A knowledge graph-guided 3D visual question answering method was proposed to capture the implicit common-sense semantic relationships between objects in the scene. By introducing external structured knowledge, the model was effectively enhanced in both semantic understanding and reasoning ability. Specifically, key semantic entities were extracted from the question text, and a knowledge graph-guided feature enhancement module was designed to obtain knowledge features using these key semantic entities. The knowledge features were fused with visual features extracted both from the question representation and from the 3D object detection network for answer prediction. Experimental results on the ScanQA dataset showed that the proposed method outperforms existing baseline models on metrics such as EM@1 and BLEU-4.

To address the issue of inconsistent thermal stress between the upper and lower devices of the same bridge arm in converters under conditions of uneven aging or layout deviations, a novel thermal balancing method for upper and lower bridge arms based on dead-time shifting adjustment was proposed. By monitoring the on-state voltage of insulated gate bipolar transistor (IGBT) online, the real-time junction temperature of IGBTs was deduced based on the mapping relationship between the on-state voltage and junction temperature of IGBTs. Then, the temperature difference between the upper and lower devices of the same bridge arm was determined. When the thermal stress difference between the upper and lower bridge arms exceeded the threshold, the effective conduction time of the upper and lower bridge arms was adjusted by fine-tuning the position of the dead time, thereby actively regulating the power loss distribution of the upper and lower bridge arms. This achieved an increase in the loss of the bridge arm with lower temperature and a decrease in the loss of the bridge arm with higher temperature without affecting the total system loss, thus improving the thermal stress consistency of power devices in the converter and enhancing system reliability. Theoretical analysis and experimental results showed that the proposed method did not affect system efficiency and had no significant negative impact on output performance. Under typical operating conditions, the average temperature difference between the upper and lower devices was reduced by approximately 14% after adjustment, effectively improving the thermal stress consistency of the upper and lower bridge arms.

A railway foreign object detection algorithm for complex weather conditions based on an improved RT-DETR, named FRP-DETR, was proposed. A feature complementary mapping module (FCM) and Pzconv units were introduced, which built complementary paths between shallow spatial details and deep semantic information to compensate for the limitations of single-scale feature representation, with the two components working collaboratively to enhance the perception capabilities of small and edge targets. A railway perception modulation fusion module (RMFM) was introduced, which was based on adaptive channel attention and spatial modulation mechanisms to enhance the model’s response to key semantic information in railway scenes. The original downsampling module was replaced with pinwheel-shaped convolution (PSConv), which enhanced edge texture extraction through multi-directional asymmetric padding and separable convolution. A railway foreign object detection dataset containing four weather conditions was constructed based on the SaMam style transfer method, by transferring sunny railway foreign object images to rainy, foggy, and snowy scenes. Experimental results showed that compared to the original RT-DETR-R18 model, this method achieved improvements of 1.65 percentage points and 3.4 percentage points in mAP@0.5 and mAP@0.5:0.95, respectively, with a 63.5% reduction in parameter count and an inference speed of 88 frames per second. The results verified that FRP-DETR achieved high accuracy, lightweight design, and real-time performance in railway foreign object detection under complex weather conditions.

Significant progress in multimodal large language models (MLLMs) has driven advances in visual question answering, visual understanding, and reasoning tasks, and their potential for deployment on resource-constrained edge devices is increasingly recognized. However, large model sizes and the substantial costs of deployment and inference remain major barriers to practical adoption. Optimizing MLLMs for edge devices has become a critical research direction in this field. A comprehensive survey of recent advances in optimizing MLLMs for edge deployment was presented, along with the associated challenges and development trends. The research evolution of MLLMs on edge devices was reviewed, with particular emphasis on model architecture optimization and inference scheduling strategies. In model architecture optimization, techniques including visual information compression, sparse attention, and mixture-of-experts models were specifically analyzed. System-level optimizations involving computation scheduling, hardware adaptation, compilation optimization, and cloud-edge collaboration were investigated to enhance inference efficiency and energy efficiency. Furthermore, the key challenges of these models in practical applications were discussed, and a variety of task scenarios ranging from assistive to collaborative and autonomous types were covered, categorized by the perspective of autonomy levels. Finally, current limitations were summarized and future research directions regarding standardized deployment, efficient computing and storage, and multi-modal fusion optimization were outlined.

Current industrial image anomaly detection methods generally face challenges such as high dependency on anomalous samples, insufficient realism of synthetic samples, and limited perception capability for complex defects. To An industrial image anomaly detection method based on diffusion synthesis and feature mining was proposed to address these issues. Accordingly, a category-sensitive selective diffusion anomaly synthesis module was designed to generate pseudo-anomaly samples with adjustable intensity and category adaptability through controllable perturbation and category-sensitive loss, which could effectively alleviate data scarcity. Meanwhile, a multi-stage feature mining framework was constructed, including contrast-driven feature selection, multi-dimensional perception attention reconstruction, and residual refinement selection modules, enabling dynamic screening of anomaly-sensitive features and enhancement of structural details. Experimental results demonstrated that the proposed method achieved outstanding performance on the MVTec AD and MPDD datasets, with image-level AUROC scores of 99.7% and 98.4%, and pixel-level AUROC scores of 99.0% and 98.7%, respectively, validating its effectiveness and robustness.

Discrete element method (DEM) and Tavares breakage model were employed to investigate the rock fragmentation mechanism and the statistical characteristics of fragments under impact loading. Drop hammer impact crushing process of limestone specimens with different aspect ratios (1.0—2.0) and sizes (diameter 30—50 mm) was systematically simulated. Image J image processing technology was used to quantitatively analyze the fragment count, particle size distribution, and shape indices. The results showed that increasing the aspect ratio and the size significantly raised the crushing energy threshold, which led to a delay in the peak normal force, a reduction in its amplitude, and a slowdown of the overall fragmentation process. The equivalent particle size of fragments showed poor alignment with the sieve analysis curve, which indicated that the fragments commonly exhibited irregular geometric configurations such as flaky and rod-like shapes. Fragments from specimens with lower aspect ratios (1.0) and smaller sizes (diameter 30 mm) had lower roundness, more pronounced angularity, and greater surface roughness. A quantitative relationship model among geometric parameters–crushing response–fragment morphology can provide theoretical basis and regulatory pathways for the design of impact-resistant structures, the optimization of blasting parameters, and the control of particle morphology in rockfill grading.

Autonomous driving simulation platforms play a vital role in the development, testing and validation of autonomous driving systems. A systematic review of the classification and key technical pathways of mainstream simulation platforms was presented, covering aspects such as environment modeling, sensor simulation, vehicle dynamics modeling, perception algorithm evaluation, V2X communication and cloud-based simulation. Core challenges and research progress related to synthetic data generation, cost-effective algorithm training, cross-domain generalization and platform scalability were analyzed, emphasizing simulation technologies based on computer vision and artificial intelligence. Future development trends were discussed. Simulation platforms are evolving toward higher realism, interactivity and closed-loop validation with the advancement of emerging technologies such as generative AI, neural rendering and multimodal learning, gradually forming a comprehensive pipeline that integrates data generation, algorithm training and performance evaluation. Simulation platforms will continue to play an essential role in enhancing the generalization capability of autonomous driving system, accelerating product deployment, and establishing standard testing and validation framework in the future.

A lightweight detector SCGI-DETR was proposed based on an enhanced RT-DETR in order to address challenges in grape leaf disease detection—complex background, missed detection of small target, and resource-constrained deployment. The efficient StarNet backbone was employed to reduce parameter count and computational cost, enabling lightweight deployment. A feature pyramid CGSFR-FPN was designed. Spatial feature reconstruction was combined with multi-scale feature fusion in order to strengthen global context modeling and improve localization of multi-scale lesions in cluttered scenes. The Inner-PowerIoU v2 loss was constructed, which integrated global convergence acceleration and local region alignment in order to speed up bounding-box regression and enhance small-object detection performance. SCGI-DETR attained 91.6% precision, 89.8% recall and 93.4% mAP@0.5 on a grape leaf disease dataset, which improved 2.6, 2.4 and 2.3 percentage points over the baseline, and reduced parameters and computation by 46.2% and 64%, respectively. Results demonstrate that the improved algorithm achieves lightweight implementation while delivering superior detection performance, meeting deployment requirements for mobile and embedded devices.

Two key technical improvements were proposed to address the issue of decreased target detection accuracy caused by target occlusion and fog interference in intelligent autonomous driving scenarios. Firstly, an improved detection method was proposed for the problem of target occlusion. A lightweight MobileNetV3_small with an integrated enhanced attention mechanism was used as the SSD backbone feature extraction network. The detection accuracy in occlusion scenarios was enhanced by combining a multi-scale feature fusion mechanism and an adaptive hyperparameter Soft-NMS algorithm. The confidence loss function was reconstructed by improving the adaptive Focal Loss to alleviate the imbalance between positive and negative samples and the sensitivity to noisy labels. Secondly, an improved lightweight AOD-Net defogging method was proposed to address the problem of target blurriness in foggy images. The image contrast was effectively enhanced and texture details were improved, by constructing a multi-scale feature extraction network based on depthwise separable convolution, optimizing the cross-layer connection structure, and introducing a boundary enhancement module. Finally, an end-to-end collaborative optimization of the defogging network and the detection network was achieved through a joint loss function, providing a more reliable optimization path for the detection task of occluded targets in foggy images. Experimental results showed that the joint optimization model improved the target detection performance in foggy occlusion scenarios, achieving efficient detection with an accuracy of 93.85% and a detection speed of 47.61 frames per second, and demonstrating excellent model robustness.

A detection model named dynamic kernel perception YOLO (DKP-YOLO) was constructed to address the problems of contextual information loss, multi-scale distress coexistence, and insufficient detail restoration in pavement distress detection from a UAV perspective. A lightweight crack context module was proposed to mitigate the variability of crack morphology and the loss of contextual information in complex scenes. This module enhanced feature discriminability through heterogeneous parallel depthwise separable convolutions and a dynamic feature fusion mechanism. In order to tackle the challenges of complex backgrounds and multi-scale distress, a pavement distress perception module was designed, which improved the multi-scale distress perception capability using parallel multi-scale convolutions and a dual-attention mechanism. A large-kernel perception feature fusion module was introduced to overcome the limitations of traditional convolutions in modeling long-range dependencies and responding to small targets. This module captured global contextual relationships by integrating extra-large receptive field convolution with channel-spatial attention. Finally, a lightweight dynamic upsampling operator was incorporated into the neck network to reduce detail loss and semantic ambiguity during upsampling. Experiments on the UAV-PDD2023 dataset showed that the proposed model significantly improved detection accuracy while effectively reducing model complexity. Compared with the baseline model, DKP-YOLO achieved a better balance between being lightweight and maintaining high detection performance, making it more suitable for UAV-based pavement distress detection.

A heterogeneous smoke and flame visual detection method integrating visible and infrared images was proposed, in order to address the problems of low detection efficiency, poor accuracy, and consequent delays in fire disposal and misallocation of resources in the early stage of tunnel fires caused by lighting interference and structural environment. For infrared images, an OTSU threshold segmentation optimized by a grey wolf optimizer fused with differential evolution was applied, and the maximum entropy algorithm was combined to improve flame region segmentation accuracy. Considering the tunnel scene characteristics, YOLO11 was adapted, and the backbone network was optimized using a Swin-Transformer modified with a neural attention mechanism to enhance weak feature extraction. During feature fusion, a lightweight BiFPN was employed to adapt to multi-scale smoke and flame features, and an adaptive WIoU loss function was introduced to improve localization accuracy under fuzzy boundaries. At the decision layer, weighted fusion and conflict resolution were used to achieve dual-modal collaborative decision-making. Experimental results demonstrated that the proposed method achieved at least a 23.55% reduction in average localization error compared with single-modal detection. Compared with nine fusion algorithms such as evidence theory, the recall, F1, and mAP values of the proposed method reached 92.52%, 90.83%, and 93.47%, respectively, all exceeding those of the compared algorithms, while maintaining the smallest localization error fluctuation. The inference speed of the model increased by an average of 14.61%. The research results not only contribute to tunnel fire prevention and control but also effectively improve overall tunnel safety management and emergency response capability.

A small object detection algorithm FMCM-YOLO based on feature enhancement and fusion attention mechanism was proposed, aiming at the challenges of limited feature extraction, foreground-background confusion, and severe missed and false detections in small object detection in optical remote sensing images. Firstly, a four-head detection model was designed and a small target detection layer was added to detect numerous small objects in optical remote sensing images. Secondly, a feature enhancement module was proposed in the backbone network, which improved feature extraction capability by designing a multi-branch convolutional structure and introducing dilated convolution of different sizes. Thirdly, channel and spatial attention mechanisms were incorporated into the neck network, and a residual structure was introduced to focus on small objects, facilitating the distinction between targets and backgrounds. Finally, MPDIoU was adopted as the model’s loss function to accelerate convergence and enhance detection performance for small objects. Experimental results demonstrated that the mAP50 of the proposed algorithm on the two public datasets, USOD and AI-TOD, reached 89.9% and 60.6% respectively, which were 2.8 and 5.9 percentage points higher than those of the baseline algorithm YOLOv5m. Especially, the mean average precision for extremely tiny, tiny, and small objects increased by 2.1, 6.5, and 5.1 percentage points, respectively. These results proved that the FMCM-YOLO algorithm effectively improved the detection performance of small targets in optical remote sensing images.

A collaborative control approach integrating multi-agent proximal policy optimization (MAPPO) and maximum pressure control (MPC) was developed to address traffic control challenges at mixed traffic intersections involving connected autonomous vehicles (CAVs), human-driven vehicles (HDVs), and pedestrians. A hierarchical state space and an action space were designed through the formulation of a decentralized partially observable Markov decision process (Dec-POMDP). Reward functions considering the balance among safety, traffic efficiency, phase switching frequency, and regulatory compliance were introduced. The proposed model was validated under low, medium, and high traffic flow conditions using a centralized training with decentralized execution framework on the SUMO simulation platform. Experimental results demonstrated that the proposed MAPPO-MPC method significantly improved the throughput (by up to 44.3%) while reducing the queue length (by up to 49.3%), average vehicle delay (by up to 43.2%), and pedestrian waiting time (by up to 43.53%). Moreover, the model exhibited more substantial performance advantages when the CAV penetration rate exceeded 50%, outperforming the traditional Webster-based methods and the baseline models.

A lightweight micro-expression recognition method based on optical flow and convolutional vision Transformer was proposed to solve the problems of short duration, low motion intensity and insufficient sample size of micro-expressions. The optical flow and optical strain of human faces between the onset frame and the apex frame were extracted to highlight the movement of facial muscles, thereby effectively reducing the texture interference and lowering the feature dimension. The adversarial domain adaptation method based on identity domain was adopted to further remove the irrelevant components in the micro-expression features by making full use of the subjects’ labels. A lightweight multi-stage CNN-Transformer hybrid model named MiER-CvT, including the convolutional embedding layer, the convolutional Transformer block and the SeqSoftmax layer, was constructed to enhance the model’s capabilities of local representation and information integration for micro-expressions. The experimental results showed that the proposed method achieved a UF1 score of 0.9171 and a UAR score of 0.9192 on the MEGC 2019 dataset, and the parameter number and computational complexity of MiER-CvT were 7.5 M and 0.1 G, respectively. Compared with the existing methods, such as MiMaNet, the proposed method has the advantages of high precision and light weight.

An improved lightweight small object detection method based on YOLOv8s was proposed to address the challenges of complex background interference, small object sizes, and limited device resources in unmanned aerial vehicle (UAV) image object detection. A high-resolution detection head was reconstructed to improve the sensitivity to tiny objects. A dynamic multi-scale aggregation network was designed to adaptively adjust the receptive field for scale-variant targets, while an efficient multi-scale feature deep fusion module was proposed to integrate features from different semantic levels. An efficient channel-spatial attention module was introduced to further refine fused features. Experiments on the VisDrone2021 dataset demonstrated that the proposed method achieved improvements of over 5.0% in average precision, precision at different intersection over union thresholds, and detection precision for small and medium-sized objects, with a 16.25% reduction in parameters. Tests on the DOTA dataset validated the generalization ability of the proposed model. Results demonstrate that the proposed method improves the detection performance while achieving the model light weighting, indicating its practical value for object detection in UAV aerial images.

A novel detection framework integrating automatic fluid target annotation and a lightweight YOLOv8n model was proposed, aiming at addressing the challenges of low annotation efficiency and insufficient model adaptability in real-time debris flow monitoring. An open-source debris flow dataset of 8 064 images was built. This dataset was achieved using multi-scale dynamic capture frames, which leverage fluid motion continuity, along with adaptive frame extraction and an improved binary classification model. To enhance the extraction of dynamic debris flow features and improve the irregular boundary localization, the LGSA module and C2f_GhostNetV2 structure were developed. The Shape-IoU loss function was also introduced into the model architecture. Experimental results demonstrated that the proposed YOLOv8-Mudslide model achieved an mAP@0.5 of 86.7%, which was 4.7% higher than that of the baseline model, and the detection speed reached 230.89 frame/s. This case provides reliable technical support for the real-time monitoring of debris flow disasters, and its framework can be further extended to other intelligent detection fields of fluid targets.

A multi-stage surrogate model named LSTM-LFM-SRU was developed to address the high computational cost and limited real-time capability of traditional hydrodynamic models. The urban flood process was divided into two stages: one-dimensional drainage through pipe networks and two-dimensional surface inundation. Long short-term memory (LSTM) networks and a combination of a low-fidelity physical model (LFM) with a super-resolution U-shaped convolutional neural network (SRU) were adopted to replace the corresponding stages. A flood prediction framework was constructed by combining data-driven methods with physical information. The multi-stage surrogate model was applied to a case study in the Xiasha area of Hangzhou. A 20-hour intense rainfall-induced inundation event was simulated within 5.31, with its computational efficiency reaching 12.7 times that of the high-precision hydrodynamic models. The one-dimensional LSTM-based surrogate achieved an average Nash-Sutcliffe Efficiency value of 0.94, while the two-dimensional LFM-SRU model reduced the Root Mean Square Error by 17.2% compared with the LFM-U-net model.

To address the core issue of efficiency degradation in concentrator photovoltaics (CPV) caused by high heat flux, microchannel cooling technology has emerged as a critical research direction. Advances in this field were systematically reviewed, with a focus on three mainstream configurations: single-layer, manifold, and jet-impingement structures. The effects of factors such as channel geometry, working fluid, and flow regime on pressure drop, temperature rise, and cooling efficiency were analyzed. Furthermore, the strengths and weaknesses of various flow channel designs, including serpentine, pin-fin, and fractal configurations, were compared. Studies have shown that single-layer straight channels are prone to high pressure drop and temperature non-uniformity. Derived serpentine and fractal channel designs can enhance heat transfer by perturbing the flow. The manifold configuration effectively reduces flow resistance and improves temperature uniformity, often serving as a foundational platform for integrating other cooling technologies. The hybrid jet-impingement and microchannel approach demonstrates potential in addressing localized ultra-high heat flux, though its system complexity and energy consumption still require optimization. Looking ahead, future research on this technology should progressively advance from the millimeter scale to the micro-scale, enabling the development of lightweight and efficient portable concentrator photovoltaic devices. The focus should shift from pursuing “cooling intensity” to emphasizing “intelligent heat transfer architecture,” deeply integrating bionics and artificial intelligence to design adaptive microchannels. Furthermore, the research perspective should transition from optimizing individual components toward a systemic, co-design approach that couples photonic, thermal, electrical, and mechanical multi-physics fields.

The development trajectory of coupled heat and moisture transfer (HAMT) models was elucidated, and the applicability and limitation of different formulation were compared in order to address the complexity of HAMT mechanism in building envelope, as well as the diversity of modeling approaches and experimental methods with markedly different levels of engineering applicability. Typical experimental methods and property measurement techniques were summarized, and their differences in reproducibility and controllability were analyzed. Engineering applications of HAMT were synthesized, with particular attention given to the specific challenges of the regions characterized by high humidity, abundant rainfall, and intermittent operation patterns. Existing research gaps were identified and future research directions were proposed in light of the ongoing enhancement of building envelope performance and recent advance in artificial intelligence methods, with the aim of providing systematic reference and methodological insight for subsequent study and the transition towards net-zero energy building.

An encirclement decision-making method based on adaptive curriculum learning and multi-agent proximal policy optimization (MAPPO) algorithm was proposed to address the problem of encirclement and game confrontation for unmanned surface vehicle (USV) swarms in complex marine environments. For maritime encirclement and defense tasks, a combat simulation environment including dynamic targets and multiple reefs was constructed, and the judgement criteria for successful encirclement and task termination were defined. To improve the decision-making efficiency in complex adversarial environments, a normalized state space representing the motion states of both friendly and hostile parties, multi-scale reward functions, and a continuous action space were designed. An adaptive curriculum scheduler was introduced into the centralized training and decentralized execution framework to dynamically adjust the environmental complexity and the noise intensity of action exploration. The results of multiple sets of adversarial simulation experiments indicate that, compared to the baseline methods (i.e., no curriculum learning and traditional curriculum learning methods), the proposed method effectively improves the encirclement success rate, shortens the average task completion time and the average encirclement path length, reduces the number of collisions, and exhibits good generalization ability and adversarial adaptability.