Journal of Hunan University Natural Sciences

Accurately distinguishing Indonesian banknote denominations remains challenging for older adults and people with visual impairments, especially after the recent release of a redesigned currency. To address this gap, we propose an intelligent recognition system based on deep transfer learning with systematic hyperparameter optimization. Our approach fine-tunes a pre-trained ResNet-50 backbone while simultaneously calibrating the learning rate, batch size, and training epochs via an am-optimized grid search. Extensive experiments across multiple convolutional neural-network architectures confirmed that the ResNet-50 model, trained with a learning rate of 0.0001, batch sizes 20 and 15 epochs, achieved state-of-the-art performance: 99.29 % accuracy, 99.32 % precision, 99.29 % recall, and 99.29 % F1-score. These findings demonstrate that carefully tuned transfer-learning pipelines can deliver near-perfect classification of the new Indonesian banknote series while retaining strong generalization to previously unseen images, thereby offering a practical assistive tool for the visually impaired and elderly.

Keywords: Deep transfer learning; ResNet-50; Convolutional neural networks; Indonesian banknote recognition; Hyper-parameter optimization; Assistive computer vision.

https://doi.org/10.55463/issn.1674-2974.52.5.19

This study examines subjectivity and polarity in multilingual WhatsApp messages exchanged during the employees’ union election at the Mehran University of Engineering and Technology (MUET), Khairpur campus. As social-media platforms generate vast volumes of unstructured multilingual text, sentiment analysis an intersection of natural language processing (NLP) and machine learning (ML) offers a rigorous means of extracting collective opinion. We introduce a novel analytical framework for code-switched discourse in Sindhi and Urdu, thereby contributing to both multilingual NLP research and electoral studies. A corpus of WhatsApp messages reflecting anticipated election outcomes was collected, cleaned, and annotated. Five supervised classifiers—feed-forward neural networks, nonlinear support vector machines, random forests, decision trees, and k-nearest neighbours were trained to detect sentiment polarity and subjectivity. Across all models, Candidate 1 received markedly more favourable sentiment, posting an average positive-sentiment score of 61.48 % versus 37.11 % for Candidate 2. Candidate 1 likewise exhibited superior aggregate performance metrics (mean accuracy = 89.92 %, recall = 86.52 %, precision = 85.12 %, F1-score = 85.81 %). These findings demonstrate the feasibility of fine-grained sentiment mining in under-resourced languages and yield actionable insights for campaign strategists. Future research will enhance classification accuracy through advanced deep-learning architectures and will track dynamic sentiment patterns across multiple social-media platforms throughout the electoral cycle.

Keywords: multilingual sentiment analysis; code-switching; WhatsApp; natural language processing; machine learning; Sindhi; Urdu; electoral behaviour.

https://doi.org/10.55463/issn.1674-2974.52.5.18

The purpose of the article is to develop and validate a hybrid biomedical signal monitoring system that integrates edge-computing capabilities with Microsoft Azure cloud services. The article describes a new edge-and-cloud architecture based on the EmotiBit wearable sensor and a Raspberry Pi gateway, enabling continuous acquisition, local buffering, and scalable cloud synchronization of multi-modal biometric signals. Using hardware evaluation (signal accuracy, power consumption, wireless connectivity), Azure Blob Storage, Azure SQL Database, Power BI dashboards, and Azure Stream Analytics, the authors demonstrate a seamless pipeline for real-time visualization and post-hoc analysis of electrocardiogram (ECG), photoplethysmogram (PPG), galvanic skin response (GSR), and additional vital signs. We illustrate the proposed system by conducting a rigorous battery performance analysis under three workloads continuous sensing alone, sensing with local storage, and combined local plus cloud upload and comparing operational endurance across battery capacities ranging from 1,200 mAh (≈12 h runtime) to 6,000 mAh (≈45 h runtime).
Our proposal allows improvement in data delivery reliability to ≥99.5 % under network latencies up to 200 ms and reduces baseline clinical monitoring load by up to 30 %. The new method for performance evaluation is confirmed by quantitative runtime and reliability calculations. New research results supplement existing telemedicine paradigms by demonstrating enhanced patient mobility, near-real-time anomaly detection, and scalable data management, and can be used for remote patient monitoring, clinical diagnostics, and large-scale health studies. This paper is novel because it jointly optimizes edge-device design, battery endurance, and cloud analytics within a unified, deployable Azure framework.

Keywords: EmotiBit; Raspberry Pi; Microsoft Azure; Edge Computing; Internet of Things (IoT); Real-time Monitoring; Telemedicine.

https://doi.org/10.55463/issn.1674-2974.52.5.17

The Sine–Cosine method has emerged as a robust analytical approach to derive solitary wave solutions for nonlinear dispersive partial differential equations. In this work, we systematically employ this technique to the Camassa–Holm hierarchy, encompassing the classical Camassa–Holm, Degasperis–Procesi, Fornberg–Whitham, and Fuchssteiner–Fokas–Camassa–Holm equations. Each member of the hierarchy models shallow-water wave propagation under specific integrability conditions, exhibiting rich dynamical behavior. By applying an appropriate wave transformation, we reduce the governing equations to ordinary differential equations and construct exact travelling-wave solutions in terms of trigonometric and hyperbolic functions. The solutions obtained include compacton like profiles and classical sech² and cosh² structures, with explicit expressions for wave speed and amplitude as functions of model parameters. Comparative analysis highlights the effectiveness and simplicity of the Sine–Cosine method relative to more elaborate techniques such as the Hirota bilinear formalism and the inverse scattering transform. Our contributions lie in the unified application of this method across the entire hierarchy and the presentation of a comprehensive classification of solitary wave families.
These results extend existing literature and provide valuable benchmarks for numerical simulations and theoretical investigations of nonlinear wave dynamics in fluid mechanics and related fields. Studies illustrate the influence of nonlinearity and dispersion coefficients on wave morphology.

Keywords: Camassa–Holm hierarchy; Sine–Cosine method; solitary wave solutions; nonlinear dispersive equations; integrable systems; exact analytical solutions.

https://doi.org/10.55463/issn.1674-2974.52.5.16

The purpose of this article is to propose a novel machine learning–enhanced data-driven model predictive control (DD-ML-MPC) to improve robustness and performance under uncertainty. The article describes a DD-ML-MPC method, based on neural-network predictors trained on historical I/O data and integrated within an MPC framework, enabling more accurate forecasts and tighter constraint satisfaction when plant models are unavailable. Using numerical simulations on a MIMO LTI system corrupted by Gaussian noise, the authors compare three strategies—Model-Based MPC (MB-MPC), Data-Driven MPC (DD-MPC) and DD-ML-MPC—by assessing root mean square error, convergence time, constraint violation rate and computational overhead. The proposed DD-ML-MPC is demonstrated on a four-state, two-input, two-output plant subjected to stochastic disturbances. Our DD-ML-MPC achieves a 20 % reduction in constraint violations and a 15 % decrease in tracking error relative to MB-MPC.
The evaluation methodology is validated through Monte Carlo analysis. The evaluation criteria also include stability margins and computational efficiency under real-time constraints. New research results bridge model-based and data-driven paradigms, offering practical applications in autonomous vehicles, process control and robotics. Results demonstrate the potential for field deployment and suggest future extensions to nonlinear and time-varying systems. This paper is novel because it introduces a unified control paradigm that integrates deep learning predictors into MPC, yielding robustness and scalability improvements.

Keywords: Model Predictive Control (MPC); Data-Driven Control; Machine Learning Control; Neural Network.

https://doi.org/10.55463/issn.1674-2974.52.5.15