Ngoc-Dau Mai

Ensuring meat freshness is critical for food safety, quality assurance, and supply-chain sustainability. Conventional inspection methods are labor-intensive, time-consuming, and susceptible to human error. To address these limitations, this study proposes a non-invasive, battery-free system for real-time meat freshness monitoring that integrates far-field radio-frequency (RF) energy harvesting with deep learning. The system employs a compact conformal meandered antenna as the RF energy harvester, providing structural flexibility and stable performance, with a measured gain of 2.31 dBi at 915 MHz. Experimental characterization demonstrates a maximum RF-to-DC power conversion efficiency of 54.16% at 10 dBm input power, generating up to 3.13 V across the load. To enhance model robustness, data augmentation and optimization are performed using a sequential Variational Autoencoder–Bidirectional Long Short-Term Memory (VAE–BiLSTM) framework. Multivariate gas-sensing data, including temperature, equivalent carbon dioxide, and total volatile organic compounds, are collected from pork, chicken, and fish samples over five days under both ambient and refrigerated conditions. These signals are analyzed using a BiLSTM network with an attention mechanism, achieving classification accuracy exceeding 94% across three freshness states: fresh, semi-fresh, and spoiled. The proposed system provides a reliable and scalable instrumentation framework for characterizing RF energy harvesting performance and enabling practical battery-free freshness monitoring in real-world sensing applications.

Automated emotion identification via physiological data from wearable devices is a growing field, yet traditional electroencephalography (EEG) and photoplethysmography (PPG) collection methods can be uncomfortable. This research introduces a novel structure of the in-ear wearable device that captures both PPG and EEG signals to enhance user comfort for emotion recognition. Data were collected from 21 individuals experiencing four emotional states (fear, happy, calm, sad) induced by video stimuli. Following signal preprocessing, temporal and frequency domain features were extracted and selected using the ReliefF approach. Classification accuracy was assessed for PPG, EEG, and combined features, with combined features yielding superior results. An XGBoost classifier, optimized with Bayesian hyperparameter tuning, achieved 97.58% accuracy, 97.57% precision, 97.57% recall, and a 97.58% F1 score, outperforming support vector machine, decision tree, random forest, and K-Nearest Neighbor classifiers. These findings highlight the benefits of multimodal physiological sensing and optimized machine learning for reliable emotion characterization, with implications for mental health monitoring and human-computer interaction.

Stress is a critical factor that increases the risk of traffic accidents and should be addressed in driver monitoring and intelligent transportation systems. Electroencephalography (EEG) offers a direct, non-invasive measure of neural activity, making it a promising modality for real-time stress recognition. However, most existing studies have primarily focused on binary classification within constrained experimental settings, which limits reproducibility and practical relevance. In this study, the DEAP dataset was repurposed for driver stress detection via a Russell’s circumplex model–based label transformation, yielding three stress levels (no, low, high). Four state-of-the-art EEG models—EEGNet, EEGNeX, FBCNet, and TSception—were benchmarked under a subject-balanced stratified split strategy, with automated hyperparameter optimization using Optuna’s TPE sampler and Focal Loss to mitigate class imbalance. Experimental results revealed clear trade-offs among accuracy, inference latency, and computational complexity. FBCNet achieved the lowest inference latency (2.1 ms, 60,523 samples/s) with the smallest model footprint, whereas EEGNet demonstrated balanced performance with 78.8% accuracy. EEGNeX attained 81.6% accuracy at the cost of increased complexity, while TSception achieved the highest accuracy (98.9%) and strong generalization, confirming its suitability for EEG-based stress detection. By combining dataset relabeling, reproducible evaluation protocols, and comprehensive benchmarking, this study suggests TSception as the most effective model for EEG-based stress recognition and provides practical guidance for integrating EEG-driven stress monitoring into driver safety and intelligent transportation systems.

This study presents an innovative framework for remote stress monitoring that integrates in-ear electroencephalography (EEG), hybrid quantum-classical convolutional neural networks (QC-CNN), and Internet of Medical Things (IoMT) technologies. Leveraging in-ear EEG signals, the proposed system addresses the limitations of conventional scalp EEG systems, such as bulkiness and user discomfort, while enabling continuous, unobtrusive monitoring. The QC-CNN model synergistically combines quantum computing and deep learning to enhance stress classification performance. Key contributions of this work include: (1) the development of a comprehensive in-ear EEG system comprising both hardware and software components; (2) the introduction of a novel time–frequency signal-to-scalogram conversion technique based on the fractional adaptive superlet transform (FASLT); (3) a comparative evaluation of the QC-CNN model against traditional neural networks; and (4) the implementation of an IoMT-enabled platform for real-time remote stress monitoring and management. The system was validated through experiments involving stress-inducing tasks such as the Stroop color-word test, mental arithmetic, and memory recall. Results demonstrated strong classification performance, with 0.906 accuracy using leave-one-out cross-validation (LOOCV) and 0.942 using 10-fold cross-validation, alongside high sensitivity, specificity, precision, and F1-scores. The Beta frequency band emerged as the most informative for stress detection, and right-ear EEG signals yielded superior accuracy compared to left-ear recordings. This study underscores the potential of integrating in-ear EEG, QC-CNN models, and IoMT frameworks for effective real-time stress monitoring, offering a promising avenue for personalized healthcare and stress management.

Monitoring mental states is essential for understanding physiological and neurological health, offering insights that enhance diagnostics and therapy. However, current systems are often bulky, immobile, and reliant on external devices for deep learning (DL), limiting real-time applications. This study introduces a compact, real-time mental state monitoring system with four innovations. First, a wearable ear-centric device captures multimodal biosignals-behind-the-ear (BTE) electroencephalography (EEG) and in-ear photoplethysmography (PPG)-enhancing portability. Second, a high-resolution superlet transform (HRST) converts these signals into spectrogram images to enrich input features for DL models. Third, the PhysAttenConvLSTM model, designed for edge deployment using tiny machine learning (TinyML), integrates convolutional neural networks (CNNs), long short-term memory (LSTM), self-attention, and a physics-informed merge layer (PIML) to classify affective states efficiently. All processes—from signal acquisition to classification—occur directly on the edge device, enabling low-latency, real-time operation without external computation. Fourth, the Internet of Medical Things (IoMT) framework is implemented for data storage, remote access, and system management. Multimodal data were collected from 25 participants during social media-based mental state induction tasks. The system achieved 92.07% accuracy in leave-one-out cross-validation (LOOCV) and 95.05% in ten-fold cross-validation (CV), with strong specificity, sensitivity, precision, and $F1$ scores. This research advances edge DL, multimodal biosignal fusion, and IoMT integration, offering an ultraportable, energy-efficient solution for real-time mental state monitoring, with potential in affective computing and personalized mental healthcare.

Integrating computer vision and physiological analysis in driver drowsiness detection (DDD) is a promising technology for accurately identifying drowsy states while driving, thereby preventing potentially dangerous accidents. This study proposes a multimodal DDD system with a deep neural network that combines computer vision-based face expression analysis and electroencephalogram (EEG) data analysis. Key contributions include: 1) providing a comprehensive hardware, firmware, and software design for the DDD system to acquire behind-the-ear (BTE) EEG signals, rather than conventional scalp EEGs, due to their convenience and practicality; 2) proposing a powerful and lightweight GAN-based auto-denoising method to eliminate artifacts from EEG signals during signal acquisition, significantly influencing the quality of the obtained result; 3) developing a multimodal DDD network by combining EEG analysis and computer vision-based face expression identification to improve performance in monitoring and early detection of the driver’s drowsiness while engaging in traffic. The study employs the relative root mean squared error (RRMSE) in both temporal and spectral domains to quantitatively assess the performance of the proposed approaches in artifact removal. The proposed GAN-based auto-denoising network outperforms other comparable approaches, with an RRMSE (temporal) of 0.210 and RRMSE (spectral) of 0.161. The proposed trained multimodal model with GAN-based auto-denoising is superior to other models with different denoising approaches in driver drowsiness detection across all five-evaluation metrics, with an accuracy of 95.33%, specificity of 95.48%, sensitivity of 95.17%, precision of 95.47%, and an F1-score of 95.32%. The experimental results demonstrate the practicality and feasibility of our proposed DDD system.

Automated meat monitoring is crucial for ensuring safety, maintaining quality, reducing economic losses, adhering to regulations, managing inventory effectively, and minimizing environmental impact. This article presents a self-powered system for classifying meat freshness using ensemble learning and a battery-free sensor tag powered by energy harvested from a 3D quasi-Yagi antenna (QYA). The QYA design features a slotted-bowtie structure at the driven element and a meandering structure at the reflector, optimized through generative adversarial networks to address various parameters and reduce design time. Constructed from five 2D QYAs arranged in a cube and an RF harvester PCB base, the 3D QYA considerably enhances the energy harvesting performance. Compared to its 2D counterpart, the 3D QYA significantly enhances coverage area, gain, energy-scavenging efficiency, and transmission distance, achieving an average gain of 11.43 dBi over a 25 MHz band across prediction, simulation, and measurement. RF harvesting tests at 915 MHz achieved a maximum power conversion efficiency of 68.2% at 9.5 dBm input power. This energy supports the high power demands of the integrated gas sensor for battery-free monitoring of meat freshness. Over five days, data on equivalent carbon dioxide, total volatile organic compounds, temperature, and storage time were collected and analyzed using ensemble learning models with five primary learners. The system accurately classified meat freshness into fresh, semi-fresh, and rotten categories, tracking transitions between states with over 96% classification accuracy. Our results confirm the effectiveness of this system as a practical solution for industrial applications and improving health and safety standards in meat consumption.