Gongcheng Kexue Yu Jishu/Advanced Engineering Science

Abstract : Soil used in earthwork is typically rated as “excellent-to-good” when it effectively supports building foundations, embankments, pavements, and other civil infrastructure. Expansive clays are problematic soils due to their tendency to swell and shrink under varying moisture “conditions”, and represent a significant challenge in geotechnical engineering, often resulting in considerable economic losses for construction projects. These soils exhibit high vulnerability to failure mechanisms, including excessive settlement and subgrade instability, which exacerbate their performance issues compared to other soil types. Given the scarcity of suitable construction sites, improving the engineering properties of these soils is imperative for their application in construction projects. Stabilization techniques play a crucial role in enhancing the characteristics of native or granular soils used in pavement layers. This study presents an experimental investigation into the combined effects of admixtures, specifically RoadCem and cement, on the geotechnical properties of expansive soils. The findings revealed that the incorporation of these stabilizers resulted in an increase in optimum moisture content (OMC), a decrease in maximum dry density (MDD), and significant improvements in unconfined compressive strength (UCS) and California Bearing Ratio (CBR)..
Full article

Abstract : Across all industries, digital service platforms have a significant problem from customer attrition. In order to improve churn prediction accuracy and cross-domain generalization, this paper suggests a cross-domain deep learning architecture that makes use of transfer learning. Through knowledge transfer across several industries, including food delivery, streaming, and mobility services, the suggested method enhances predictive performance in areas with less labeled data, allowing for more resilient and flexible churn control solutions. The first dataset used to train a Fully Connected Neural Network (FCNN) had 388 samples and 55 characteristics related to food delivery. The trained model was then applied to two different domains to assess its generalization ability: Uber (50,000 samples, 14 features) and Netflix (1,000 samples, 26 features). Training was conducted using the Synthetic Minority Oversampling Technique (SMOTE) to rectify the imbalance in classes. Significant performance improvements across domains were shown by the transferred model. With a churn recall of 0.79 and an accuracy of 66.73% on the Uber dataset, it outperformed XGBoost and CatBoost by 16.2% and 41.1%, respectively. The recall for Netflix was 1.5% better than CatBoost and 23.2% better than XGBoost. With a churn recall of 0.90 in the source domain (food delivery), the model outperformed XGBoost and CatBoost by 15.4% and 5.9%, respectively. In target domains, the suggested FCNN with transfer learning improved churn recall by up to 10.1%, consistently outperforming both baseline and hybrid models. Because it allows for the early identification of at-risk clients utilizing information transferred from related domains, this method is especially advantageous for sectors with minimal labeled data. In society's expanding digital economy, the concept helps to improve service continuity and minimize corporate losses by increasing proactive client retention..
Full article

Abstract : Maintaining nutrition, increasing yield, and minimizing an overabundance all depend on prompt identification and accurate classification of pests and diseases in tropical fruit crops. However, current deep learning approaches frequently have significant computing expenses and little flexibility to actual farming operations. In order to accomplish excellent classification precision with considerably fewer processing necessities, this study proposes ESA-ResNet34-Lite+, a novel lightweight, attention-enhanced deep learning framework that can be used in mobile and Internet of Things-based farming systems. Optimizing extracting features and concentrating on disease-specific behaviours while lowering the complexity of the model is achieved by integrating separable convolutions by depth and Efficient Spatial Attention (ESA) modules within a modified ResNet34 foundation. Furthermore, greater levels utilize Convolutional Block Attention Modules (CBAM) to improve multi-scale visualization of features and accuracy in classification. To train and validate the approach, a carefully selected dataset of roughly 3,500 tagged photos of tropical fruit crops including important guava diseases like Phytophthora, Scab, and Styler was utilized. Investigations show that ESA-ResNet34-Lite+ outperforms standard models like VGG16, MobileNet, and ResNet34 by 2–4% in classification accuracy, with an overall accuracy of 95.7%, precision, recall, and F1-scores of 95.6%, 95.8%, and 95.7%, respectively. The model also achieves a 70% reduction in FLOPs and an 85% reduction in specifications when compared to traditional architectures, highlighting its effectiveness and appropriateness for real-time field implementation. These results demonstrate that ESA-ResNet34-Lite+ offers an adaptable and economical intelligent agricultural approach by striking a good balance between accuracy and computational cost. Through the demonstration of exceptional detection performance on a difficult tropical crop dataset, this study creates a useful and trustworthy foundation for precision gardening, facilitating prompt disease control and encouraging ethical farming practices..
Full article

Abstract : Effectively processing range queries for big, distributed data sets remains a perennial problem throughout today’s data systems, when queries are multifeatured in particular. Traditional indexes, such as B-Trees, KD-Trees, and R-Trees, often don’t work well in distributed or high-dimensional environments due to scalability and integrative limitations. This research paper proposes AVL Tree based Multi-Feature Query Engine framework for optimal multi-feature range queries. Maintaining logarithmic time per query when coupled with collection size, AVL trees represent a highly efficient indexing model, most prominently when resultant sets are modest in size. It’s deployed in Python and describes building local AVL trees on partitioned data and mapping it to a distributed environment utilizing MapReduce’s Hadoop implementation. It suggests a highly efficient filtering of big collections by performing range criteria in a staged mode across features, which severely reduces execution time when compared to linear scans. It describes the usability of balancing trees within distributed querying systems and spans scalability between in-memory indexes and big data systems..
Full article

Abstract : Digital Twin (DT) technology has rapidly evolved into a transformative innovation bridging the physical and virtual worlds through real-time data integration and simulation. Initially applied in aerospace and manufacturing, DTs are now extensively employed in power electronic systems (PESs) to enhance modeling accuracy, predictive maintenance, and operational efficiency. This study explores the implementation of DTs within power electronics, emphasizing their role in simulation, health monitoring, and performance optimization. The findings reinforce the potential of DT-based frameworks in achieving high-fidelity virtual representations that enable intelligent control, fault prediction, and optimized performance in power electronics applications..
Full article