ABSTRACT This paper designs and investigates an STM32-based multicore load-balanced communication platform tailored for industrial field environments. The platform fully leverages the low power consumption and high integration characteristics of the STM32 microcontroller, incorporating multiple optimised communication modules. These include an anti-interference RS485 communication system based on the Modbus protocol, a high-speed communication mechanism between the STM32 and an FPGA coprocessor via the FSMC parallel interface, and a wireless laser communication subsystem equipped with multipath channel estimation and parallel equalisation capabilities. For wireless communication, the platform employs dynamic clock frequency adjustment and sleep mode strategies, significantly reducing the power consumption of both the processor and the RF module, achieving a low power level of 10–30 μJ under continuous data transmission. Experimental results obtained using the SMARTConvert test platform show significant improvements in communication stability, energy efficiency, and throughput. These results validate the feasibility and advantages of the STM32 multicore architecture in achieving load-balanced communication for industrial applications.

ABSTRACT A target detection model based on the upgraded version of YOLO-v4 is proposed to address the problems in image recognition of power engineering equipment, such as poor visual effects, large recognition gaps, and inconsistent image styles. A target detection dataset of infrared temperature difference photos of power engineering equipment was created, and the MSR_CR method was employed to enhance the contrast and clarity of the images in obscured weather conditions. The Focal_loss function is employed to address the classification challenges arising from the unequal recognition of image data. The experimental results show that after iteration to about 5000 times, the model performance stabilises, the loss factor converges to about 0.2, and the recognition accuracy is close to 100%. The image enhancement method is tested to achieve homogeneous recognition of multiple types of power engineering equipment with 96.31% recognition accuracy and 71 frames/second detection rate. The experiment verifies the scientificity of the novel enhancement recognition method.

ABSTRACT Preschool dance education is essential for promoting physical, cognitive, and social development in young children. Although traditional teaching methods are effective, they often lack interactive elements that sustain engagement and improve learning outcomes. Virtual Reality (VR) technology offers a promising solution by providing immersive, personalised, and engaging learning experiences that complement conventional dance instruction. This study aims to evaluate the impact of VR based lessons in preschool dance education by comparing VR and traditional approaches in terms of movement accuracy, rhythm synchronisation, engagement, and motivation. The VR module integrates real time motion tracking and posture analysis to guide children during dance activities. It provides immediate corrective feedback and rhythm cues, enabling children to adjust movements while practising and strengthening coordination and accuracy. A pre test post test experimental design was employed, with one group receiving VR based instruction and another receiving traditional dance lessons. Quantitative analyses, including paired t tests and independent t tests, were conducted to assess differences in movement accuracy and rhythm synchronisation between groups. Qualitative feedback from teachers, parents, and children was analysed using thematic analysis to evaluate engagement and usability. Results showed that the VR group achieved significantly higher post test movement accuracy scores than the traditional group, along with improved rhythm synchronisation and engagement levels. These findings indicate that VR based dance education provides a more engaging and effective learning experience for preschool children.

ABSTRACT Tennis techniques are characterised by high openness, rapid speed, and complex limb coordination. Traditional evaluation methods based on manual observation or laboratory motion capture struggle to support efficient, precise, and continuous monitoring in real training scenarios. To address challenges in pose estimation during high-dynamic movements – such as occlusion, complex backgrounds, and temporal noise – this paper proposes PoseCResNet-R, a 2D human pose estimation model based on multi-scale feature fusion and spatio-temporal consistency correction. PoseCResNet-R performs 2D pose estimation, while STSPose refines the 3D motion sequence with temporal upsampling, ensuring smooth motion reconstruction. By integrating a Transformer-based temporal structure, we further develop STSPose, a 3D pose refinement model that achieves high-precision reconstruction of continuous 3D motion sequences from sparse 2D keypoints. Experimental results demonstrate that this framework achieves superior keypoint localisation stability, motion sequence smoothness, and inference efficiency in open tennis training scenarios. Building upon this foundation, this paper reveals the coupling mechanisms between lower-body power generation, trunk rotation, and upper-body acceleration from a kinematic chain transmission perspective. It analyzes quantitative relationships between joint angle changes and racket velocity/ball strike height, providing structured motion feedback for tennis technique training.

ABSTRACT A personalized recommendation system for graphic design helps users quickly discover relevant and inspiring content aligned with their individual interests and creative activities. Traditional approaches, such as collaborative filtering, are widely adopted but suffer from data sparsity, cold start problems, limited contextual awareness, and poor adaptability to evolving user preferences. To address these limitations, this study proposes a hybrid recommendation framework that combines collaborative filtering with enhanced personalization driven by user interaction history and explicit feedback. The process begins with comprehensive data collection, including user generated designs and detailed interaction metadata. Preprocessing techniques, including resizing and normalization, ensure consistent input quality. Meaningful visual features are extracted using color histograms and layout analysis, while design styles are identified through Convolutional Neural Networks. Scalability for large datasets is achieved using a Hadoop based big data pipeline supported by MapReduce. Additionally, the system integrates an Artificial Intelligence Generated Content module trained with a fine tuned Generative Adversarial Network optimized using AdaGrad++. This component enhances creative generation and personalized recommendations. The proposed hybrid model effectively learns user preferences and delivers highly relevant graphic design suggestions. Experimental results demonstrate strong performance, achieving 98.25 percent accuracy, 99.25 percent recall, and a 98.75 percent F1 score. Overall.

ABSTRACT Based on the fuzzy decision algorithm, this paper analyses the personalised strategy in college English teaching and designs an optimal learning path for students. A personalised learning path recommendation system is built through data collection and analysis, and multidimensional data such as students’ class participation, homework achievement, learning time and learning interest are comprehensively considered. Combined with a fuzzy decision algorithm, different students are recommended to meet their actual needs of learning paths to improve the learning effect. The experimental results show that a personalised learning path recommendation system can improve students’ learning motivation, class participation and academic performance. In low-motivation student groups, the effect of personalised strategies is weak, and future research should focus more on these student groups, optimise the system algorithm and recommendation strategies, and better adapt to individual differences.

ABSTRACT This paper focuses on the architectural environment design of Buddhist temples in northern Anhui Province, and explores the application of integer planning algorithms in the design of temple environments and their effects on the transmission of cultural context. The integer planning model is used to optimise the spatial layout of the temple building to maximise the impact of the transmission of cultural context. The integer planning model is optimised by the proposed gradient method and genetic algorithm. Fuzzy hierarchical analysis is introduced to construct an evaluation index system to analyse the transmission effect of cultural mood quantitatively. The relative affiliations of the temple environment design plan based on integer planning and the temple environment design plan based on the general method are 0.0576 and 0.0259, respectively. According to the principle of maximum affiliation, the integer planning model design scheme is better at conveying the cultural context. In addition, the architectural environment of Buddhist temples should be designed for the functional principle as the basis, the cultural principle as the essence, and the artistic principle as the method.

ABSTRACT Ensuring tourist safety in dynamic outdoor environments is critical, especially in regions prone to natural hazards such as wildfires, landslides, and flash floods. With global tourism growth, intelligent systems capable of real-time monitoring and early warning are increasingly necessary, as conventional approaches lack adaptability and predictive capability. This study proposes an integrated intelligent framework for tourism safety monitoring and early warning by combining ecological sensing, mobile data acquisition, and advanced machine learning. Real-time hazard data and anonymized tourist movement patterns are collected and pre-processed using Noise-Resistant Adaptive Normalisation (NRAN) to handle noise and missing values. Hierarchical Spatio-Temporal Encoding (HSTE) is applied for effective feature extraction. Hazard prediction is performed using a Binary Gannet Optimiser-driven Dynamic Random Forest Tree (BGO-DRFT), which adaptively responds to evolving environmental and visitor conditions. The system integrates IoT sensors, UAVs, and mobile applications for real-time alerts and response planning. Experimental results demonstrate improved accuracy, reduced false alarms, and faster response times compared to existing methods.

ABSTRACT This study integrates IoT and deep learning to develop a comprehensive framework for Green Port development aimed at promoting sustainable logistics that supports regional economic growth in coastal and inland areas of China. Using spatial econometrics and panel data analysis, the research investigatexs the complex coupling between logistics and economic subsystems, highlighting spatial heterogeneity and the significant influence of local economic factors. The study employs a three-dimensional spatial model to examine the evolving dynamics of logistics and business flows. At the same time, an evaluation index system is constructed to assess the coordination of Green Port subsystems. ARIMA-BP predictive modelling for the 2020–2025 period suggests a trend towards improved coupling and coordination across most regions, providing strategic insights for sustainable logistics investments and regional integration. This work contributes to the understanding of Green Port logistics development and offers policy recommendations for regional economic alignment.

ABSTRACT Financial risk prediction is a crucial indicator for maintaining the stability and growth of Chinese listed firms, as accurate forecasting can help reduce the impact of financial distress. This paper proposes a novel deep learning–based approach using the TabNet architecture, optimized with the Multiple-Strategy Siberian Tiger Optimization (MSSTO) algorithm, to predict financial risk in Chinese companies. The proposed model aims to enhance prediction accuracy and computational efficiency by optimizing hyperparameters through MSSTO, which simulates the natural hunting behavior of Siberian tigers to achieve effective global exploration and local exploitation of the search space. The scalable attention mechanism of TabNet is integrated into the framework to improve feature selection and model interpretability, enabling deeper insights into key financial risk factors. Model performance is evaluated using multiple metrics, including R-squared, Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Square Error (RMSE). The experimental results demonstrate strong predictive capability, with an R-squared value of 0.9766, indicating that the model explains 97.66% of the variance in the data. Overall, the findings highlight the effectiveness of the proposed approach in supporting accurate financial risk prediction and informed decision-making for financial analysts, investors, and policymakers.