This paper proposes an end-to-end “large model annotation—small model distillation—onboard inference” framework to address a critical engineering bottleneck in the field of anomaly detection for intelligent connected vehicles: the tension between the strong reasoning capabilities of large language models and the severe resource constraints of edge computing. Specifically, the LongCat-Flash-Lite model (68.5 billion total parameters, approximately 3 billion activated parameters) is used as the teacher model, which performs three-level semantic annotation (Level 0: normal driving, Level 1: suspicious behaviour, Level 2: anomalous behaviour) on a balanced sample set pre-filtered by a TabNet coarse classifier, using few-shot prompting. Subsequently, the annotations serve as supervised signals to efficiently fine-tune the lightweight Qwen3-1.7B student model via Low-Rank Adaptation (LoRA), thereby transferring the large model’s anomaly-discrimination knowledge to the compact student model. Ultimately, the distilled student model performs inference independently on the onboard side without invoking a cloud-based large model API. In simulation experiments based on real vehicle telemetry data (originally 82.82 million records, downsampled to 1 million records), the proposed method achieved the following key results: (1) In the TabNet coarse-filtering stage, the binary classification accuracy reached 98.62% with a macro-average F1 score of 97.15%; (2) The LongCat-Flash-Lite teacher model annotated 120,000 balanced samples with an average confidence of 0.9174, and the distribution of the three annotation levels was 26.3%, 39.2%, and 34.6%, respectively; (3) After LoRA distillation, the student model’s accuracy improved from 27.50% to 32.10% (an absolute improvement of 4.60 percentage points), the macro-average F1 score increased from 0.1624 to 0.2468 (a relative improvement of 52.0%), and precision and recall improved by 7.49 and 2.83 percentage points, respectively. These experimental results validate the feasibility of transferring knowledge from large models under low-computational-power constraints and provide a new technical pathway and engineering reference for recognising anomalies at the edge of intelligent connected vehicles.
With the popularity and research of 5G and 6G technologies, a new era of everything being connected is gradually being realized. As technology advances, there is a surge in computation-intensive and latency-sensitive tasks. Meanwhile, traditional ground-based base stations are difficult to recover quickly when facing complex terrain or when equipment is damaged, resulting in service disruption. Unmanned Aerial Vehicles (UAVs) have become an important tool for providing communication and computing services in disaster scenarios due to their high manoeuvrability, low cost and rapid deployment capabilities. Traditional cloud computing is difficult to meet, and the development of mobile edge computing (MEC) is limited by resource constraints and other issues. Unmanned Aerial Vehicles (UAVs) have become important mission relays or terminals due to their high mobility, low cost, and ability to respond quickly to communication needs. Therefore, the combination of UAVs with edge computing has significant research value. Research for UAV combined with MEC mainly focuses on UAV-assisted computing or UAV relaying. In this paper, we study the task offloading problem under the scenario of a UAV acting as a server and a relay at the same time. We design an offloading scenario in which the UAV and the ground base station collaborate, and the transmission between the terminal and the UAV adopts a probabilistic channel. We also propose a sub-timeslot task division method. We cleverly build the joint optimal power allocation, proportional allocation, and UAV flight trajectory into an optimization model that minimizes the energy consumption of the system. The problem is non-convex in nature, and a three-stage algorithm based on alternating iterations is proposed as our solution to transform the non-convex constraints into convex ones. Simulation results show that the proposed algorithm can effectively reduce system energy consumption under the influence of task volume and flight period compared with other solutions.
Deep learning has achieved significant success in image classification. However, training deep learning models solely based on datasets often leads to limited improvement in their performance. Self-distillation, as a novel technique in knowledge distillation, utilizes the model’s structure to construct multiple branches and improves performance through distillation. However, existing self-distillation methods often utilize deep-level features to guide shallow-level features, neglecting the value of shallow-level features. Moreover, most methods rely on branches with deeper layers as the source of knowledge, which limits the utilization of knowledge from branches with shallower layers. Based on this, we propose a knowledge fusion mutual self-distillation (KFMSD) framework. KFMSD constructs multiple branches at different depths based on the structure of the original network, encouraging mutual learning between the original network and each branch. The framework includes a knowledge fusion module (KFM) that can fuse features from deep and shallow layers, enabling effective knowledge transfer between them. We validate the proposed method on three datasets, CIFAR-10, CIFAR-100, and Tiny-ImageNet, and compare it with other knowledge distillation methods. Experimental results demonstrate that KFMSD can further enhance model performance.
Remote sensing images are characterized by high spatial resolution and complex scenes, featuring large variations in object scales and the presence of visually similar or easily confused classes, posing significant challenges for accurate semantic segmentation. To address these issues, this paper proposes a novel semantic segmentation network for remote sensing imagery that integrates a hybrid attention mechanism with multi-scale feature fusion, built upon the standard encoder-decoder architecture. Within the skip connections, a spatial-channel hybrid attention module is introduced to simultaneously highlight salient spatial regions and recalibrate channel-wise feature importance, thereby enhancing the fidelity of feature propagation. In the decoder, a multi-scale fusion module composed of convolution kernels with diverse receptive fields is designed to effectively aggregate contextual information across scales, facilitating more precise boundary recovery and class discrimination. Experimental results demonstrate that the proposed method achieves a 4.51% improvement in mIoU over the standard U-Net on the UAVid dataset, and a 3.34% gain on the Aerial Imagery dataset, validating its effectiveness in handling multi-scale and complex urban scenes.
This paper employs a random forest regression model to predict elevator traffic flow in hospital inpatient buildings, providing a scientific basis for optimizing the scheduling and management of hospital elevator systems. Through a comprehensive analysis of the traffic flow characteristics of elevators in these buildings and the factors influencing pedestrian flow, we have developed a prediction model that incorporates various elements, including time distribution characteristics, population composition, and spatial distribution. The simulation results demonstrate that the random forest regression model effectively captures the periodic changes and fluctuations in elevator traffic flow, achieving high prediction accuracy. Predict elevator traffic flow using a simulated dataset (500 samples), when compared to support vector machines (SVM) and back propagation (BP) neural networks, the model exhibits lower mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE) values, along with higher R² values, confirming its superior performance in predicting elevator traffic flow in tertiary hospital inpatient buildings.
This paper addresses the challenge of frequency deviation and economic charging coordination in multi-agent electric vehicle (EV) parking systems with AC microgrid characteristics and low-bandwidth communication. A distributed finite-time control algorithm is proposed, leveraging load prediction to coordinate inverters and optimally share EV charging rates under constraints such as state of charge and total power demand. Each agent requires only local and neighbor information, reducing communication overhead and enhancing scalability. Theoretical analysis proves global asymptotic stability, and RT-LAB semi-physical simulations confirm the algorithm’ s effectiveness in achieving both frequency synchronization and economic dispatch.
With the rapid development of artificial intelligence (AI), smart agriculture has become a key direction for agricultural transformation. However, in Taiwan, the adoption rate of smart farming management systems among agricultural practitioners remains relatively low. This study aims to identify the key factors influencing adoption intentions and to analyze the functional requirements of big data-driven smart farming applications (APPs). By extending the Technology Acceptance Model (TAM) [1], this research incorporates perceived privacy and perceived security as external variables [2]. A qualitative approach was adopted through semi-structured in-depth interviews with farmers, production and marketing group leaders, cooperative managers, and agricultural association representatives. Government open data were also utilized to supplement the APP design analysis. The findings reveal that perceived usefulness and perceived ease of use remain the primary determinants of adoption, while privacy and security concerns significantly affect users’ trust and their intention to adopt the system. Interviewees emphasized that an effective smart farming management system should include core functions such as farmland and crop management, pest and pesticide monitoring, inventory control, cost management, and real-time alerts. In addition, privacy and security dimensions should be given due consideration. Therefore, this study integrates privacy and security constructs into the proposed framework, thereby addressing an important research gap in smart agriculture adoption. Practically, the results provide empirical insights for system developers and policymakers to design user-centered smart farming applications and to promote agricultural digital transformation. Overall, this study not only enhances academic understanding of adoption mechanisms in smart agriculture but also offers valuable implications for policy formulation and system development.
Software-Defined Networking (SDN) decouples the data plane from the control plane, centralizing control logic and enabling programmability, which provides far greater flexibility and manageability than traditional networks. However, this architecture also makes the SDN controller a prime target for attacks that can bring down the entire network, rendering intrusion detection in SDN critically important. Further, imbalance in the relevant dataset will cause the model to learn features of the majority classes while neglecting those of the minority classes, which will severely compromise the detection rate of the intrusion detection model. To address this issue, this paper proposes a hybrid sampling strategy: applying random undersampling to majority-class samples while employing SMOTE-ENN oversampling for minority-class samples. This approach generates a more balanced dataset and enhances the detection rate for minority classes. To mitigate diverse threats in SDN, this work introduces the SA-BiGRU intrusion detection framework. The model employs Sequential Attention (SA) for dimensionality-aware feature selection to eliminate redundant information, followed by BiGRU for capturing critical temporal dynamics in network traffic. Extensive experiments on the CICIDS2017, NSL-KDD, and CICDDOS2019 datasets demonstrate that SA-BiGRU achieves superior detection accuracy and precision, reliably identifying a wide range of attack types and outperforming baseline methods.
Background: Artificial intelligence integration into regional innovation systems presents fundamental questions about policy-driven technological transitions. While AI’s general-purpose capabilities offer substantial potential, real-world impacts depend critically on local digital infrastructure, institutional capacity, and network connectivity. China’s National AI Innovation Pilot Zone program provides a large-scale natural experiment to examine how AI interventions reshape regional innovation architectures, yet causal evidence on systemic effects and underlying mechanisms remains limited. Methods: Using longitudinal data from 264 Chinese cities (2010-2023) and staggered difference-in-differences with two-way fixed effects, we estimate causal effects of AI policy on regional innovation ecosystem resilience. Regional innovation ecosystem resilience was quantified across five computational dimensions via entropy-weighted aggregation of 23 indicators. Mechanism analysis decomposes policy effects through human capital upgrading and financial resource reallocation pathways. Heterogeneity analysis examines moderating roles of internal mobility, external connectivity, and digital infrastructure. Results: The AI policy significantly increases ecosystem resilience. Contrary to direct technology diffusion models, effects operate through systematic restructuring of human capital and financial flows, transitioning regional economies from factor-driven to knowledge-intensive growth modes. Policy effectiveness exhibits strong network dependency: internal mobility determines reconfiguration speed, while external connectivity shapes performance ceilings. Digital infrastructure depth emerges as a critical conditioning variable. Implications: AI policy functions as network-mediated system architecture rather than isolated technological intervention. Effectiveness depends fundamentally on pre-existing connectivity topology across physical, digital, and institutional dimensions, challenging technology-push models that emphasize intervention intensity over structural context. Findings establish a computational framework for modeling policy-driven socio-technical transitions and demonstrate the necessity of network-aware AI governance. Results suggest that future research should integrate real-time digital trace data with network science methodologies to capture dynamic ecosystem evolution.
This paper proposes a hybrid modeling method that employs dual weighting of attributes and instances for the analysis and prediction of the teacher evaluation data. This method extends the Averaged One-Dependence Estimator (AODE) model by incorporating attribute weights and instance weights. Instance weighting integrates instance weights into the calculation of conditional probabilities, while attribute weighting incorporates attribute weights into the classification formula of the classifier. The improved weighted model, attribute and instance weighted one-dependence estimator (AIWODE), combines an attribute weighting method based on information gain with an instance weighting method based on the mode of the instance set and the similarity to other instances. The experimental data is derived from teacher recording data within the Hubei Province AI-Assisted Teacher Project. By comparing various Bayesian approaches and employing ten-fold cross-validation to validate model performance, the experiments assessed four evaluation metrics: accuracy, recall, AUC, and F1 score. Experimental results show that AIWODE, with its dual weighting mechanism, significantly outperforms existing methods in teaching evaluation tasks.
In the field of dominance robust multi-objective optimization (MOP), how to accurately quantify the robustness of solutions under uncertainty is the challenge in the identification of dominant robust optimal solutions. This paper proposes a historical solution clustering-guided robust Multi-objective optimization algorithm based on decision variable assortment (RMOEA-HSA). It combines decision variable assortment (DVA) with historical solution clustering while introducing a dominance robustness enhancement metric (RobustDR), aiming to accurately quantify the dominance robustness by using systematic analyses of the aggregation patterns of historical solutions. Subsequently, this study develops three strategies: first, employs the aggregated performance of historical solutions to quantify the dominance robustness of individuals; second, an integrated approach combining K-neighbor clustering with an adaptive search mechanism to avoid local optima; last, a DBSCAN and cosine similarity-driven population reduction (DCPR) method for elite individual selection. Superior stability is demonstrated through nine benchmark tests against four existing state-of-the-art algorithms.
Coating systems are frequently used in the automotive and mechanical engineering industries to improve wear resistance and service life. However, premature failures such as cracking, delamination, and interface delamination remain common due to micro-structural defects and residual stresses generated during processing. In this study, an experimental algorithm was developed to evaluate the mechanical properties and failure mechanisms of coatings and to elucidate the relationships between micro-structure, residual stresses, and performance. Coated samples were produced under controlled application conditions and grouped according to process and/or post-processing parameters. Micro-structural characterization enabled the quantitative assessment of layer thickness homogeneity, porosity/void density, and interface quality. Residual stresses were determined using a standardized procedure with repeated measurements to minimize uncertainty. Mechanical performance was evaluated using microhardness testing, wear testing under controlled contact conditions, and adhesion testing. Subsequently, a failure pattern analysis was performed to identify damage initiation sites and the dominant mechanisms. The results show that reducing porosity and improving interfacial continuity increase the stability of the properties and decrease data variability. Residual stresses, primarily compressive stresses, slow down cracking and delamination, while tensile stresses promote premature cracking and delamination. These results demonstrate the feasibility of an integrated process control that combines defect reduction and residual stress management to improve coating reliability.
Based on clarifying the theoretical mechanisms underlying the coordinated development of digitalization and greening, this study takes A-share listed high-carbon enterprises in China from 2015 to 2024 as its sample. Employing a comprehensive methodology that integrates text mining, Word2Vec machine learning, the entropy weight method, a coupling coordination model, kernel density estimation, the Dagum Gini coefficient, and spatial autocorrelation analysis, it measures the level of coordinated digital-green development among these enterprises and systematically examines its temporal evolution, industrial heterogeneity, and spatial distribution characteristics. The findings reveal the following: First, the level of dual-transformation coordination among high-carbon enterprises continuously improved during the study period, with the coordination grade progressing from moderate imbalance to the brink of imbalance. This improvement exhibits a phased evolution pattern characterized by ‘bottom clearing - middle expansion - top breakthrough’. Nevertheless, high-level coordination exhibits significant fragility, indicating that the industry remains entrenched in a pivotal yet precarious transition toward a fully integrated state. Second, coordinated development demonstrates significant industrial heterogeneity, with three types of intra-industry distribution patterns emerging: overall convergence, leading differentiation, and concentrated lagging. These patterns can be attributed to differences in technological complexity, policy consistency, and market structure. Third, regional disparities primarily stem from high internal differentiation within the eastern region. Furthermore, a consistently strengthening positive spatial autocorrelation is observed at the provincial level, resulting in an unbalanced pattern characterized by a Yangtze River Delta ‘high-high’ agglomeration, western ‘low-low’ agglomeration, and ‘high-low’ anomaly zones in some developed provinces. Based on the conclusions, policy recommendations are proposed, including industry-specific measures, tiered enterprise cultivation, regional collaborative linkages, and the development of standards and market mechanisms. Such efforts are pivotal for aligning industrial transitions with national ‘Dual Carbon’ goals and catalyzing high-quality economic transformation.
Electric vehicles (EVs) exhibit significantly different energy consumption characteristics under different driving conditions, posing significant challenges to traditional energy management strategies. Most existing energy management systems (EMSs) are based on fixed rules or predefined driving cycles and, therefore, cannot adapt to dynamic changes in driver behavior under real-world driving conditions. To overcome this limitation, this paper proposes a driver-based adaptive energy management strategy for EVs. First, representative driver behavior characteristics are extracted from vehicle operation data, and a real-time driver behavior recognition model is built to classify driving styles. Based on the identified behaviors, an adaptive energy management mechanism is then developed in which key control parameters and power allocation strategies dynamically adapt to the driver’s driving characteristics. The proposed method is integrated into a hierarchical energy management framework and evaluated through comprehensive simulations under various driving cycles and driver profiles. The simulation results show that the proposed strategy effectively reduces energy consumption while maintaining stable charge state profiles and driving characteristics. Compared with traditional energy management systems, the proposed approach has excellent adaptability, robustness and energy efficiency, highlighting its potential for practical application in battery electric vehicles.
In the context of digital economy, this study constructs the evaluation index system of college students’ entrepreneurial ability based on a SWOT model, aiming to analyze the realization path of college students’ entrepreneurship. It is found that digital literacy (weight 0.203) is the core advantage of college students’ entrepreneurship, which significantly improves the efficiency of entrepreneurship; active innovative thinking (weight 0.126) and strong learning ability (weight 0.126) are the key abilities; and the dividend of the digital economy (weight 0.117) provides important opportunities for entrepreneurship. However, lack of experience (weight 0.081) is the main disadvantage. Through the SWOT portfolio analysis, four types of entrepreneurial strategies were proposed: the SO strategy, the WO strategy, the ST strategy, and the WT strategy, which provide systematic guidance for university students’ entrepreneurship.
In this paper, a method for encoding English world appearance features based on matrix operations is proposed [1]. Existing Soundex or hash encodings exclusively target pronunciation or uniqueness, leaving the visual morphology of English words unquantified. We address this deficit with a matrix-based appearance encoder: the 26-letter alphabet is partitioned into two visual-shape classes [2]; each glyph is fused with class-specific feature matrices via Hadamard products [3] to produce 62-bit letter codes, which are subsequently aggregated and compressed into a fixed 62-digit word fingerprint by a merging rule [4]. Experiment – The experimental protocol is not specified in the present abstract [5]. The encoding preserves visual similarity, reduces code length to below 50 % of the original character count [6], and, to our knowledge, constitutes the first numerical representation of English glyph structure. By overcoming the phonetic-only or uniqueness-focused limitations of prior techniques, the proposed method fills a critical gap in appearance-based word encoding and demonstrates substantial potential for OCR error correction [7] and the retrieval of visually similar words, thereby offering AI systems that handle spelling variants or handwritten input [8] a new technical pathway for multimodal natural-language processing.
Wave-driven unmanned surface vessels (USVs) have become promising long-term ocean observation platforms due to their ability to utilize wave energy for propulsion. This study systematically analyzes the propulsion mechanism of wave-driven USVs, focusing on the underwater propulsion vehicle as the core energy conversion component. Based on hydrodynamic theory, the propulsion mechanism is simplified, and a dynamic model of the underwater propulsion vehicle in the vertical plane is established. A three-dimensional geometric model is created and meshed using ANSYS ICEM software. Numerical simulations are performed using ANSYS CFX software to evaluate hydrodynamic and forward thrust characteristics. Furthermore, a data-driven deep learning framework is introduced to identify and optimize key hydrodynamic parameters in the fundamental equations. Finally, an experimental platform is built in a wave simulation tank to verify the numerical simulation and optimization results. Experimental comparison results demonstrate the effectiveness of the proposed modeling and parameter optimization methods.
The key advantage of using multi-modal image fusion techniques is to reduce information loss and obtain a highly informative fused image to support illness diagnosis. Herein, a new multi-modal image fusion algorithm is proposed based on hybrid filter decomposition (HFD) and multiple-feature preserving. First, the HFD framework is based on a weighted-mean-curvature filter and combing the advantage L1 norm decomposition and L2 norm decomposition (hybrid l0 − l1 layer decomposition) to effectively separate the features of the source images, and the source images are decomposed into fine-scale detail, small-scale detail layer feature, median-scale detail layer, and large-scale detail layer images; Second, the contrasting salience, fuzzy entropy, sum-modified Laplacian, laws of the texture, and multiple-direction gradient are used to active measure of the different features, and then the different fusion rules are adopted to fuse the multiple-features. Finally, the final fused image is obtained by linearly combining the multiple-level-feature fused images. Experimental results demonstrate that the qualitative and quantitative results. The proposed method excels in preserving more contrast, brightness, detail, and edge features, which proves advantageous for follow-up image processing task.
Modern object detection models deployed on edge and embedded platforms are increasingly exposed to sporadic numerical faults, such as low-level bit-flips in stored weights or intermediate activations. These disturbances—arising from memory instability, transient hardware behavior, or environmental factors—can lead to abnormal feature propagation and severe degradation of detection accuracy. This work investigates the vulnerability of convolutional object detectors to such bit-level perturbations and proposes a lightweight static–dynamic stabilization framework to enhance robustness. We first analyze the numerical sensitivity of IEEE 754 FP32 representation and construct a fault model that considers both static weight corruption and dynamic activation disturbances. A layer-wise sensitivity evaluation is then carried out on YOLOv8n, revealing that shallow backbone layers and mid-level feature-fusion modules are particularly susceptible to small bit-level deviations. To improve robustness with minimal overhead, we introduce a two-part stabilization strategy. On the static side, convolution kernels are decomposed via truncated SVD and protected using block-wise Reed–Solomon coding applied to sign and exponent bytes, enabling compact storage and effective correction of perturbed weights. On the dynamic side, a rectified weighted Gaussian (RWG) activation is adopted to suppress extreme activations induced by numerical faults while preserving normal feature statistics. Experiments on RSOD demonstrate that the unprotected YOLOv8n detector quickly collapses under moderate bit-flip rates, whereas the proposed static–dynamic scheme maintains stable performance (mAP50 ≥ 0.87 at flip prob = 10⁻⁴) with less than 7% runtime overhead and nearly 30% model size reduction. These results suggest that the proposed approach provides a practical, low-overhead robustness enhancement for object detectors subject to bit-level numerical disturbances.
As a clean and renewable resource, solar energy offers a significant advantage over nonrenewable energy sources due to its abundance. Photovoltaic (PV) systems represent a promising technology for the direct conversion of sunlight into direct current electricity. Identifying PV model parameters is a complex nonlinear optimization problem. Traditional metaheuristic algorithms often encounter several challenges. They may converge slowly, get stuck in local optima, or require complex parameter tuning to perform well. In response to these challenges, this study proposes a novel parameter estimation technique, IALA. It combines the Artificial Lemming Algorithm (ALA) with a triangular walking strategy and an inverse cumulative Cauchy distribution operator. The triangular walking strategy enhances the exploration ability of ALA and reduces the risk of getting trapped in local optima. At the same time, the inverse cumulative Cauchy distribution operator strengthens its exploitation mechanism, accelerating convergence toward the optimal solution. The proposed algorithm is used to estimate the optimal parameters of single-, dual-, triple-, four-diode PV models and compared with several existing well-established algorithms. Systematic comparisons over 100 independent runs, the improved ALA (IALA) performs better than other techniques. It provides more accurate and robust parameter estimation for various PV models. Furthermore, its robustness is validated through statistical assessments, including the wilcoxon signed-rank test.
