To investigate the evolution of concrete compressive strength in a Na2SO4-NaHCO3 composite salt - flowing water environment typical of karst tunnel groundwater, accelerated erosion tests were conducted based on the representative ionic system of Guangxi. The influences of sulfate concentration, carbonate concentration, seepage velocity, and exposure age were examined, and a BP neural network model was developed using these environmental variables as inputs. Results show a unimodal two-stage behavior of 'initial enhancement followed by deterioration.' Across 18 conditions, peak strength ranged from 46.2-54.2 MPa (1.2%-13.6% higher than 30 d), with 7.8%-19.9% reduction by 90 d. Increased flow velocity advanced the peak age by about 15 d and accelerated post-peak degradation. Under high concentration and velocity (6% + 18%, 0.3 m/s), strength declined from 49.9 MPa to 40.1 MPa, revealing a strong concentration - velocity coupling effect. The BP model achieved high correlation (R2 = 0.836 at 30 d; 0.958 at 90 d) with stable and accurate predictions, providing a practical tool for durability assessment of tunnel linings.
This work presents the design, fabrication, experimental testing, and numerical analysis of a 15-cm bio-inspired robotic fish that swims by a median-paired-fin (MPF) traveling-wave mechanism. A twin-crankshaft/three-rocker transmission, fabricated entirely by fused-filament printing (PLA+ fin skeleton) and cast silicone membranes, drives the fin with only two low-cost brushed DC motors. A simplified moving-mesh model implemented in COMSOL Multiphysics (R) couples a piecewise kinematic input to an extremely fine, dynamically remeshed domain and reproduces the fluid-structure interaction in 60 min of CPU time per case. Experiments conducted in a 1 m & times; 0.6 m & times; 0.4 m tank show that increasing the supply voltage from 3 V to 5 V raises the mean swimming speed from 1.44 cm s-1 to 1.93 cm s-1; the model predicts these speeds within +/- 5%. Flow-field visualizations reveal that the performance gain is accompanied by larger vortex structures and higher turbulent kinetic energy, identifying a trade-off between maximum speed and directional stability.
This research explores regression modeling for interval-valued data by combining the Bivariate Center and Range Model (BCRM) with Ordinary Least Squares (OLS), Principal Component Regression (PCR), and Partial Least Squares Regression (PLSR), respectively. Using the center-range representation for interval-valued variables expands the predictor space, leading to multicollinearity that may undermine model stability and interpretability. To address this, PCR and PLSR are integrated into the BCRM framework to reduce dimensionality through latent components while maintaining the inherent structure of interval-valued variables. The proposed methods are tested on three real-world datasets: Taiwanese temperature records, exchange rates, and air quality data. Results indicate that the Bivariate Center and Range Model with OLS provides a highly accurate baseline, with average RMSE values of 0.8111, 0.7678, and 22.7234 for the respective datasets. Compared to BCRM using OLS, the Bivariate Center and Range Model using the PLSR model achieves a better balance between prediction accuracy and model complexity, reducing the number of explanatory variables by about 25-30% while keeping prediction errors close to those of the BCRM-OLS baseline. These results suggest that the proposed approach offers an effective and reliable framework for regression analysis involving high-dimensional interval-valued data.
Alzheimer's disease (AD) is a common neurological condition that causes brain cells to atrophy, progressively impairing a person's ability to think and react independently. The possibility of reducing the growth of this mental illness increases with early identification. MRI is a key biomarker used for measuring disease progression and allows for noninvasive longitudinal assessments. The proposed work suggests a deep learning (DL) based lightweight CNN approach for early AD identification. The suggested model was assessed and validated using the ADNI dataset, which includes three different types of MRI, such as AD, MCI, and CN. Further, the proposed scheme consists of a comparison of distinct machine learning and transfer learning models, which include Alexnet, DenseNet201, GoogleNet, InceptionV3, ResNet18, ResNet50, ResNet101, SqueezeNet, VGG-16, VGG-19, and Xception. The lightweight CNN for the multiclass classification of AD MRI images proposes the best pre-trained model to accurately predict the patient's stage. It is observed that the best accuracy (99%) was attained utilizing a lightweight CNN, whereas machine learning (98%) and transfer learning (96%) achieved lower results. Our model outperforms previous studies, reducing doctors' burden and speeding decisions, but reliance on the ADNI dataset highlights the need for future research using diverse, generalizable datasets.
This study develops a finite element framework that couples nonlinear Stokes flow with a basal sliding-driven erosion law to reconstruct the evolution of the glacier bed in a subtropical, low-latitude alpine cirque. Applied to the well-preserved Cirque No. 2 on Taiwan's Xue (Syue) Mountain, this model provides the first physics-based quantitative constraints on glacier erosion in a low-latitude environment. Quasi-steady Stokes-sliding simulations over a 1,000-year interval reproduce the downward migration of the ice-bed interface and yield erosion velocities consistent with empirical estimates. By comparing the simulated bed evolution with the present cirque topography derived from publicly available contour data, we identify parameter combinations that are consistent with the similar to 9,000-year glaciation interval inferred for the Last Glacial Maximum. The results demonstrate how sliding efficiency, erosion constants, and ice geometry jointly regulate cirque erosion, offering a quantitative basis for reconstructing alpine glacier-bed evolution in data-sparse mountain regions.
Considering the bulldozer AT gear design method is rare in literature and time-consuming, a fast topology and parameter synthesis method for design of multi-speed ATs with the aid of FGE-based graph library is introduced in this paper. First, all non-isomorphic topology graphs of PGTs with multi-DOF are enumerated. Second, all gears of each PGTs with given number of brakes and clutches are synthesized and the PGTs is transformed into 1-DOF PGTs at different gears in order to derive its gear ratios based on the FGE composition process. A graph library is built based on the FGE composition process. Third, numerous gear ratios of each PGTs are synthesized quickly by identifying their isomorphic structures in the graph library and matching their nodes with those of the isomorphic structures. Compared with traditional method, the calculation time has been greatly reduced and the design efficiency has been improved. Finally, the rational topology schemes are sorted out by completely searching all possible combinations of gear sequences and optimizing their characteristic parameters with NSGA-II. The efficacy and feasibility of the method are verified by comparing the innovative 6-speed bulldozer AT scheme with existing Komatsu scheme.
This paper presents a multimodal emotion recognition framework that integrates enhanced ShuffleNet V2 with an efficient channel attention (ECA) mechanism to mitigate modal redundancy and semantic misalignment. For video modality, an improved ShuffleNet V2 extracts spatial features, while for audio modality, a CNN-BiLSTM network coupled with ECA modules captures temporal dependencies and reduces feature redundancy. A simple feature concatenation followed by fully connected layers is employed for cross-modal semantic alignment. Evaluated on the CREMA-D dataset, the proposed framework achieves a mean classification accuracy of 77.73% +/- 1.19%, with particularly high recognition rates for anger (86.1%) and happiness (83.3%). Cross-dataset validation on IEMOCAP Session 1 yields an accuracy of 73.26%, confirming generalization capability. Moreover, the model attains 84 frames per second with only 2.74 million parameters, demonstrating superior computational efficiency and robustness. These results indicate the framework's practical potential for real-time human-computer interaction and mental health monitoring.
