Currently, 35.2% of mothers suffer from hand and wrist pain after giving birth. Their physical problems are often related to the posture in which they carry their baby. Therefore, correct posture when carrying the baby is important to avoid postpartum physical problems such as tendonitis. However, determining the infant-carrying posture requires feedback from an expert, (e.g., a midwife), and is time-consuming. To overcome this problem, an infant-carrying posture determination (ICPD) method is proposed using an RGB camera and the BlazePose pose estimation model. With the ICPD method, a person carrying an infant can easily determine the quality of their posture when carrying an infant. To achieve a more accurate determination, the ICPD method normalizes the infant-carrying posture and selects features based on the area under the receiver operating characteristic curve, which is a widely used performance measure in classification models. The postures of 28 mothers while carrying infants was experimentally determined to validate the proposed system. The experimental results confirmed that ICPD was more accurate on the test dataset than conventional methods, both with and without feature selection.

In view of the inadequacy of interactivity and ecological aesthetics in park landscape, the existing methods are often difficult to achieve accurate terrain analysis and design, resulting in unsatisfactory landscape effects. In order to make up for this shortcoming, the research innovatively integrates sparrow search algorithm (SSA) and self-organizing mapping algorithm (SOM) to conduct in-depth analysis of park landscape topography. SSA, with its powerful global search ability, can quickly lock the optimal solution region. SOM uses unsupervised learning to effectively mine the underlying rules in terrain data. The combination of the two not only improves the accuracy and efficiency of terrain analysis, but also lays a solid foundation for the construction of park landscape environment interaction design model. The root-mean-square error of the algorithm is only 0.01 m², the accuracy is as high as 98.1%, and the F1-value is also as high as 97.8%. Compared with previous studies, the algorithm has better fitting performance. In addition, the model is applied to the actual park landscape design, and an interactive park landscape optimization design scheme is proposed based on the design concept of forest walk and ecological protection area. After evaluation, the program scored 9.1 points for interactivity and 9.2 points for sustainability, fully proving that it meets the interactive needs of tourists while also taking into account ecological protection and sustainability. Compared with previous studies, the innovation of this study lies in the successful integration of SSA and SOM algorithms, which significantly improves the accuracy of terrain analysis, and verifies the practicability and effectiveness of this method through practical application. Overall, the method proposed in the study can analyze topography more accurately than traditional methods and design park landscapes that are both interactive and aesthetically pleasing.

This paper addresses the optimal consensus problem in uncertain switching multiagent systems. The inherent uncertainty and time-varying structure of local tracking error system render conventional methods ineffective for deriving optimal control protocols. To overcome these challenges, we introduce a reference model for each agent and construct a modified augmented local tracking error (ALTE) system. This approach transforms the optimal consensus problem into two sub-problems: 1) model reference control (MRC) between agents and their reference models; 2) distributed optimal stabilization of the modified ALTE system. We propose a new control scheme that combines filtered tracking error with equivalent input disturbance method to achieve MRC. To realize distributed optimal stabilization of the modified ALTE, we introduce a deep deterministic policy gradient method based on value iteration. Through theoretical analysis, we demonstrate that the multiagent system achieves a near Nash equilibrium, which is further validated by numerical simulation.

Existing aspect-based sentiment analysis (ABSA) methods do not sufficiently enhance multiple subtasks with syntactic knowledge in a joint framework. In this paper, we propose an ABSA method that utilizes a multitask learning framework to enhance syntactic knowledge fully. The method first builds on a dependency relation embedded graph convolutional network to learn syntactic dependencies and the dependency types between words in a sentence fully. Second, to make better use of the syntactic information between aspect and opinion words, we extend the adjacency matrix based on dependency parsing to establish the direct relationship between aspect and opinion words. Finally, an information passing mechanism is exploited to ensure that our model learns from multiple related tasks in a multitask learning framework. The results of experiments on three public datasets, namely LAP14, REST14, and REST15, show that the proposed method has better performance than the baseline method.

This work introduces an innovative machine learning algorithm based on the minimalist machine learning paradigm, called matrix transformations bootstrap. Evaluated on 15 medical datasets, ranging from 3 to 1626 attributes, the methodology incorporates matrix transformations like rotation and shearing to improve dataset separation in binary classification tasks. Additionally, random feature selection is applied via the bootstrap method, resulting in two new attributes that can be visualized on the Cartesian plane while achieving substantial dimensionality reduction. The results show significant classification performance improvements over traditional algorithms like k-NN, SVM, Bayesian models, ensembles, neural networks, and logistic functions, evaluated using balanced accuracy, recall, and F1-score.

With the advances in artificial intelligence and computers, sensor-based human behavior recognition technology has been gradually applied to many emerging cross-cutting fields such as smart healthcare and motion monitoring. First, we design a deep learning model for human behavior recognition single mode based on lifting wavelet transform (lifting scheme convolutional neural networks-gated recurrent unit, LSCG) to address the problem of inaccurate and insufficient feature extraction from sensor data in the human behavior recognition network model. The structure of the LSCG network model consists of a wavelet decomposition module and a feature fusion module. Then, we further address the limited ability of a single modality for human behavior recognition by designing a multimodal human behavior recognition model based on the LSCG model (multimodal lifting scheme convolutional neural networks gated recurrent unit, MultiLSCG). The structure of the MultiLSCG network model consists of a feature extraction module and a multimodal feature fusion module. The feature extraction module consists of the LSCG model, which enables the model to extract features from different modal human behavior data. The multimodal feature fusion module enables the model to obtain more features from the multimodal behavior signals by extracting the global feature information of the human behavior signals and the local feature information of the human behavior signals. Finally, the experimental results show that in the public dataset OPPORTUNITY, the accuracy of the motion pattern dataset reaches 91.58%, and the accuracy of the gesture recognition dataset reaches 88.53%, which is higher than the existing mainstream neural networks, on the UCI-HAR and WISDM data sets, the accuracy of our proposed model reached 96.38% and 97.48%, which further verified the validity and applicability of our proposed model.

Drowsy driving is a major contributor to traffic accidents, making real-time monitoring of driver drowsiness essential for effective preventive measures. This paper presents a novel method for detecting driver drowsiness through facial video analysis and non-contact heart rate measurement. To address the challenges posed by varying lighting conditions, the algorithm integrates RGB (red, green, and blue) and multi-scale reinforced image color space techniques. This combination enhances the robustness of heart rate signal extraction by generating spatio-temporal maps that minimize the impact of low light. A convolutional neural network is used to accurately map these spatio-temporal features to their corresponding heart rate values. To provide a comprehensive assessment of drowsiness, a differential thresholding method is utilized to extract heart rate variability information. Building on this data, a dynamic drowsiness assessment model is developed using long short-term memory networks. Evaluation results on the corresponding dataset demonstrate a high accuracy rate of 95.1%, underscoring the method’s robustness, which means it can greatly enhance the reliability of drowsiness detection systems, ultimately contributing to a reduction in traffic accidents caused by driver fatigue.

Sampled-data control for dynamic programming of continuous-time system, which can facilitate to implement control actions under networked environments, is rarely considered in most existing works. In order to address this issue, an event-triggered dynamic programming sampling control (ET-DPSC) approach is investigated for networked path-following control of autonomous vehicles. The first goal is to establish the sampled-data-based event-triggered path-following control model using Hamilton–Jacobi–Bellman equation. Secondly, the asymptotic stability criterion in an input-to-state sense should be tackled by exploiting Lyapunov theory and input delay approach. As a third goal, the sampled-data controller based on dynamic programming method should be synthesized. Compared to most existing ADP-based control strategies, the proposed ET-DPSC approach not only guarantees the stability of path-following control but also provides significant benefits for control implementation under communication-constrained environments. In addition, Zeno behavior is naturally excluded by using periodic discrete-time sampling control fashion. At last, Simulink and CarSim joint simulations are conducted to show effectiveness of the proposed ET-DPSC scheme by comparing with linear quadratic regulation without considering input delay.

Impacted third molar extraction, particularly of mandibular teeth, is a common procedure performed to alleviate pain, infection, and misalignment. Accurate diagnosis and classification of impaction types are crucial for effective treatment planning. This study introduces a novel algorithm for automatically measuring the impaction angles of mandibular third molars (T32 and T17) from orthopantomogram (OPG) images. The proposed method is based on deep learning techniques, including segmentation and key point detection models. It categorizes impactions into Winter’s classification: distoangular, mesioangular, horizontal, vertical, and other on both sides, using the measured angles. The proposed method used 450 OPGs, achieving high mandibular molar segmentation accuracy with dice similarity coefficients (DSC) values of 0.9058–0.9162 and intersection over union (IOU) scores of 0.82–0.84. The object keypoint similarity (OKS) for detecting the four corner points of each molar was 0.82. Angle measurement analysis showed 80% accuracy within ±5° deviation for distoangular impaction of T32 and within ±8° for T17. The F1-scores for mesioangular classifications were 0.88 for T32 and 0.91 for T17, with varying performance in other categories. Nonetheless, the predicted angles aid in identifying impaction types, showcasing the method’s potential to enhance dental diagnostics and treatment planning.

This paper proposes a remote learning monitoring method based on learning behavior time series data to effectively monitor learning progress of students. This method integrates multi-scale feature extraction, a variational information bottleneck module, and a variational autoencoder to enhance feature diversity and clustering performance. Tests indicate that the proposed multi-scale full convolution algorithm model achieves a Precision of 0.887, an F1 score of 0.922, an area under the curve of 0.883, and a Recall of 0.960, outperforming benchmark algorithms such as Naive Bayes and chaotic lightning search algorithms in leak prediction. The improved unsupervised algorithm achieves a Precision of 0.888, a Recall of 0.944, an F1 score of 0.915, and an Accuracy of 0.861, surpassing benchmark algorithms. This study offers a high-precision solution for remote learning monitoring, which holds practical value in enhancing teaching quality, addressing learning challenges of students, and providing theoretical support for optimizing the learning environment. Future research will focus on further optimizing algorithm models.

After learning, the object-detection algorithm can automatically detect whether the riders of electric mopeds are wearing helmets, thereby saving regulatory labor costs. However, the complex environmental background and headwear similar to helmets can easily cause a large number of false negatives and false positives, increasing the difficulty of detection. This paper proposes the YOLOv8n-Improved object-detection algorithm. First, in the neck part, the algorithm uses a simplified weighted bi-directional feature pyramid network structure to remove single input nodes, add connection edges, and attach path weights according to the importance of features. This structure optimizes the algorithm’s multiscale feature-fusion capability while improving computational efficiency. In the head part, the algorithm uses the scale-sensitive intersection over union loss function to introduce the vector angle between the predicted and ground-truth boxes, redefining the penalty metric. This improvement speeds up the convergence process of the network and improves the accuracy of the model. After comparative validation on the test set, the YOLOv8n-Improved algorithm shows a 1.37% and 3.16% increase in the average precision (AP) metric for electric moped and helmet detection, respectively, and a 2.27% increase in the overall mean AP metric, with a reduction in both false negatives and false positives for the two categories.

Power system data possess many characteristics and indicators, having certain high dimensions and redundant information, which can easily increase the calculation and storage overhead. To reduce the dimension of power data, eliminate redundant information, and reduce the delay time, a data clustering algorithm is proposed. Firstly, an algorithm based on PCA and kernel local Fisher identification is used to reduce the dimension of large multidimensional samples and enhance the accuracy of subsequent clustering. Thereafter, the redundant data are processed after dimension reduction is processed to optimize the data quality by introducing a bloom filter structure. In the graph model, data clustering is completed based on the parallel processing of redundant data. Simulation results show that the correctness and stability of this method are over 85%, and the delay time is decreased, representing good application prospects.

This study proposes a fuzzy clustering algorithm based on fuzzy classification maximum likelihood, t-distribution, and Tsallis entropy regularization. The proposed algorithm is shown to be a generalization of the two conventional algorithms, not only in the use of their objective functions, but also at their algorithmic level. The robustness of the proposed algorithm to outliers was confirmed in numerical experiments using an artificial dataset. In addition, experiments using 11 real datasets demonstrated the superiority of proposed algorithm in terms of the clustering accuracy.

Measuring agreement among participants in group decision-making problems is critical to such processes. This paper introduces a novel consensus index derived from the Gini coefficient, which avoids the need for traditional aggregation matrices, simplifying calculations while maintaining robustness. The proposed Gini Consensus Index demonstrates properties of reciprocity and boundedness, making it a reliable alternative to traditional distance-based measures. Through a comparative statistical analysis using the Wilcoxon test, the GCI performed similarly to established methods but with computational advantages and enhanced stability. These features make it a promising tool for consensus evaluation in fuzzy preference frameworks.

With the reform of teaching methods, hybrid online and offline teaching modes have been used increasingly in college courses. In this setting, the factors affecting academic performance are more complex, making it more challenging to predict students’ performance. Therefore, there is an urgent need for higher-performance prediction algorithms. This study briefly analyzed college students’ learning in ideological and political courses. Then, the learning features of college students in the courses were extracted using the Super Star platform and teaching system. Feature selection was carried out based on the information gain rate, while the training set was balanced using the synthetic minority oversampling technique (SMOTE). Moreover, the seagull optimization algorithm (SOA) was applied to optimize the hyperparameters of eXtreme Gradient Boosting (XGBoost) to develop the SOA-XGBoost algorithm for early warning of performance. Experiments were performed on the collected datasets. It was found that the effect of the SOA-XGBoost algorithm on the early warning of performance improved significantly following SMOTE processing. The F1-value reached 0.955 and the area under the curve value was 0.976. The SOA exhibited superior performance in hyperparameter optimization compared with other algorithms such as the grid search. The SOA-XGBoost algorithm also achieved the best results in early warning of performance. These results confirm the effectiveness of the proposed SOA-XGBoost algorithm for early warning of performance, and the method can be widely applied in practice.

This paper proposes a method for generating comic style Chernoff face with generative adversarial network (GAN) as a first step towards the generation of data comics from multi-dimensional data. The proposed method converts Chernoff face into comic style face images based on the combination of CycleGAN and Pix2Pix. Since both Chernoff face graph and comic images do not have enough information for direct conversion, the Chernoff face graphs are converted into photo style face images and then converted into comic images. A questionnaire asking to rank face images according to the specified impressions is conducted to evaluate the proposed method. The result of the questionnaire shows that the proposed method achieved the same level of consistency among answerers’ judgments as original Chernoff face. It is also confirmed that the proposed method can express the difference in attribute values with mouth parts.

To address the need for enhanced educational quality and effective student management, this study introduces a novel model for predicting student engagement and delivering personalized recommendations by integrating a GRU-Attention network with an L-DMF recommendation algorithm. Our approach employs a GRU-Attention network to analyze student behavior data and accurately predict engagement levels. The attention mechanism enhances the model’s ability to prioritize significant features, resulting in an impressive prediction accuracy of 98.15%, surpassing traditional classification methods such as decision trees, support vector machines, and random forests. In addition, the author proposes an L-DMF-based recommendation system that utilizes student behavior data to generate tailored suggestions. The model’s performance was compared with leading recommendation algorithms, including LibFM, KGCN, and DRER. The results demonstrate that our approach provides more accurate and contextually relevant recommendations. By effectively incorporating both spatial and temporal features of student behavior, our model achieves superior results in both engagement prediction and recommendation tasks. Overall, the dual focus on precise engagement forecasting and personalized recommendation highlights the model’s efficacy in enhancing educational management and student support.

This study aims to explore tourists’ sentiment tendencies and focal points by analyzing online reviews of 5A-level tourist attractions. After conducting data cleaning, word segmentation, stop-word filtering, and part-of-speech tagging, we preprocessed the review texts and utilized the ROSTCM6 software for sentiment analysis. The study found that most tourists hold a positive attitude toward their experiences at 5A-level attractions, though there remains room for improvement in certain facilities and services. This research provides valuable feedback for attraction managers to enhance the visitor experience.

5G mobile communication technology can satisfy the needs of network quality of service (QoS) for vehicle-to-everything (V2X) in ideal conditions. However, complex intelligent transportation scenarios may lead to fluctuations in 5G QoS, resulting in passive and lagging degradation of the service level of V2X services. To address the challenge of aligning service requirements with network conditions, it is crucial to explore schemes for predicting and managing QoS fluctuations. This paper proposes a vehicle route planning scheme to improve the quality of experience for V2X services by QoS prediction based on probability distribution detection (PDD). We design a distribution detection algorithm to tackle the issue of improving QoS prediction accuracy by calculating probability confidence weights of the outcome of two different QoS prediction models. Simulation evaluations show that the proposed PDD-based prediction method significantly enhances the accuracy of predictions. We have achieved 0.128 mean absolute error, with 0.189 root mean square error, in predicting the network throughput. Furthermore, in comparison to the routes selected by the length-based route planning scheme, the proposed route planning strategy can enhance the network throughput by at least 5.3 kbps.

The current fire-detection methods rely primarily on smoke and temperature detection, which are generally performed in the late stage of fire and thus cannot provide a timely reminder in the early stage of fire. The continuous development of artificial intelligence has enabled machine-vision fire detection. This study proposes a convolutional neural network target-detection algorithm, i.e., You Only Look Once version 4 (YOLOv4), to detect small targets. It offers outstanding characteristics and enables scenic-spot monitoring via the video extraction of real-time fire detection using a significant amount of fire data. The diverse fire scenes can provide accurate and timely detection in the early stage of fire, thus providing favorable early warning and alarm function.