Current K-12 artificial intelligence (AI) education is progressively shifting toward regularized, project-based instruction. However, pedagogical practices still suffer from obscured learning paths, insufficient perception of cognitive processes, and a lack of high-order reflection. Meanwhile, the routine use of large language models (LLM) in basic education often tends to be outcome-oriented, providing direct knowledge outputs. To some extent, this may reduce students' opportunities for autonomous trial and error as well as their space for cognitive construction, thereby constraining the development of their self-regulated learning (SRL) abilities. As a core metacognitive factor that supports the solving of complex engineering problems, the healthy development of SRL abilities directly determines students' capacity for strategy transfer and higher-order thinking in uncertain technological contexts. To address these challenges, an LLM agent-driven support framework for self-regulated learning was proposed embedded within the infrastructure of a parallel teaching laboratory. Synthesizing the planning-performance-reflection cycle of SRL, the framework operationalized complex AI projects into structured, node-based task spaces monitored by multi-role agents. Under the artificial systems, computational experiments, and parallel execution (ACP) approach, the framework leveraged artificial systems to capture multimodal process logs, conducted computational experiments to infer students' strategic deviations and cognitive obstacles, and utilized parallel execution to deliver heuristic prompts and teacher-facing evidence. Finally, a design-oriented case study was presented grounded in the Intelligent Mobility AI Laboratory and the school-based "AI+" curriculum at Tianjin No. 2 High School, illustrating the integration paths across various experimental projects and establishing a systemic foundation for subsequent empirical studies driven by fine-grained process data.
Traditional occupational models were limited by a static perspective centered on the biological human. To overcome this, a "tangram"-inspired dynamic triadic collaboration framework was proposed that preconceived occupations as networks of digital humans, robots, and biological humans. By employing task-atomic decomposition and dynamic weighting quantification, the framework systematically unpacked how technological penetration reconstructed occupational task bundles. This approach breaks through the static "human task bundle" paradigm of occupational science and advances the discipline toward a human-machine collaboration modeling paradigm, which provides a scientific foundation for occupational forecasting, policy formulation, and educational restructuring in the artificial general intelligence era.
The unmanned aerial vehicle (UAV) path planning problem is highly challenging due to the high dimensionality of the search space and presence of complex operational constraints. To address these limitations, an enhanced differential evolution (DE) algorithm was proposed that integrates three synergistic strategies. First, a Poisson distribution based mechanism was introduced to estimate the mutation success rate in the next generation, enabling adaptive allocation of search resources toward more effective mutation strategies. Second, cosine similarity constraint was employed to guide the evolutionary direction of the population, thereby reducing directional divergence and improving convergence efficiency. Third, a staged adaptive parameter control scheme was designed, which enhances global exploration in the early phase and accelerates convergence in later phase while maintaining population diversity. By collaboratively integrating adaptive strategy selection, evolutionary direction guidance, and dynamic parameter adjustment, the proposed method achieves a balanced trade off between exploration and exploitation. Experimental results of the CEC2017 benchmark suite and multiple UAV path planning simulation scenarios constructed from real digital elevation model (DEM) data demonstrate that the proposed algorithm achieves competitive performance in convergence speed, stability, and solution quality.
The rise of artificial intelligence is not merely a consequence of advances in computing technology, but the result of a long-term convergence of philosophical reflection, mathematical logic, and cognitive science. From the perspective of knowledge origins, this paper systematically examines the intellectual foundations and disciplinary evolution of artificial intelligence. At the philosophical level, it traces early forms of intelligent activities to the Eastern tradition of "Fu, Bi, Xing" and the Western trivium of "Grammar, Logic, Rhetoric", and further interprets intelligence through ontology, parallel intelligence, and the triadic framework of Being, Becoming, and Believing. At the mathematical level, the paper reviews the role of the three foundational crises of mathematics in promoting formalization, computability, and complexity theory, and reconstructs a logical trajectory from Hilbert's Program, Gödel's Incompleteness Theorem, the Turing Machine, Cybernetics, and McCulloch-Pitts' Model of Artificial Neurons to Category Theory and the Langlands Program. At the level of cognitive science, it discusses the Lighthill Report, the AI Winter, Longuet-Higgins's naming of cognitive science, Miller's cognitive revolution, as well as the hexagonal framework of cognitive science, showing how cognitive science helped artificial intelligence move beyond symbolic limitations toward deep learning, large models, and embodied intelligence. The paper argues that artificial intelligence is a product of the sustained interaction among humanistic thought, formal rationality, and cognitive mechanisms. Reconstructing its knowledge origins provides a fundamental framework for understanding the boundaries, limitations, evolution, and future directions of intelligence.
To address the limitation that global path planning algorithms struggle with dynamic obstacle avoidance and local path planning algorithms were susceptible to local optima, an improved hybrid algorithm was proposed combining the ant colony optimization (ACO) and dynamic window approach (DWA) for indoor service robots. Firstly, aiming at the low search efficiency of the traditional ACO algorithm, the heuristic factors were optimized, and the search neighborhood was reconstructed based on the differential drive robot motion model to enhance the global searching efficiency. Secondly, a novel obstacle avoidance strategy and path polyline optimization method were adopted to solve the deficiency that the original ACO-generated paths fail to guarantee reliable safety distances around obstacles. Finally, the improved ACO algorithm was integrated with the dynamic window approach to achieve accurate and stable dynamic obstacle avoidance. Experimental results demonstrate that compared with the conventional ACO-DWA algorithm, the proposed hybrid algorithm enables service robots to travel safely in complex scenarios with random dynamic obstacles.
Efficient and accurate detection of foreign object intrusion on transmission lines based on UAV aerial images and deep learning algorithms enables rapid identification of potential safety hazards, thereby improving the lean operation and maintenance capability of power grids. The research progress of deep learning-based foreign object intrusion detection methods for transmission lines using UAV aerial images was reviewed. First, deep learning-based foreign object detection methods for UAV aerial images were systematically reviewed from four dimensions: Faster R-CNN, SSD, the YOLO series algorithms, and other derivative algorithms. Second, an improved model for UAV aerial image-based foreign object intrusion detection in transmission lines was elaborated. It illustrated the application strategies and performance optimization effects of core technologies in detection tasks, including attention mechanisms, multi-scale feature fusion, and lightweight networks. Subsequently, the DETR-based foreign object detection models built on the vanilla Transformer architecture were investigated, as well as large Transformer models tailored for the power industry and their engineering applications. Finally, in view of the current challenges in UAV aerial image-based intelligent foreign object detection for transmission lines, future research prospects were proposed.
In autonomous driving perception, accurate understanding of interaction signals is key to safe, human-like decisions. Real-time detection of these signals is difficult to achieve by current high-precision models. Effective modeling of dynamic temporal signals is generally lacking. However, effective modeling of dynamic temporal signals, such as turn signal flashes and pedestrian gestures. As a result, missed detections and false alarms occur frequently in complex scenarios, creating safety risks. To address this, a lightweight-temporal-knowledge distillation collaborative optimization model for autonomous driving edge devices is proposed in this paper, named Lightweight Temporal Knowledge-distilled YOLO (YOLO-TKD). The model is designed to resolve the conflict between the need for temporal modeling in interactive signal detection and the constraints of computational resources. The YOLOv8n baseline is adopted, with MobileNetV2 first utilized as the backbone network to compress the model size. Subsequently, a Temporal Convolutional Network (TCN) module is introduced to capture dynamic dependencies between consecutive frames. Finally, a Knowledge Distillation (KD) strategy is incorporated to maintain recognition performance. Experimental results on the dataset show that the proposed method achieves precision, recall, and mean average precision of 87.5%, 86.4%, and 84.3%, respectively, which represent improvements of 3.2%, 2.6%, and 1.1% over the baseline YOLOv8n. At the same time, the model parameter count is reduced by 26.3%. The proposed YOLO-TKD not only significantly reduces resource consumption but also effectively improves the accuracy and temporal robustness of interaction signal detection, providing a viable technical solution for autonomous driving edge devices.
The detection of strip mending in pavement diseases provides crucial technical support for intelligent road maintenance. Existing deep learning-based semantic segmentation models often suffer from issues such as large model parameters, slow inference speed, and loss of detailed features in strip mending detection tasks, making it difficult to balance accuracy and real-time performance. To address these challenges, this paper proposes RT PSPNet, a lightweight and high-precision real-time semantic segmentation model. Based on PSPNet, RT PSPNet incorporates multi-dimensional improvements: firstly, structured pruning is applied to the ResNet50 backbone in terms of depth and channel number, resulting in a shallow lightweight ResNet that significantly reduces parameters while preserving more detailed features; secondly, a lightweight Compact Pyramid Pooling Module(CPPM) module is designed to replace the original PPM module, which reduces pooling branches and introduces depth-wise convolutions to further enhance feature extraction efficiency; finally, a multi-scale feature fusion mechanism and a multi-head composite loss function are introduced to improve the model's ability to segment slender targets completely and enhance training stability. Experimental results on a self-built strip mending dataset show that RT PSPNet improves the segmentation accuracy (mIoU) by 7.06 percentage points compared to the original PSPNet, while reducing the parameter count by 96.4%, computational cost by 76.1%, and achieving an inference speed of 147 fps, and reaching 258 fps while using mix precision speeding strategy, which is 8.6 times that of the baseline model. Compared with many mainstream and lightweight state-of-the-art models, RT PSPNet demonstrates comprehensive advantages in balancing three aspects: accuracy, speed, and model light-weighting.
Brain language decoding provides a critical pathway for restoring communication in patients with aphasia and serves as an important driving force for the development of brain-computer interface(BCI)technology.In recent years,substantial progress has been made in multi-level brain language decoding based on both invasive and non-invasive brain signals.However,different linguistic levels vary markedly in brain signal representation,task complexity,and model se-lection,which has led to fragmented technical pathways and research emphases,while a systematic level-based synthesis is still lacking.Based on this,the major advances in brain language decoding research from 2008 to 2025 were reviewed across different linguistic levels,including phonemes,words,phrases,sentences,and discourse.The influence of brain signal modality,task setting,and model architecture on decoding accuracy was analyzed,and further key issues such as multi-level decoding,open-vocabulary decoding,and the integration of pre-trained language models were discussed.On this basis,several future research pathways were proposed for the development of language brain-computer interfaces and intelligent language rehabilitation,with the aim of providing theoretical support and technical reference for related inter-disciplinary studies.
To address the challenges faced in the analysis of information needs in time-sensitive task scenarios, an intelligent parsing method based on ontology is proposed. Firstly, an ontology model is designed to formally describe tasks and forces, Secondly, a semantic transformation and logical reasoning method based on the ontology model is established to automatically generate a structured list of information needs from unstructured task descriptions. The experimental results show that this method can generate a standardized and detailed list of information needs. Compared with traditional manual analysis, the efficiency of information assurance is significantly improved.
Artificial intelligence of things (AIoT) has become an important enabling technology for the chemical industry, especially in online monitoring and intelligent inspection, equipment management, and predictive maintenance, providing strong support for the transition toward Industry 5.0. From the perspective of AIoT system architecture, the recent progress of key technologies was reviewed, including data acquisition at the perception layer, data transmission at the network layer, data processing at the edge and platform layers, and control and actuation at the execution layer. Typical application scenarios in the chemical industry were further analyzed to summarize the deployment effectiveness and technical pathways of AIoT in real production environments. The analysis indicates that AIoT can significantly enhance state perception, anomaly identification, predictive analysis, and collaborative control under complex operating conditions. However, challenges remain in heterogeneous data management, model generalization and interpretability, secure communication, and large-scale engineering deployment. Future research should focus on cloud-edge-end collaborative intelligence, lightweight and adaptive deployment, and physics-data hybrid modeling for safety-oriented closed-loop optimization.
Fine-grained few-shot image classification aims to recognize subtle inter-class differences under limited annotated samples and has been widely applied in intelligent recognition, ecological monitoring, and autonomous driving. However, existing convolutional architectures are constrained by fixed receptive fields and local modeling schemes, resulting in insufficient characterization of multi-scale feature relationships. Although attention-based or frequency-domain methods have improved the discriminability of fine-grained features, limitations still exist in modeling cross-scale dependencies and feature fusion. To address these issues, a Mamba-wavelet-based multi-scale modeling method for few-shot fine-grained image classification was proposed. Specifically, a multi-scale feature relation network (MSFRNet) based on Mamba state space modeling was constructed. The proposed network consisted of two core modules, namely a wavelet-guided Dynamic Mamba multi-scale feature extraction (WDMFE) module and a cross-scale attention fusion (CAF) module. In the WDMFE module, a wavelet-guided dynamic adaptive Mamba structure was introduced to enhance frequency perception and contextual modeling across different scales. In the CAF module, multi-scale features were integrated through channel and spatial attention mechanisms to achieve cross-scale feature complementation. Experimental results on benchmark datasets, including CUB-200-2011, Stanford Dogs, and Stanford Cars, demonstrated that higher classification accuracy was achieved and stable performance improvements were obtained. It is concluded that the proposed network effectively enhances fine-grained feature representation and cross-task generalization ability, and provides a scalable framework for multi-scale modeling in few-shot fine-grained classification.
The Macao national observation and research station for coastal ecological environment (Macao field station) played a critical role in advancing research on coastal ecological environments and climate change. Its primary research focused include monitoring and modeling coastal environmental processes, investigating the mechanisms and environmental impacts of pollutant migration and transformation across multiple environmental media, developing theories and technologies for the prevention and control of compound environmental pollution, and exploring regulatory mechanisms and ecological restoration strategies for coastal environments under climate change. However, traditional environmental monitoring systems faced several persistent challenges, such as difficulties in acquiring and integrating large-scale multimodal data, the complexity of environmental process modeling, limitations in sensing equipment, and insufficient capacity for large-scale data processing and analytics. Furthermore, existing disaster early-warning and emergency management systems often struggle to respond effectively to rapidly evolving emergencies. To address these issues, a parallel-intelligence-driven framework for holistic intelligent observation and emergency management was proposed, grounded in the ACP (artificial societies, computational experiments, and parallel execution) method and large language models. The system constructed artificial models of ecological observation and emergency scenarios, evaluated response strategies through computational experiments, and enabled dynamic coordination of monitoring and emergency actions via parallel execution. By integrating key technologies such as parallel sensing, holistic perception, cloud-edge collaborative computing, coordinated control among biological humans, digital humans, and robots, and social radar, the proposed system established a comprehensive observation and emergency-management architecture that satisfied the requirements of security, safety, sustainability, sensitivity, serviceability, and intelligence (6S). This framework substantially enhanced the scientific rigor, foresight, and responsiveness of coastal ecological monitoring and disaster emergency management, providing robust technical support and scientific foundations for environmental governance and risk mitigation.
To address the challenges faced by kindergarten through twelfth grade (K-12) teaching laboratories in operational efficiency, safety assurance, and intelligent support, parallel intelligence was introduced, and a parallel teaching laboratory framework centered on the artificial systems, computational experiments, and parallel execution (ACP) method, while integrating the three-class human system. Through the construction of artificial systems, the execution of computational experiments, and the implementation of parallel execution, traditional teaching experiments were transformed into virtual-real integrated processes that were computable, analyzable, and verifiable. By systematically integrating next-generation AI technologies, including large language models and vision-language-action (VLA) models, autonomous operation and safety-controllable experimental teaching were supported. The architecture, key technologies, and operational mechanisms of the parallel teaching laboratory were analyzed, and the feasibility and application value of the framework in real educational settings were validated through the construction practice of the Smart Transportation AI Laboratory at Tianjin No. 2 High School. The results show that the parallel teaching laboratory provides a feasible pathway for promoting the evolution of K-12 experimental teaching from static facilities to intelligent systems.
As a model of urban park design worldwide, the Olympic Forest Park (OFP) serves as a paradigm of China's ecological civilization and a vital hub for urban well-being. To maximize the park's ecological value and address the growing public demand for healthy lifestyles, the "Parallel OFP" framework was proposed. Grounded in the ACP (artificial systems, computational experiments, and parallel execution) methodology, Parallel OFP utilized descriptive intelligence for digital modeling, predictive intelligence for ecosystem analysis, and prescriptive intelligence for operational guidance, greatly facilitating the smart management and value expansion of OFP. On this basis, we construct a metaverse architecture integrating scenario engineering, multi-agent modeling, and multi-modal large language models was constructed to serve as the operational backbone bridging virtual simulation and physical execution. The deployment of Parallel OFP was discussed in ecological monitoring, health promotion, and environmental education, highlighting its potential to transform traditional management paradigms. Ultimately, by fusing the artificial and physical realms, Parallel OFP aims to drive innovation in ecological conservation and promote healthier lifestyles, contributing to the sustainable development of urban parks.
The intelligent equipment system serves as the physical carrier for the transformation of operational forms toward cognitive adaptation. To maintain functional continuity within highly uncertain and strongly adversarial battlefield environments, the system must possess an endogenous "learnability". Learnability as the ontological capability of an intelligent equipment system was defined to cope with uncertainty and sustain functional continuity, constructing a three-layer theoretical framework that integrates ontology, operational mechanisms, and implementation paths. Firstly, the intrinsic logic of learnability becoming a rigid requirement for system survival was demonstrated in three dimensions, such as historical inevitability, technological reshaping, and complexity criticality. Secondly, the connotations of learnability at three hierarchical levels were analyzed, such as the intelligent node level established "robust autonomy" as the existence baseline, the intelligent system level achieved "emergent collaboration" as structural gain, and the intelligent system-of-systems level realized "cognitive evolution" for rule dominance. Furthermore, corresponding three-layer structured implementation paths were proposed, elaborating on the coupling mechanisms and closed-loop feedback mechanisms ranging from individual representation deduction and system structure self-organization to system-of-systems rule evolution. This framework aims to provide a systematic theoretical framework for the top-level design, effectiveness evaluation, and capability generation of intelligent equipment systems, offering theoretical exploration for understanding and constructing the "cognitive life forms" of the future battlefield.
Forestry knowledge is highly specialized and broad in scopeand is particularly sensitive to regulations and practical standards. To address the knowledge gaps, terminological ambiguities, and factual inaccuracies that general-purpose large language model (LLM) exhibit in forestry scenarios, an integrated framework of "data synthesis-model training-systematic evaluation" was proposed. Based on a general base model, domain-specific instruction fine-tuning was conducted to obtain LinYun, a domain-adapted model for forestry. Experimental results demonstrate that LinYun significantly outperforms general-purpose models of comparable scale in forestry-related tasks, and in some tasks approaches or even surpasses the performance of much larger models.
Urban transportation systems are rapidly evolving into CPSS (cyber-physical-social system), driven by the continuous integration of autonomous vehicles, unmanned aerial vehicles, and diverse intelligent agents. This evolution has dramatically increased system complexity, dynamics, and coupling, rendering traditional human-centric research paradigms insufficient for timely understanding and response to fast-evolving system behaviors. To address these challenges, an autonomous framework called "AI Transportation Scientist" was proposed to revolutionize transportation research through parallel intelligence. The architecture leveraged a synergy between large language model and multi-agent system across four functional layers (interaction, cognitive, experimental, and support). At its core, a dynamic routing engine adaptively scheduled intelligent agents to tackle mechanism discovery, strategy validation, and system optimization. By implementing a full-chain collaborative closed loop—encompassing problem identification, simulation, and feedback optimization—the framework enabled the autonomous discovery of transportation laws and the continuous evolution of control strategies. This research establishes a scalable technical paradigm for advancing transportation science within CPSS environments, ensuring both efficient problem-solving and innovative strategy iteration.
The randomness of individual behavior, the complexity of social relationships, and the lag in prevention and control strategies make the transmission process of infectious diseases highly dynamic and uncertain. This poses new challenges to the analysis of the evolution of infectious diseases. Existing research has made significant progress in disease transmission, behavioral evolution, and policy intervention through multi-layer network modeling and individual behavior analysis. However, there is still insufficient systematic modeling of cross-layer closed-loop mechanisms, which makes it difficult to fully explain the coupling relationship between individual behavior, social network structure, and prevention and control strategies. This paper describes the dynamic interaction process of “policy, behavior, dissemination, and feedback” as a whole, and proposes a computational model for the evolution of infectious diseases based on social networks and individual behavior, which includes three interconnected levels: individual level, organizational level, and social level. Finally, the effectiveness of this model was verified by analyzing and calculating the number of surviving agents of various types under different strategies and the final stable state achieved in the experiment.
Accurately understanding and predicting pedestrian crossing intent is crucial for ensuring the safety of autonomous vehicles. Existing approaches are often limited to visual motion cues such as pedestrian trajectories or body poses, while overlooking interactive signals like gestures and head orientations, making it difficult to capture key cues of pedestrian-vehicle interaction. To address these limitations, ARPCI (accurate reasoning for pedestrian crossing intent) was proposed, a multi-feature fusion framework designed for pedestrian intent inference. Specifically, a pedestrian feature module was developed that first focused on skeleton-based features to capture motion trends, and further leveraged MobileNet to extract head pose features. Combined with YOLOv8n for gesture recognition, pedestrian-vehicle interaction signals were captured more comprehensively by the model. In addition, a scene encoding module and a self-vehicle feature module were introduced to integrate contextual and ego-dynamic information, thereby enhancing adaptability to complex traffic environments and improving prediction accuracy. Extensive experiments on the widely used JAAD dataset show that the approach achieves an accuracy of 88%, surpassing several state-of-the-art counterparts. Moreover, the ablation studies provide further evidence of the effectiveness of the proposed input features.
