Saliency object detection (SOD) aims to identify the most visually prominent regions in an image that attract human attention. Despite significant advancements in deep learning-based SOD methods, existing approaches still face challenges in effectively integrating multi-scale, multi-semantic, and multi-spatial features, leading to incomplete saliency prediction, blurred boundaries, and poor generalization to complex scenes. To address these issues, this paper proposes a novel SOD framework that combines multiple-perspective feature fusion (MPFF) and deep adversarial networks (DAN), named MPFF-DAN. First, a multi-perspective feature extraction module is designed to capture complementary information from three critical perspectives: (1) the spatial perspective (low-level features with precise spatial localization); (2) the semantic perspective (high-level features with strong category-aware representation); (3) the scale perspective (multi-scale features to adapt to objects of varying sizes). Second, an adaptive feature fusion network (AFFN) is proposed to dynamically weight and aggregate the multi-perspective features, leveraging a dual-attention mechanism (channel attention + spatial attention) to enhance the discrimination of salient regions while suppressing background noise. Third, a deep adversarial network is integrated into the framework, where a generator (based on an improved U-Net) generates highquality saliency maps, and a discriminator (a multi-scale convolutional neural network) distinguishes between the generated saliency maps and ground-truth masks. The adversarial training paradigm drives the generator to produce more realistic and boundary-preserving saliency results. Extensive experiments are conducted on five benchmark datasets using four evaluation metrics. Quantitative and qualitative results demonstrate that MPFF-DAN outperforms 15 state-of-the-art (SOTA) methods. It maintains high efficiency with a computational complexity.
Accurate quantification of natural landscapes is of great significance to the study of ecological benefits. At present, most systems are still dominated by human intervention. This paper addresses the limitation of existing landscape design systems that prioritize visual aesthetics over ecological functionality and lack accurate ecosystem simulation. It proposes a natural landscape generation model combining L-system theory, complex network analysis, and Generative Adversarial Networks (GAN). The system encodes L-system rules as input to a GAN generator, enabling the creation of realistic and ecologically meaningful landscape structures. By analyzing landscape characteristics, topology, and growth functions, the model ensures effective visualization of complex landscape components. Experimental results demonstrate that the proposed L-system-GAN model achieves significantly higher resolvability and operates 5.4 times faster than traditional GAN approaches. Additionally, it meets commercial standards in generation quality, computational efficiency, and user experience, with strong potential for deployment in embedded systems. Therefore, subsequent optimization directions can focus on extreme scenario robustness and energy consumption control. The proposed framework utilizes publicly available terrain and landscape-related datasets, with key training parameters including controlled learning rate, batch size, and iteration settings for GAN optimization. Performance is evaluated using fractal dimension (FD), resolvability (SVR), and generation time metrics, ensuring consistent and verifiable assessment of model effectiveness.
Mental health issues among university students are increasing and negatively affect academic performance and quality of life. Traditional clinical interviews and surveys are time-consuming, costly, and often delayed, limiting timely intervention. This study proposes a rapid mental health screening system using machine learning to classify students into low, moderate, and high-risk groups by processing both structured and unstructured data. The framework integrates ELMo contextual embeddings to capture semantic meaning from text and CNN to extract high-level features. Demographic, lifestyle, and behavioral data are also incorporated. The proposed ELMo-CNN model achieved strong performance with 96.82% accuracy, 96.14% precision, 95.47% recall, 95.80% F1-score, and 0.9821 ROC-AUC, outperforming baseline models such as SVM and CNN. The results indicate that the model is effective for fast and accurate mental health screening in real-world educational settings.
The large-scale integration of distributed photovoltaic (DPV) systems introduces significant spatio-temporal variability into distribution networks, posing challenges to voltage stability, line loading, and protection coordination. This paper proposes a spatio-temporal analysis and modeling framework to quantify the safety impacts of DPV contributions. By integrating spatial topology characteristics with temporal generation-load fluctuations, the proposed model evaluates dynamic safety margins under varying penetration scenarios. A composite load model incorporating ZIP components, induction motor dynamics, and a simplified PV subsystem is established, and a reinforcement learning-enhanced Grey Wolf Optimizer (RL-GWO) is developed for high-precision parameter identification. Case studies demonstrate its effectiveness in identifying high-risk nodes and revealing critical spatio-temporal coupling effects, providing analytical support for distribution network planning and secure operation, while showing that appropriate DPV penetration can significantly improve voltage profiles and reduce network losses.
Environmental pollution has become increasingly severe due to multiple contributing factors, among which the rapid growth of industrial waste is particularly significant. In practice, the glass industry generates substantial waste daily, primarily in the form of grinding glass powder. While cullet can be recycled through remelting, the ground glass powder is often directly discharged into the environment or disposed of in landfills. Therefore, the treatment and valorization of industrial waste, particularly glass powder, have become research topics for mitigating environmental pollution. In this study, waste glass was used as a silica source to synthesize calcium silicate hydrate and investigate the effect of the CaO/SiO2 ratio on phase formation. Waste glass powder was mixed with lime, pelletized, and subjected to hydrothermal treatment at 180 oC for 10 hours. The formation of Xonotlite and Tobermorite was confirmed by XRD, FTIR, SEM, and EDX analyses, demonstrating the feasibility of synthesizing calcium silicate hydrate from waste glass powder. Mechanical and physical properties, including diametral tensile strength, water absorption, and bulk density, indicated that the optimal CaO/SiO2 ratio for synthesizing calcium silicate hydrate from waste glass is 0.9.
This study examines how digital language support services improve communication, visitor interaction, and service quality at multilingual tourist attractions. A mixed-method approach was used, involving quantitative data from 200 respondents through structured questionnaires and qualitative insights from 30 participants using interviews and focus group discussions. Quantitative analysis included descriptive statistics, Cronbach’s Alpha, Pearson correlation, and multiple regression, while qualitative data were analyzed through thematic coding and frequency mapping. The findings indicate moderate to high satisfaction with digital language services. Technology-related satisfaction recorded the highest mean score (3.14), whereas translation accuracy showed the lowest (3.01), suggesting a need for improvement. Correlation analysis revealed that language barriers negatively affect tourist satisfaction (r = -0.369), while effective digital language services positively influence satisfaction (r = 0.417). Regression analysis (R² = 0.712) showed that usability is the strongest predictor of tourist satisfaction. The study emphasizes the importance of accurate, user-friendly, and context-aware digital language systems in multilingual tourism environments.
Electric vehicles (EVs) are undergoing rapid expansion, leading to an explosive increase in ownership. This surge, however, poses unprecedented challenges to power grid operation and charging infrastructure planning. In particular, accurately forecasting charging demand has become increasingly difficult. Conventional graph-based or single-modal models often fail to capture the complex and nonlinear nature of charging behaviors because (1) transportation networks exhibit pronounced spatiotemporal heterogeneity, and (2) charging stations are characterized by intricate, multi-scale interdependencies. To overcome these limitations, this study introduces a multimodal hypergraph-based charging demand forecasting framework. The proposed approach constructs multidimensional spatiotemporal hypergraphs from heterogeneous multimodal data, thereby uncovering higher-order correlations among charging stations across different spatial and temporal scales. By dynamically modeling these evolving inter-station relationships, the framework enhances the accuracy of charging demand prediction and provides a robust analytical foundation for the optimal planning and management of EV charging infrastructure. Building upon the proposed methodology, this study validates its effectiveness using a real-world EV charging demand dataset collected from an urban area and further develops an intelligent charging station operation system. Extensive experiments and ablation analyses demonstrate that the proposed multimodal hypergraph based framework consistently outperforms unimodal and conventional graph-based baselines, confirming its superior capability in capturing complex spatiotemporal dependencies and improving forecasting accuracy.
Wind power forecasting is very crucial to ensure stable renewable energy (RE) integration and the stability of power systems. In this research work, several machine learning (ML) models, such as linear regression (LR), random forest (RF), gradient boosting (GB), extreme GB (XGBoost), and grid search cross-validation (GS-CV) optimized XFGBoost are analyzed in terms of their performance for wind power forecasting using a real-life large dataset having weather and temporal features. Performance of these models is analyzed using metrics such as mean absolute error (MAE), root mean square error (RMSE), and correlation factor (R2). It is found that XGBoost with GS-CV approach consistently performs better than other models with the minimum prediction error and maximum R2 value. The results show that hyperparameter-optimized XGBoost can greatly enhance forecasting accuracy and generalization, providing a practical data-driven solution for wind power forecasting and decision-making.
The rapid growth of digital music collections has increased the demand for automatic composer identification and musical style recognition. This study proposes a hybrid framework using the MAESTRO v3.0.0 symbolic MIDI dataset. EB-SMR preprocesses MIDI by extracting pitch, onset, duration, and velocity, followed by tempo, pitch, and duration normalization. The data are converted into piano-roll representations. Handcrafted symbolic features capturing melodic, rhythmic, and dynamic properties are combined with Convolutional Neural Network (CNN) -based deep features, enabling improved feature learning for composer-aware music generation and style recognition tasks. These complementary features are fused into a unified representation and classified using a Hybrid Feature-Based Softmax Neural Network (HFSNN), while the Lion optimizer is employed as a training strategy to improve convergence efficiency during optimization. The framework incorporates composer-aware music generation by sampling musical events from class-conditional distributions to produce stylistically coherent MIDI sequences. Experimental results show that the hybrid framework achieves 0.97, outperforming CNN-only and handcrafted-feature models. Combining symbolic descriptors and CNN features improves style/composer recognition, enables composer-aware MIDI generation, but is limited to piano MIDI; future work explores multimodal and multi-instrument representations.
In the context of the knowledge economy and the transformation of higher education, cultivating innovation, entrepreneurship, and moral responsibility has become a strategic priority. Existing innovation and entrepreneurship education platforms often fail to address diverse student profiles and rarely integrate moral education into personalized learning pathways. This study proposes an integrated teaching platform that unifies innovation, entrepreneurship, and moral education through multimedia networks and neural network-driven cross-modal semantic retrieval. The platform consists of a student learning space, teacher management modules, a multimedia resource repository, and real-time feedback mechanisms. A hybrid neural network model is introduced to map multimodal educational resources and student submissions into a shared semantic space using weak semantic label generation, bidirectional feature crossing, attention-enhanced GRU modules, and position encoding. In this framework, weak semantic labels are automatically generated pseudo-labels derived from semantic category probability distributions of unlabeled multimodal samples and are iteratively incorporated into model training to enhance semantic representation learning. This framework enables personalized content recommendation, dynamic interest tracking, and moral dilemma feedback. Experiments conducted on the Wikipedia, Wikipedia-CNN, NUS-WIDE, and domain-specific I&E-EDU datasets demonstrate that the proposed method consistently outperforms baseline approaches, including MRCR-SNN. On the I&E-EDU dataset, the proposed model improves mean Average Precision (mAP) from 28.3% to 32.7% (+4.4 percentage points), Precision@10 from 43.1% to 49.4% (+6.3 percentage points), and increases student engagement by 26.5%.
To address limitations in real-time responsiveness, spatial perception, and long-term stability in emission monitoring of coal-fired power plants, this paper proposes an edge–cloud collaborative video-based monitoring framework. The system integrates video acquisition, image preprocessing, multi-scale target detection, emissionstate recognition, and cloud-based collaborative analysis into a unified pipeline. An improved YOLO-FPN model is developed to enhance the detection of plume regions, flame instances, and abnormal operating zones. In addition, a multi-node edge collaboration strategy with dynamic model updating is introduced to reduce latency and improve adaptive performance over time. Experiments conducted on a real industrial testbed demonstrate improved detection accuracy, reduced processing delay, and lower network bandwidth consumption, while maintaining stable online deployment capability.
To address illumination-induced color recognition instability, ball stacking and jamming during continuous delivery, and reduced reliability of intelligent delivery robots in complex environments, this paper proposes a robot intelligent delivery method integrating an anti-jamming storage structure and an adaptive CLAHE-LAB spatial feature fusion algorithm. First, a passive anti-jamming ball storage structure based on ball-diameter constraints is designed for continuous ball storage and delivery. A top light shield reduces external illumination interference, a ball-diameter-matched multi-hole inlet array disperses the ball entry path, and inner-wall antibridging limiting ribs disrupt stable support states caused by multi-ball stacking, thereby reducing the risk of jamming and discharge interruption. Second, an adaptive CLAHE-LAB spatial feature fusion algorithm is proposed to overcome false and missed detection caused by traditional RGB/HSV threshold segmentation under weak, strong, and non-uniform illumination. By decoupling luminance and chromaticity in LAB space, enhancing local contrast, and fusing color saliency features, the separability between blue targets and the background is improved. Finally, an experimental platform is built to evaluate color recognition, segmentation accuracy, jamming frequency, and continuous delivery success rate. Experimental results show that the proposed method improves blue target recognition stability, alleviates ball stacking and jamming, and enhances continuous delivery reliability.
New energy vehicles (NEVs), powered by sustainable energy, are emerging as a cleaner alternative to conventional vehicles. Their performance and safety depend heavily on the health of critical components, making accurate fault diagnosis essential. This study proposes an advanced fault detection framework for NEVs, focusing on fault identification and classification in the drivetrain using deep learning techniques. Sensor data from electric vehicles, including current, voltage, motor speed, and environmental conditions, are pre-processed through normalization and missing-value imputation to ensure consistency. Feature extraction is performed using Wavelet Transform (WT) and Fast Fourier Transform (FFT) to capture both transient and steady-state behaviours. The proposed Enriched Crow Search Optimizer-based Adaptive Gated Long Short-Term Memory (ECS-AG-LSTM) model enhances LSTM adaptability and fault classification accuracy through optimization. Experimental results demonstrate superior performance, achieving 98.5% accuracy, 98.31% recall, 98.42% precision, 98.51% F1-score, and an R 2 value of 0.956 . The findings confirm that ECS-AG-LSTM provides a robust and reliable solution for improving NEV safety and operational dependability.
Extreme climatic conditions pose significant operational challenges to power systems by increasing fault occurrences, communication failures, voltage instability, and large-scale outages. This study proposes a highresilience integrated emergency interconnection and communication framework for reliable power system operation during extreme weather events. The proposed system combines Optimal Power Flow (OPF) control with hybrid communication technologies, including satellite, unmanned aerial vehicle (UAV), and Mobile Ad Hoc Network (MANET)-based communication systems. In addition, a hybrid artificial intelligence fault detection model integrating Random Forest (RF) and Support Vector Machine (SVM) with ensemble learning is developed for accurate fault identification. The methodology is implemented on the IEEE 14-Bus System using the Newton–Raphson load flow method. Extreme climatic conditions are modeled through scenariobased simulations, while Monte Carlo simulations are employed to generate synthetic disturbances and faults. Communication performance is evaluated through delay and packet loss analysis within hybrid fiber–IoT communication networks. The framework also incorporates rule-based communication selection, optimizationdriven coordination strategies, load shedding optimization, and power balancing mechanisms to maintain system stability during emergencies. Experimental results demonstrate strong resilience and communication performance, achieving a Voltage Deviation Index (VDI) of 0.04281 p.u., Resilience Index (RI) of 0.7241, Recovery Time (RT) of 5.64 hours, and Energy Not Supplied (ENS) of 2.85 MWh. The communication system records latency between 0.005 s and 0.018 s, throughput of 99.34 Mbps, and a Communication Reliability Index (CRI) of 0.9782. The hybrid fault detection model achieves 98.83% accuracy, confirming the effectiveness of the proposed framework for resilient smart grid operations.
This research provides a numerical evaluation of the impact response of cold-formed rectangular hollow steel tubular (RHST) columns, using finite element models created and validated in ABAQUS/Explicit. A parametric analysis was carried out by varying three major parameters: cross-section properties (wall thickness and sectionsize), impact angles (0 ◦, 45◦, and 90◦, where 0◦corresponds to axial impact, 45◦ to oblique impact, and 90◦ to transverse impact orientations), and impact locations (mid-span, L/2, and quarter-span, L/4), while maintaining an approximately consistent global slenderness level across all models. It is observed that the ultimate impact load increases substantially with increasing wall thickness and section size, while thicker, larger sections exhibit better deformation control. Impact orientation also affects the response mode, with 0◦ and 90◦impacts resulting in higher impact loads and global deformations, whereas a 45◦ impact reduces global response but increases local instability. Impact location also affects impact response, with L/4 impacts generally resulting in higher ultimate impact loads but lower peak displacements than L/2 impacts. Failure modes change from local denting in thin-walled sections to a combination of global buckling and shear deformation in thicker-walled sections. This research can guide designers in designing RHST columns for accidental lateral impact, with consideration of ultimate impact capacity and deformation response.
Short-term heating energy consumption prediction is essential for intelligent control and efficient energy allocation in centralized heating systems. To address the limited nonlinear modeling ability of traditional time-series methods, the insufficient long-term dependency learning of conventional machine learning models, and the weak adaptability of existing LSTM-based models to multi-source heterogeneous features, this study proposes an improved AFW-Attention-LSTM model integrating adaptive feature weighting and local attention mechanisms. The model follows a four-stage framework of feature optimization, temporal extraction, keysegment enhancement, and prediction output. The adaptive feature weighting module dynamically adjusts the importance of environmental, operational, and building-related features to reduce redundant information interference, while the local attention mechanism emphasizes critical temporal segments such as extreme operating conditions and period transitions. Experiments using 121 days of real operational data from a residential community in a severe cold region show that AFW-Attention-LSTM outperforms standard LSTM, GRU, and Attention-LSTM under both normal and extreme operating conditions. In 1–24 h forecasting tasks, it achieves lower MAE and RMSE, with a maximum prediction accuracy of 96.7% for 1-hour forecasting under normal conditions. The model also demonstrates good training efficiency and prediction stability, providing an effective technical solution for accurate short-term heating energy consumption forecasting and supporting intelligent scheduling and optimized operation of centralized heating systems. However, since the data were collected from a single residential community, further validation across different regions and heating systems is required.
This paper presents an integrated electric-gas energy system fault recovery method based on bidirectional coupled devices. Simulation examples of an IEEE 13-node electric system and a 6-node natural gas system are used to validate the effectiveness of the method. YALMIP is used to model the simulations in the MATLAB environment, while CPLEX is used to solve the optimization issue. Three recovery strategies are compared under the system failure scenarios: independent failure recovery, failure recovery with GFT coupling only, and failure recovery taking into account both GFT and P2G coupling. The power and natural gas systems are bi-directionally coupled through one GFT and one P2G device. The outcomes demonstrate that the recovery strategy that incorporates the consideration of two-way coupled devices can greatly enhance the recovery of natural gas and electricity loads. Scenario 3’ s total recovery is 41,488.12, which is 66.64% and 55.94% of Scenario 1 and Scenario 2’ s respective recoveries. Through a detailed analysis of the changes in network topology and the output of energy conversion equipment during the recovery process, the resilience and effectiveness of the method presented in this study under various fault scenarios are confirmed.
The study focuses on developing a new methodology that would give students the tools they need to participate in the worldwide, digitalized music market as well as integrating intercultural competency and entrepreneurial abilities into undergraduate music education. It further asks whether a curriculum combining intercultural awareness and entrepreneurial competences truly adheres to its intent of equipping students to take responsibility for their learning by actively engaging in applied art studies. A mixed-method action research approach with 30 undergraduate music students, the framework achieved 75–78% improvement in intercultural competence, 70–74% improvement in entrepreneurial skills, and 15–20% improvement in project performance, demonstrating its effectiveness as an integrated competency-based music education model. It stresses the need for flexible learning environments centred on students to nurture autonomy and self-development, creativity, and flexibility. The model contends on different sources of motivation while focusing mostly on intrinsic motivation being able to nurture entrepreneurial behaviours-what is being assumed is that others will need some scaffolding to support students of different levels of self-regulation.
Artificial intelligence (AI) has significantly advanced machine translation through deep learning-based approaches. This study presents a comparative evaluation of Neural Machine Translation (NMT) and TransformerBASE models for Chinese-English automatic translation using a parallel corpus containing 200,000 bilingual sentence pairs. The dataset was pre-processed through sentence cleaning, tokenization, and Byte Pair Encoding (BPE)-based subword segmentation to improve translation quality. Both models were implemented and evaluated using Bilingual Evaluation Understudy (BLEU), METEOR, and Translation Edit Rate (TER) metrics. Experimental results demonstrated that the Transformer model outperformed the conventional Neural Machine Translation (NMT) model, achieving improvements of +6.30 BLEU and +5.60 METEOR while reducing TER by 3.77. Ablation analysis further confirmed the importance of multi-head self-attention and positional encoding in improving translation performance. The findings indicate that Transformer-based architectures provide higher translation accuracy, semantic fluency, and computational efficiency, making them suitable for modern multilingual translation applications.
To investigate the differences in thermogravimetric characteristics and heat release behavior of various tobacco leaves, flue-cured, oriental, and Cuibi No. 1 tobaccos were analyzed using thermogravimetric analysis coupled with a programmed heating platform and gas collection system. Partial least squares discriminant analysis (OPLS-DA) was employed to identify key differential substances across heating stages. The thermal weight-loss process for all samples comprised four stages, with characteristic parameters varying significantly among tobacco types. Oriental tobacco exhibited a higher weight-loss rate at stage II, while Cuibi No. 1 demonstrated a superior comprehensive pyrolysis index of 3.36 × 10−4 (%/min ·◦C2). Kinetic analysis revealed that the F1.5-order chemical reaction model adequately described the decomposition processes. Activation energy at stage II ranged from 54.30 to 89.20 kJ/mol, with flue-cured tobacco showing significantly higher activation energy at stage III than the other two types. Stage II was identified as the primary phase for aroma product formation, where the total release amount followed the order: flue-cured tobacco (15,278.56 µg/g) > oriental tobacco (14,974.00 µg/g) > Cuibi No. 1 (13,190.34 µg/g), with aroma components accounting for over 84% of the total aroma compounds for each variety. OPLS-DA identified 20, 10, and 5 key differential substances for stages II, III and IV, respectively. These findings provide data support for understanding quality characteristics of tobacco materials and may offer guidance for product formulation design.
