Model compression is widely used to deploy large neural networks on resource-constrained edge devices. Among existing techniques, low-rank composition is theoretically grounded in approximation theory and provides a strong basis for preserving model performance after compression. However, in practice, even state-of-the-art low-rank methods such as Tucker, Tensor Train, and Tensor Ring decomposition may still introduce substantial reconstruction errors. These errors are difficult to recover using conventional fine-tuning strategies, especially under high compression ratios. To address this issue, this paper proposes Progressive Knowledge Distillation-Incorporated Low-Rank Compression (PKD-LRC). The proposed framework introduces three key components. First, a residual-compensation mechanism is incorporated into convolutional compression to reduce information loss. Second, a dynamic temperature adjustment strategy is used to improve the fine-tuning of low-rank models and enhance prediction accuracy. Third, a progressive learning paradigm is developed to balance compression ratio and classification performance. Theoretical analysis is conducted to investigate the effectiveness of the residual compensation mechanism for compressing convolutional neural networks (CNNs). The experimental results on the CIFAR-100 dataset show that PKD-LRC achieves a 0.35% higher Top-1 accuracy with a 7.75 & times; compression ratio compared to the uncompressed VGG-16 model. On the ImageNet dataset, the proposed method yields a 0.20% Top-1 accuracy gain with a 2.78 & times; compression ratio over the baseline ResNet-18 model. Above results indicate that PKD-LRC effectively mitigates reconstruction errors and enhances the performance of compressed models through adaptive fine-tuning.
When subcritical fuel is injected into a supercritical environment, transcritical phase transition may occur, which serves as a fundamental problem in advanced combustion engines. In this work, gradient theory modeling of transient transcritical phase transition of hydrocarbon fuel droplets is performed under various ambient conditions. Results show that the evaporative mass transfer and the temporal evolution of liquid-gas interfacial stability jointly determines the phase transition pathway. Under supercritical ambient conditions, the rapid increase of interfacial temperature reduces the liquid-gas density contrast and weakens the thermodynamic non-uniformity across the interface. This leads to a sharp decrease in the grand-potential density difference and the tangential-pressure depression relative to the equilibrium pressure, thereby driving surface tension decay and the transition from a sharp liquid-gas interface to a diffuse mixing layer. The transient characteristics of the interface is largely affected by the ambient temperature, pressure, and fuel vapor mass fraction. Specifically, the ambient temperature accelerates heat and mass transfer, the ambient pressure enhances gas dissolution and promotes interfacial thickening, while the ambient fuel vapor is capable of suppressing subcritical evaporation yet promoting transcritical mixing at elevated temperature and pressure. Finally, a phase transition criterion is proposed based on the residual liquid-core mass ratio, showing good agreement with the phenomenological classification obtained from experimental results.
Hydrogen offers a credible route to zero-carbon combustion in internal combustion engines and can deliver high thermal efficiency, but its physical and combustion properties create persistent challenges in fuel delivery, storage, and NOx control. This work reports experiments on a boosted single-cylinder spark-ignition engine running on hydrogen direct injection (DI) as the primary fuel, with intake-port water injection used as a NOx mitigation strategy and further enriched with hydrogen-charged nanobubbles. The objectives are to quantify how far water injection can suppress NOx and to test whether the additional hydrogen carried by nanobubbles can recover, or even improve, the power and efficiency lost through water dilution. Bulk nanobubbles (50-500 nm) remain suspended in water for weeks because their negligible buoyancy and strongly negative zeta potential suppress rise and coalescence; despite established uses in agriculture, water treatment, and biomedicine, their application as a hydrogen carrier in a working engine has received very limited experimental attention. A bespoke generator was used to charge water with hydrogen nanobubbles, and their size, concentration, and zeta potential were measured by dynamic light scattering (Malvern Zetasizer Ultra); by optimising the generator, a maximum concentration of 5.12 & times; 10(11)nanobubbles per millilitre was achieved. Hydrogen nanobubble water injection was then compared directly with conventional water injection. Relative to conventional water, nanobubble water raised indicated thermal efficiency by about 4% and torque by about 5%. NOx fell substantially in both cases, with a slightly greater reduction for nanobubble water; this difference is attributed principally to the lower measured lambda of that campaign rather than to a distinct NOx-suppression mechanism. Three independent diagnostics - residual exhaust oxygen, wideband lambda and exhaust-gas temperature - shift consistently in a direction that is compatible with additional in-cylinder hydrogen release from the nanobubble suspension. It must be emphasised, however, that such a release was not measured directly. The present dataset establishes the performance and emissions outcome; the underlying mechanism remains a hypothesis that requires dedicated optical and pressure-decay diagnostics before it can be regarded as demonstrated.
Coupled with vision-based inspection techniques, unmanned aerial vehicles (UAVs) have been extensively applied to power transmission line inspection. UAV images typically cover a wide field of view, they often contain complex backgrounds in power transmission line inspection, which make the accurate detection and localization of small-size defects challenging, especially on resource-constrained devices. To tackle this issue, this paper proposes YOLO-TDH, a novel lightweight object detection framework derived from YOLOv8. The proposed YOLO-TDH incorporates three key components to improve the detection of small-size defects: (1) an Enhanced Feature Integration Module (EFIM), which strengthens multi-scale feature extraction and captures fine-grained details required to distinguish defective components from normal ones; (2) an Enhanced Feature Fusion Module (EFFM), which optimizes feature information flow and reduce the loss of critical details for small-size defects; and (3) a novel Transformer-based decoder, which models global context and alleviates ambiguity among overlapping components. The ShapeIoU loss is employed to improve the bounding box regression accuracy. Experiments on a dedicated transmission line dataset show that YOLO-TDH, Insulator dataset for targeted fault verification, and the public VisDrone2019 benchmark to validate generalization capability. The results demonstrate that YOLO-TDH consistently outperforms existing state-of-the-art lightweight methods in terms of Precision, Recall, mAP@0.5, and mAP@0.5:0.95. Results show that the proposed YOLO-TDH achieves a proper balance between diagnostic accuracy and computational efficiency, providing a robust solution for real-time, fine-grained health monitoring in resource-constrained scenarios.
Historical and ethnographic evidence suggests that people around the world use supernatural beliefs to explain misfortune. The cross-cultural prevalence of this pattern may arise from our expectations that norm violations lead to sanctioning. In cases where an outcome seems punitive yet lacks an identifiable cause, individuals, motivated to fill this explanatory gap, may rely on beliefs in supernatural punishment to make sense of what has happened. If true, supernatural explanations for misfortunes should be more common when a victim is believed to have not cooperated in the past (i.e., in cases where misfortunes might be unexplained but otherwise appear deserved). Using a vignette experiment across nine diverse populations (N = 714), including African hunter-gatherers and Melanesian horticulturalists, we tested if knowledge of past bad actions increased participants' endorsements of supernatural causes for misfortune. Across all populations, participants rated a misfortune as more likely to be supernaturally caused when the victim was known to have failed to reciprocate a favour or engage in generous actions towards others in the past. This was slightly increased when the behaviour was additionally labelled normative. Responses to open-ended free-listing questions provide additional evidence that when the victim was known to be uncooperative, participants were more likely to propose supernatural causes for misfortune as well as supernatural solutions. Additional normative labelling had only a small effect. Together, these findings suggest that when misfortunes lack clear causes, people's motivation to explain interacts with moral evaluations, leading them to invoke supernatural beliefs most commonly when misfortune appears deserved.
