Personalized clothing customization often struggles to integrate precise 3D human body data with high-fidelity digital simulation. In this paper, the authors present an end-to-end framework that bridges this gap by merging conditional deep generative adversarial networks with differential geometry-driven finite element simulation. Their generation method extracts 3D principal component analysis features and style preferences into a conditional variational autoencoder, optimized by a composite loss function that combines curvature matching and size constraints to produce tailored 3D meshes and 2D patterns. During virtual display, the authors incorporated curvature matching and geodesic regularization into the finite element model to accurately simulate fabric deformation. Experimental results confirmed the framework's efficacy, reducing mean fitting error to 2.5 mm and achieving a key size deviation of 1.8%. Additionally, the authors' simulation model attained a 35.0 frames per s rendering rate with a 90.0% physical consistency score, demonstrating a robust solution for automated digital clothing customization.
Chemical process operations require fault-adaptive and energy-aware automation, yet conventional control methods often struggle to maintain robustness, fault tolerance, and energy efficiency under dynamic operating conditions. This paper presents an artificial intelligence-enhanced process monitoring and control framework for the Tennessee Eastman process that integrates a hybrid temporal convolutional neural network-bidirectional gated recurrent unit encoder for fault detection with a safety-constrained reinforcement learning-based adaptive controller. The proposed framework achieves superior fault detection performance, with an F1 score of 0.957, an area under the receiver operating characteristic curve of 0.981, and a median detection time of 10.2 s, outperforming long short-term memory and autoencoder baseline models. It also delivers improved control performance, achieving an energy index of 0.866 and a control efficiency of 1.113, while reducing energy consumption compared with the long short-term memory, autoencoder, and model predictive control baselines. These findings demonstrate that combining temporal feature learning with safety-constrained policy optimization provides a practical approach for developing more resilient, energy-efficient industrial process automation systems.
Students’ family backgrounds and behaviors can affect their learning difficulties, but teachers managing large classes often cannot monitor each student’s complex situation. Machine learning offers a solution by identifying anomalies in student performance data, enabling early detection of at-risk students. This study proposed a method that preprocessed data and applied smooth target encoding with a global mean prior and smoothing to address high-dimensional sparsity and small-sample category noise. Fitting encoding was performed on the training set to prevent data leakage into features. After training, temperature scaling calibration on the validation set produced probability outputs suitable for threshold-sensitive decision-making. Experimental results showed 92.31% accuracy, 95.24% F1 score, and 97.97% receiver operating characteristic–area under curve, demonstrating strong predictive performance for final student grades. This approach provided robust technical support for teachers to enhance teaching quality.
The rising prevalence of substance abuse and overdose incidents underscores the need for real-time public health surveillance. Social media offers valuable signals for monitoring these events; however, noisy language, slang usage, and class imbalance present significant challenges for automated analysis. To address these issues, the authors propose ATTEND, a multi-task neural network for substance classification and detection of 18 overdose symptoms, with symptom normalization to standardized MedDRA concepts. ATTEND was trained on a large multi-source corpus combining ADE Corpus V2 and the UCI Drug Review Dataset, comprising over 100,000 samples designed to emulate realistic social media communication. Experimental results show that ATTEND achieved 93.23% accuracy and 93.41% weighted-F1 for substance classification, 94.10% micro-F1 for overdose symptom detection, and 90.42% accuracy for symptom normalization, outperforming baseline multi-task models across all tasks. The framework is scalable, privacy-preserving, and suitable for real-time monitoring of drug abuse signals.
Algorithmic fairness is becoming central to public discourse as beliefs about social justice increasingly come to bear on trust in technology. Using survey and behavioral data from multiple regions, this study employs structural equation modeling and mediation analysis to examine how social justice beliefs influence public acceptance of artificial intelligence through algorithmic fairness perception. Results show both direct and indirect effects, with mechanism strength varying during societal disruptions. The findings highlight the combined role of institutional context, trust, and literacy, offering insights beyond technical approaches to algorithmic fairness. The long-term dynamics involved in this issue warrant further study.
This study addresses the static assessment and non-adaptive communication in city image management by proposing an intelligent closed-loop framework. The authors extract dynamic features from multi-source data (social media, online travel agency platforms, government databases) to construct a real-time indicator system. A hybrid Long Short-Term Memory-Analytic Hierarchy Process model enables timely evaluation and short-term prediction of city image, combining temporal pattern learning with interpretable weighting. Based on assessment outputs, an adaptive communication module using reinforcement learning and collaborative filtering dynamically optimizes resource allocation and content personalization across digital channels. Experiments on real-world data show the approach increases tourism conversion revenue by 140%, improves audience targeting accuracy by 36%, and reduces strategy adjustment latency by over 42 hours versus conventional methods. The framework offers a scalable, data-driven paradigm for intelligent urban governance and smart city communication.
Pattern recognition is key to enhancing automated decision systems, particularly for image data analysis. As visual data grows in volume and complexity, robust models for accurate pattern interpretation are essential. This study examines multi-class image recognition via advanced architectures including CNNs, hybrid CNN-RNNs, and Vision Transformers (ViT), with comparative experiments. Performance metrics include accuracy, precision, recall, and F-score. The hybrid CNN-RNN model excelled, achieving 92.24% test accuracy by fusing spatial and temporal features. Precision and recall exceeded 90%, proving effectiveness across patterns. This work demonstrates how advanced methods boost classification accuracy, aiding smarter AI for real-world tasks.
To tackle both welding deformation in large mechanical components and the poor dynamic adaptability of traditional open-loop path planning, this study proposed a digital twin-driven Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning framework for online collaborative welding path optimization and deformation control. The framework integrated data from multiple sensing sources to build a high-dimensional state space and continuous 3D action space for welding parameters. A multi-objective reward function was designed, and the framework employed a finite, element–experiment hybrid training mechanism together with the TD3 algorithm to reduce the simulation-to-reality gap and mitigate action-value overestimation. Experiments on Q345 steel box girders showed that the framework reduced maximum angular deformation to 1.82 mm/m, increased the weld qualification rate to 96.2%, and improved both energy consumption and invalid path ratio. The system also demonstrated strong robustness under disturbances, while multi-field state fusion accelerated strategy convergence and enabled closed-loop intelligent welding for high-precision manufacturing.
In order to establish a reasonable path evaluation criterion in practice and reduce errors, this paper chooses an appropriate function, such as Manhattan distance or Chebyshev distance. A new heuristically adaptive path optimization strategy is proposed that combines swap and insert operations for better solutions diversity; it incorporates a bidirectional heuristic crossover operator to increase the path's quality and speed up convergence. In addition, a dynamic fitness selection function based on the number of iterations was introduced to improve the overall exploration-exploitation trade-off ability in all stages of the algorithm, which can reduce the occurrence of early stopping. Simulation experiments were performed in a complicated environment with 12 obstructions and 42 objective points; there were 20 repetitions. Compared with the results from sparrow search algorithm (SSA), grey wolf optimizer (GWO), and the traditional nutcracker optimization algorithm (NOA) algorithm, the proposed algorithm achieved an average reduction in the optimal path length of about 8.6%, around 14%, and more than half as much as a 38.73%.
Text classification and information retrieval remain challenging due to the heterogeneity of textual representations, ranging from sparse lexical features to dense semantic embeddings. This paper presents an adaptive hybrid-kernel multiple kernel learning support vector machine that integrates heterogeneous text representations within a unified and computationally efficient framework. The proposed model combines a linear kernel on term frequency–inverse document frequency word n-grams, a spectrum kernel applied to character n-grams, and a radial basis function kernel applied to sentence embeddings. Kernel contributions are adaptively learned under a simplex constraint with entropy-based regularization to prevent kernel dominance and ensure stable fusion. A focal-hinge loss with class-balanced weighting is incorporated to address class imbalance. Experiments on benchmark text classification and information retrieval datasets demonstrate consistent improvements over strong single-kernel and multiple kernel learning baselines, while kernel approximation techniques maintain scalability with limited performance degradation.
Consumer privacy protection demands are complex and multifaceted. Traditional incentive mechanisms struggle to balance participation enthusiasm with privacy risk control, leaving consumers exposed to unfair contribution evaluations and privacy leakage risks. To address this, this paper develops a collaborative joint learning incentive mechanism integrating graph neural networks (GNN) and multi-agent reinforcement learning. The approach first constructs node relationship graphs using GNN, then measures data distribution similarity through graph convolutional networks, and finally establishes a multi-agent reinforcement learning framework where nodes act as intelligent agents. By leveraging joint reinforcement learning and a dual-objective reward function, the mechanism optimizes strategies. Experimental results demonstrate that GNN-Shapley achieves over 97% accuracy, while the privacy compensation mechanism elevates average accuracy to 98.27%. This methodology effectively alleviates participation bottlenecks and safeguards consumer rights.
Traditional mental health assessments rely on questionnaires and interviews, which require manual data collection and cannot support real-time monitoring or early warning, especially for individuals with significant mood fluctuations. This paper proposes a model that combines sentiment analysis with LSTM networks to monitor emotional states in real time and capture temporal dependencies. The model collects and cleans user text data from multiple platforms and electronic health records, then applies natural language processing techniques such as word segmentation, vectorization, and sentiment analysis. Sentiment features are extracted using SentiWordNet and Bidirectional Encoder Representations from Transformers (BERT), and further classified with a convolutional neural network. The resulting sentiment scores are arranged chronologically and fed into the LSTM model to learn long-term patterns. After cross-validation and optimization, the model achieved 92% accuracy in predicting mental health risks, with an AUC between 0.9 and 0.95, demonstrating strong performance for real-time mental health risk prediction.
Despite the potential of artificial intelligence chatbots in overcoming the tedious tasks of research, scholars have considered critical failure factors before fully integrating such tool into the research process. While extensive works have comprehensively described these factors, none have explored their interrelationships in depth. Therefore, this paper applies interpretive structural modeling analyses to an actual case study of a university in the Philippines to understand the structural relationships among factors. It is found that the authors' lack of knowledge in the research field is the most influential factor. This implies that artificial intelligence chatbots must remain an auxiliary tool in research writing and authors must still possess the firsthand, necessary knowledge to serve as the main contributor of new knowledge in the field.
Despite the potential of artificial intelligence chatbots in overcoming the tedious tasks of research, scholars have considered critical failure factors before fully integrating such tool into the research process. While extensive works have comprehensively described these factors, none have explored their interrelationships in depth. Therefore, this paper applies interpretive structural modeling analyses to an actual case study of a university in the Philippines to understand the structural relationships among factors. It is found that the authors’ lack of knowledge in the research field is the most influential factor. This implies that artificial intelligence chatbots must remain an auxiliary tool in research writing and authors must still possess the firsthand, necessary knowledge to serve as the main contributor of new knowledge in the field.
Conventional sports monitoring methods are often hampered by inconvenient data collection and by accuracy issues. This paper explores the potential incorporation of artificial intelligence technology with wearable sensors to identify a low-cost, interpretable method of human motion monitoring. A single accelerometer is used to highlight applications of the rest of the method, which derives the vertical component and uses the pre-peak threshold, the pre-trough threshold, and the time lapse between the peak and trough of this component signal as feature values. The hardware components are simpler and fewer than in past applications, but better feature values are chosen to decrease the processing complexity for classification calculations and to enable the development of a decision tree algorithm to recognize and classify human motions. Together these approaches drastically reduce problems with inconvenience and reliability in comparison with conventional methods. The results show that the proposed method reaches a 93.89% recognition accuracy rate, which allows for accurate identification and classification of simple motion categories.
This paper designs an UAV image intelligent recognition system based on a deep learning algorithm, adopts a convolutional neural network as the core architecture, and selects the YOLOv11 (You Only Look Once version 11) model to realize rapid detection and precise classification of targets in images collected by UAVs. The system applies a model lightweight technology to ensure real-time performance and a transfer learning strategy. The model is pre-trained based on ImageNet and COCO (Common Objects in Context) large-scale datasets and is fine-tuned to adapt to UAV aerial image data, improving the model’s generalization ability and cross-scene adaptability under small sample conditions. Experimental results demonstrate that the system can still maintain high recognition accuracy and real-time performance in complex terrain and multi-class target environments. The average accuracy value is equal to 84.7% when IoU (Intersection over Union) ≥ 0.5, and the inference time is between 30-35 milliseconds. It has good practical value and promotion prospects.
A comprehensive bibliometric analysis of the research in deep learning applied to knee osteoarthritis (KOA) classification is presented. The Scopus database was analyzed from 2015 to 2025 for 3,199 articles. The purpose of this study has been to understand publication patterns and identify top contributors, thematic clusters, and emerging research areas in this area. Results indicate a surge in research activity, particularly over recent years as the number of publications rises dramatically. China, United States, and India were major leading countries in terms of the research output, while the University of California and University of Stanford were identified as major contributors. Co-word analysis highlighted four key thematic clusters. In this paper, the authors examine advancements in deep learning architectures, imaging modalities used for KOA diagnosis, machine learning interpretation techniques, and data preprocessing techniques.
Through the study of ancient murals, we can get valuable information about the lifestyle, hobbies, history and culture of the ancients. However, due to the influence of human and natural factors, murals often have problems such as cracks, fading and falling off. Traditional repair methods are not only time-consuming and laborious, but also may cause more serious secondary damage during the repair process. Therefore, this paper studies the application of multimedia feature extraction technology in the restoration of Tang tomb murals, aiming at assisting the restoration of murals. Therefore, a real-time feedback method for learning and monitoring of multimedia sensor networks based on linear shift adaptive optimization is proposed, and the optimal deployment model of nodes is constructed. The fusion tracking method of sensor nodes in the cluster is used for multimedia information fusion and two-dimensional entropy feature extraction.
Clothing detection is essential in computer vision applications like smart retail, virtual try-on, behavior analysis, and surveillance. However, challenges such as diverse clothing types, complex shapes, occlusions, and cluttered backgrounds hinder performance. To address these, the authors construct a diverse clothing detection dataset (DCDD) and propose the SCS-YOLO model, designed to improve multi-scale feature extraction and enhance key feature representation. The spatial depth convolution (SPD-Conv) module captures fine-grained details using spatial-to-depth and non-stride convolutions. The content-guided attention fusion (CGAF) module introduces channel and spatial attention for better robustness, while the squeeze and excitation attention (SEA) module adaptively weights critical features. Experiments on DCDD show that SCS-YOLO achieves 84.7% mAP, outperforming the baseline by 3.2%.
The original retinal images are often affected by uneven illumination, noise and other factors, resulting in poor classification accuracy. In this paper, the original retinal image is firstly enhanced by the improved Retinex algorithm, and then the global feature extraction is performed on the enhanced image by using the ResNet-12 model to obtain the global feature vector as the node features, and then the local features of different scales are dynamically adjusted by the multi-scale adaptive aggregation module to highlight the effective information of the nodes. A graph convolutional network is utilized to update the node features, and finally the aggregated node features are input to the fully connected level for classification prediction. Experimental results on two public datasets show that the offered model improves the classification accuracy by 3.12%–15.98%, and is able to more accurately classify retinal images of different quality levels.
