKD and exhibiting exceptional robustness for continuous online monitoring on industrial edge devices. Full article

The Multi-Demand Multidimensional Knapsack Problem (MDMKP) combines upper-bound capacities and lower-bound demands, producing a restrictive feasible region. We study three swap-based repair-intensification policies in a common hybrid framework: lightweight Hash tabu search, randomly sampled tabu search, and online learning-guided tabu search. The learning-guided policy uses a linear model updated during the run to score candidate swaps, while convergence, path diversity, quality–diversity maps, and PCA describe search behavior. On the 250M benchmark group, the complete ML policy attains a mean average gap of $0.76\%$, compared with $1.30\%$ for Sampled and $1.65\%$ for Hash, at higher computational cost. On the more constrained 100G group, its mean average gap is $1.23\%$, versus $13.84\%$ and $14.17\%$, respectively. Non-parametric tests support the quality differences. Because the logical tabu memories differ, these results compare complete policies and do not isolate the effect of learning. A 165-run sensitivity screen finds no significant single-factor dependence within the tested ranges; Hash exponents and the learning score weight are the most responsive factors. The behavioral experiments do not isolate a causal benefit of diversity; they show that high recorded diversity alone is insufficient and suggest that selective intensification can be more important than broad trajectory dispersion. Future work will examine equal-memory ablations, controlled diversity interventions, dynamic penalties, balanced positive/negative online updates, and broader benchmarks. Full article

Dynamic uplink power allocation is a critical challenge in cell-free massive MIMO (CF-mMIMO) networks, where distributed access points (APs) jointly serve multiple user equipment (UEs) under mobility, time-varying propagation conditions, and strong inter-user interference. Conventional optimization-based methods can improve fairness or spectral efficiency, but they often require repeated numerical solving and are usually designed for a specific objective. Learning-based approaches can reduce online decision time after training; however, their effectiveness depends strongly on the reward design and the selected operating objective. In response to these challenges, we propose a Deep Hybrid Intelligent (DHI) architecture designed to evaluate dynamic uplink power management within cell-free massive MIMO environments. The framework uses Soft Actor-Critic (SAC) learning to generate continuous uplink transmit-power decisions and evaluates objective-specific configurations for fairness, signal-to-interference-plus-noise ratio (SINR) improvement, and spectral-efficiency enhancement. In addition, three optimization-based strategies, namely max-min fairness, max-product SINR optimization, and max-sum-rate maximization, are incorporated to analyze the trade-off among fairness, signal quality, throughput, and computational cost. Limited-memory Broyden-Fletcher-Goldfarb-Shanno with bound constraints (L-BFGS-B) optimization is employed for the max-product and max-sum-rate objectives, while the max-min strategy is evaluated through a fairness-oriented feasibility procedure. Simulation results show that the fairness-oriented configuration achieves the highest Jain’s fairness index, reaching 0.989 at 120 access points, whereas the sum-rate-oriented configuration provides stronger SINR and user-rate performance. The results also indicate execution-time reductions of 51.6%, 83.7%, and 85.0% for the evaluated max-min, max-product, and max-sum-rate strategies, respectively, compared with conventional optimization-based implementations. These execution-time gains are accompanied by a clear performance trade-off: the max-min strategy provides the strongest fairness behavior, the max-sum-rate strategy improves total spectral efficiency and user-rate performance, and the max-product strategy offers a balanced operating point between collective SINR improvement and user-service balance. Therefore, the proposed framework does not optimize only computational speed, but also clarifies the trade-off among execution time, SINR, spectral efficiency, and fairness under dynamic uplink CF-mMIMO conditions. These results indicate that this architecture serves as an adaptable platform to evaluate dynamic uplink power distribution across CF-mMIMO networks. Full article

Recent context-aware streaming 3D reconstruction frameworks provide a promising solution for online vehicle perception by maintaining anchor references, local pose windows, and trajectory memory. However, directly applying such frameworks to intelligent rail vehicles remains challenging because rail transit scenes are dominated by long straight motion, low-parallax visual observations, repetitive trackside structures, weak textures, and illumination variations. These characteristics may cause redundant context accumulation, unstable frame registration, and gradual trajectory drift. To address this problem, this paper proposes a motion-aware geometric context adaptation method for streaming 3D reconstruction of intelligent rail vehicles in low-parallax scenes. Instead of requiring task-specific large-scale retraining, the proposed method adapts the inference-stage geometric context using scale-normalized visual motion cues, including scale-normalized translational displacement, turning tendency, and inter-frame viewpoint variation. A motion-aware keyframe selection strategy suppresses redundant low-parallax frames while preserving geometrically informative observations in curved or pose-changing segments. An adaptive local pose reference window further regulates recent visual context to improve frame registration consistency. Experiments on rail transit sequences and the Oxford Spires dataset show that the proposed method achieves lower trajectory error than LingBot-Map and VIPE, while reducing redundant keyframe storage and preserving the qualitative continuity of rail-related structures. The method provides a practical motion-aware streaming 3D perception solution for rail transit inspection and digital infrastructure management. Full article

Accurate estimation of the state of health (SOH) of lithium-ion batteries is crucial for battery management systems (BMS), as it directly affects system safety, operational reliability, and remaining useful life prediction. Traditional SOH estimation methods generally rely on complete charge–discharge cycle data, which limits their practicality in real-time and online applications. To address this limitation, this paper proposes a novel voltage-charge increment curve-based dual-branch bidirectional long short-term memory network (VCIC-DB-BiLSTM) for battery SOH estimation under incomplete data. First, non-uniformly sampled battery current-voltage data are processed into standardized sequences with equal voltage intervals via voltage-charge increment curves based on ampere-hour integration and cubic spline interpolation. Subsequently, sliding window segmentation is applied to extract fixed-length curve segments from continuous voltage intervals as input features, while the corresponding complete voltage interval curves are used as labels. Finally, the VCIC-DB-BiLSTM network is designed, which uses a dual-branch structure to integrate feature extraction from both patch-processed and raw data, combined with bidirectional sequential modeling. Experimental validation on four benchmark datasets, CALCE, Oxford, XJTU, and TJU, demonstrates that the proposed method achieves competitive performance in SOH estimation under incomplete discharge data conditions, confirming its effectiveness and practical applicability. Full article

Network intrusion detection systems (NIDSs) deployed in dynamic environments face concept drift from evolving attacks and traffic patterns, causing model reliability to degrade over time. Continual learning (CL) offers an adaptive solution, yet many methods misalign drift detection, memory updating, and optimization: drift is often judged with low-dimensional statistics, while adaptation occurs in representation space, limiting consistency under buffer constraints. To address concept drift in non-stationary network traffic and catastrophic forgetting during online intrusion detection updates, we propose a continual-learning framework built upon SSF that combines feature-space Gaussian-kernel Maximum Mean Discrepancy (MMD) drift detection with gradient-matching coresets for memory admission. The proposed framework retains strategic forgetting and steady-state distillation while replacing low-dimensional drift tests with feature-space MMD and mask-based selection with gradient-matching coresets, thereby improving incremental updates under a limited memory budget. On NSL-KDD and UNSW-NB15 under a unified multi-seed streaming protocol, the proposed method improves detection performance and knowledge retention. Experimental results demonstrate that gradient-matching coreset selection is the primary contributor to the observed performance improvements, while the effectiveness of MMD-based drift scheduling and strategic forgetting depends on the underlying data distribution and drift-trigger threshold. The proposed framework employs batch-level MMD scheduling to coordinate memory admission and online optimization, providing a practical path toward robust continual intrusion detection. Full article

Background/Objectives: The digital transition in school settings is reshaping children’s learning, writing practices, and mental health trajectories. This narrative review examines whether handwriting still matters in contemporary hybrid educational environments from a multidisciplinary perspective. Methods: Evidence from neuroscience, developmental psychology, educational sciences, pediatrics, and child psychiatry was narratively synthesized, with attention to handwriting, digital exposure, learning, emotional regulation, and vulnerable populations. Results: Handwriting uniquely integrates fine motor control, visuomotor coordination, orthographic processing, attention, and embodied cognition, supporting early literacy, memory consolidation, conceptual learning, and reflective writing. Conversely, excessive or poorly mediated digital exposure may interact with attentional fragmentation, sleep disruption, online stressors, problematic use, and internalizing symptoms, particularly in vulnerable children and adolescents. Digital tools remain essential for personalization, accessibility, and compensatory support, especially for students with neurodevelopmental or learning difficulties. Conclusions: Handwriting and digital technologies should not be framed as competing educational paradigms. A developmentally sensitive hybrid model is needed, preserving handwriting during key stages of literacy and self-regulation while integrating digital tools as purposeful, individualized resources for learning, inclusion, and school mental health promotion. Full article

Maritime unmanned aerial vehicle (UAV) missions such as ship inspection, search and rescue, environmental monitoring, and emergency response often involve multi-wave task releases, time-sensitive deadlines, constrained support vessel positions, and spatially heterogeneous risk. These factors couple task allocation with path planning and make fixed dispatching rules fragile under changing mission profiles. This study develops a hierarchical cooperative planning framework for multiple UAVs over a maritime risk field. A risk-cost A* layer generates feasible routes from support vessels to task points and estimates path length, risk exposure, and sortie duration. A rolling scheduler constructs feasible UAV task candidates, while a scenario-switching-aware LinUCB hyper-heuristic selects online among deadline-first, distance-first, risk-aware, and endurance-balancing rules. A forgetting-update, one-step look-ahead, scenario memory, and lightweight switching detection are used to improve adaptation to mission profile changes. Simulations on a 28 × 40 maritime grid with two support vessels, six UAVs, 40 tasks, and nine release waves show that the proposed framework achieves the highest average effective reward (370.18), the lowest average value regret (0.61), and a best reward ratio of 0.46 over 24 random scenarios. The results should be interpreted as evidence from an idealized simulation benchmark. The main benefit is improved reward robustness under non-stationary and high-risk profiles, rather than uniform gains across all metrics or direct field-deployment validation. Full article

In 3D integrated circuits, a placement decision that looks acceptable from a 2D wirelength view can still create a local thermal or stress problem after stacking. This issue becomes more visible as the number of tiers and the density of vertical interconnects increase. We propose ThermIC, a placement framework that brings thermal and mechanical risk estimates into the placement loop rather than treating them only as post-layout checks. The novelty of ThermIC does not lie in treating graph neural networks, reinforcement learning, uncertainty-aware learning, or physics-informed regularization as individually new techniques. Instead, ThermIC contributes a placement-time coupling mechanism in which physically typed graph propagation, dense multi-constraint risk prediction, and action-level reinforcement learning feedback are jointly organized for stacked 3D-IC placement. ThermIC uses a heterogeneous graph encoder to carry thermal, stress, timing, and congestion information through the netlist; a constraint head to estimate local hotspot, stress-risk, timing-violation, and congestion probabilities; and a sequential placement policy trained with physics-informed penalties. We evaluate the method on ThermIC-Bench, a simulated corpus with more than 30,000 finite-element samples from 18 heterogeneous 3D-IC designs with 4–8 tiers. Because the present study does not include proprietary industrial circuits, silicon measurements, or a tape-out case, the experimental results are interpreted as simulation-based benchmark evidence rather than final industrial qualification. ThermIC connects the heat-kernel branch to the discretized heat-conduction equation and the stress-filter branch to linear thermo-elastic equilibrium, providing a mechanism-level basis for physical interpretability. The analysis distinguishes offline simulation/training cost from online deployment cost and reports complexity, runtime, and memory scaling for practical large-scale use. Under joint DRC, thermo-mechanical stress, and thermally coupled timing checks, ThermIC obtains an $82.1\%$ physical verification pass rate. The peak-temperature error is $3.1$ °C, the hotspot localization IoU is 0.89, and the number of placement-closure iterations is reduced by $3.7\times$ relative to the heuristic baseline. Together, these benchmark results indicate that early, differentiable multi-physics feedback can make 3D placement less dependent on late correction cycles. Full article

Vibration-based sensing systems for deployed industrial fault diagnosis often face incremental fault classes, changing degradation stages, and limited permission to retain historical sensor streams. Static fault classifiers are therefore insufficient for online maintenance settings in which a model must learn new sensor-observed states while preserving previous diagnostic knowledge under a bounded memory budget. This paper proposes ASBR-CL, an adaptive stage-aware balanced replay framework for continual fault diagnosis under a fixed exemplar-memory budget. ASBR-CL combines dataset-adaptive exemplar memory, balanced replay between current data and retained exemplars, and a conditional validation checkpoint module that is enabled only when it improves balanced stage recognition. Experiments on SEU-enhanced92, the 92-dimensional feature construction for SEU, and XJTU-bearing23, the 23-dimensional feature construction for XJTU-SY, compare ASBR-CL with DGGN/MFF same-backbone continual-learning baselines and XJTU imbalance-aware variants under five random seeds and $K=100$ training exemplars. On SEU, the selected ASBR-CL setting reports $97.49\pm 1.18$ Average Accuracy, $90.37\pm 4.53$ Final Accuracy, $90.37\pm 4.53$ Macro Recall, and $11.50\pm 5.25$ Average Forgetting. On XJTU, the conservative ASBR-CL-BoundedVal-K100 setting is not a universal Final Accuracy winner: DGGN-ER reaches a slightly higher Final Accuracy ($89.52\pm 2.68$ versus $89.05\pm 2.48$). The ASBR-CL evidence instead lies in balanced recognition and retention, with Macro Recall $75.00\pm 2.00$, Macro-F1 $67.66\pm 3.83$, and Average Forgetting $20.39\pm 4.89$, compared with DGGN-ER at $63.94\pm 7.50$, $55.48\pm 12.82$, and $41.11\pm 13.61$. Additional validation-resource, RMS-derived stage-definition, and imbalance-aware baseline analyses show that the revised XJTU claim should be framed as more stable Macro Recall, Macro-F1, and forgetting control under memory-limited continual diagnosis, not as superiority on every accuracy metric. Full article

Low-altitude economy, fine-grained surveying, emergency response, and autonomous exploration are driving Unmanned Aerial Vehicles (UAVs) from passive data-acquisition platforms toward task-executing aerial agents. This capability transition requires UAVs to operate within complex, open 3D environments under six-degree-of-freedom (6DoF) motion, strict size, weight and power (SWaP) limits, partial observations, onboard computation constraints, and safety-critical action requirements. Therefore, the central scientific problem of UAV 3D scene understanding is how a UAV agent can construct, maintain, and use a spatiotemporally coherent and uncertainty-aware 3D scene state to support localization, planning, and safe interaction. Existing surveys mainly categorize the literature according to sensor types, application scenarios, or generic 3D representations, and thus provide limited analysis of how 3D scene understanding supports agent-capability evolution under embodied aerial constraints. To address this gap, we review UAV 3D scene understanding along an agent-capability evolution from offline interpretation to online understanding and predictive reasoning. This perspective highlights the underlying tensions between representation fidelity and onboard deployability, open-vocabulary semantic coverage and calibrated trustworthiness, post-flight static reconstruction and online scene-state maintenance, and predictive reasoning and safety-bounded decision support. Finally, we discuss open challenges in closed-loop data construction, trustworthy scene-state memory, collaborative fusion, sim-to-real transfer, and reliable onboard deployment. Full article

Accurately predicting the strain time series of orthotropic steel bridge decks (OSBDs) is highly challenging due to their strong stochasticity and nonlinear characteristics. This paper proposes a hybrid prediction framework integrating wavelet decomposition with a cascaded Convolutional Neural Network and Long Short-Term Memory architecture. Initially, the raw strain signals are decoupled into temperature-dominated low-frequency trends and vehicle-induced high-frequency dynamic components using the 6-level Daubechies 10 wavelet transform. Subsequently, a deep architecture comprising three CNN layers and two LSTM layers is constructed to precisely extract and learn the local spatial features and long-term temporal dependencies of the decoupled signals. Based on real-world monitoring data, the proposed model is comparatively evaluated against baseline models, including CNN-GRU, LSTM, and Gated Recurrent Unit (GRU), across three time horizons: 24 h, 1 h, and 10 min. The results demonstrate that the proposed method consistently exhibits superior predictive performance across multiple scales. Specifically, the mean absolute percentage error (MAPE) is strictly maintained below 0.6% across all tested horizons, with an R² reaching 0.961. Furthermore, the single-step inference latency is merely 0.63 milliseconds, which is significantly lower than conventional sensor acquisition intervals. This decouple-then-predict analytical framework effectively avoids the feature interference typically encountered when a single network directly processes complex mixed signals. Moreover, while strictly satisfying real-time computational constraints, it provides an undistorted, high-fidelity data foundation for future online fatigue evaluations and continuous state tracking of bridge structures. Full article

New Radio Vehicle-to-Everything (NR-V2X) safety scheduling is difficult because burst episodes increase urgent arrivals, lower transmission success, and create retransmissions that compete for future slots. A scheduler that waits for the visible queue can react late, whereas always reserving extra safety physical resource blocks (PRBs) consumes the best-effort (BE) capacity after the stress has passed. This study proposes Memory-Based Predictive Allocation (MPA), a finite-action PRB allocation rule for safety and BE coexistence. MPA combines the deadline queue and retry state with a decayed transient-deficit memory, online success calibration, and a recoverability-aware BE cost guard. At each slot, it tests feasible safety PRB increments and chooses the action that first limits urgent safety loss, then reduces next-slot carryover, and finally avoids unnecessary PRB use. The model uses an NR-V2X resource pool interpretation and a calibrated signal-to-interference-plus-noise-ratio (SINR)-to-success mapping with hybrid automatic repeat request (HARQ)-like combining. Monte Carlo results show that MPA lowers safety misses relative to queue-reactive scheduling while preserving more BE throughput than a maximum safety reservation. In dense non-line-of-sight (NLOS) stress, MPA keeps the 95th-percentile (p95) delivered packet delay within the three-millisecond budget and preserves 0.892 normalized BE throughput, versus 0.534 under fixed maximum reservation. Full article

Driver fatigue is a major cause of road accidents, and in-cabin monitoring is increasingly embedded into the Internet-of-Things (IoT) ecosystem of modern vehicles. Deploying such monitoring directly on microcontroller-class devices is challenging: models must fit tight memory and compute budgets, provide reliable confidence estimates, and adapt online to new drivers and conditions. We propose KAN-CLUE, an uncertainty-aware continual TinyML framework for driver fatigue detection from near-infrared periocular images at the IoT edge. KAN-CLUE combines a compact convolutional backbone with a Kolmogorov–Arnold Network (KAN) classification head that outputs Dirichlet-distributed class probabilities and a principled predictive uncertainty measure. A lightweight activation-histogram mechanism provides an additional out-of-distribution (OOD) score, and both signals drive an on-device continual learning scheme that selectively updates a small subset of parameters under a KAN-specific EWC-style regularization. On the ULg DROZY drowsiness database, the quantized KAN-CLUE model uses roughly 167k parameters (about 165 kB in Flash), requires on the order of ${10}^6$ MACs, and achieves around 3.1 ms latency on a Cortex-M–class microcontroller, while reaching 97.7% test accuracy with improved calibration and OOD detection compared with softmax-based TinyML baselines. Full article

Infrared small target detection (ISTD) remains challenging in complex cluttered environments because targets usually occupy only a few pixels and exhibit weak thermal radiation with limited texture information. The problem becomes more severe in high-resolution infrared imaging systems, where sliding-window inference is commonly adopted under memory and computational constraints. However, the truncated field of view may lead to contextual information loss and increased false alarms in cluttered regions. To address these issues, we propose the Selective Contextual Uplift and Attention Network (SCUA-Net). The proposed network adopts a U-Net++-style densely nested encoder–decoder architecture to enhance multi-scale feature interaction and preserve fine-grained weak-target features. In addition, a Global-Context Calibration Coordinate Attention (GCC-CA) module is introduced to inject window-level contextual statistics into coordinate attention, thereby improving clutter suppression and localization robustness under sliding-window inference. During training, a joint optimization strategy combining Online Hard Example Mining (OHEM) and Dice Loss is employed to alleviate severe foreground–background imbalance. During inference, Gaussian-weighted fusion is adopted to reduce stitching artifacts between adjacent windows. Experimental results on NUDT-SIRST and IRSTD-1k validate the effectiveness of the proposed method. SCUA-Net achieves 99.15% Pd, 0.558 × 10⁻⁶ Fa, and 0.9570 IoU on NUDT-SIRST, while maintaining competitive performance on IRSTD-1k at 161.6 FPS on an NVIDIA RTX 4090 platform, demonstrating favorable accuracy, robustness, and real-time performance in complex infrared scenarios. Full article