Selected work

Representative Papers

Autonomous AI Surveillance: Multimodal Deep Learning for Cognitive and Behavioral Monitoring

Jul 02, 2025

This study addresses the challenge of automated classroom student attention and behavioral assessment. We propose a multimodal end-to-end intelligent monitoring framework that integrates face identity recognition (LResNet + MTCNN), mobile phone usage detection, and drowsiness behavior analysis (both based on YOLOv8), deployed on an ESP32-CAM edge acquisition platform and a PHP-based web system for real-time processing and feedback. Our key contribution lies in synergistic multi-model modeling of student attentiveness, enabling simultaneous automated attendance tracking and fine-grained behavioral cognition analysis. Experimental results demonstrate strong performance: 97.42% mAP@50 for drowsiness detection, 86.45% accuracy for face recognition, and 85.89% mAP@50 for mobile phone usage detection. The system exhibits significant advantages in accuracy, real-time responsiveness, and scalability, offering a practical, deployable solution for intelligent classroom management.

0 citationsRead paper

AWARE-NET: adaptive weighted averaging for robust ensemble network in deepfake detection

Apr 01, 2025IET Conference Proceedings

To address weak cross-dataset and cross-manipulation-type generalization in deepfake detection, this paper proposes a two-level adaptive weighted ensemble framework. At the first level, three randomly initialized models of the same architecture (Xception, Res2Net101, EfficientNet-B7) are ensembled via mean fusion to reduce variance. At the second level, learnable backpropagation-based weighting dynamically assigns architecture-level weights according to each model’s reliability, enabling adaptive fusion. This work is the first to jointly integrate hierarchical weighting with diversity enhancement via random initialization. The method achieves state-of-the-art performance on FF++ and CelebDF-v2 (AUC = 100.00%, F1 = 99.95%). For cross-dataset generalization, it attains AUCs of 88.20% and 72.52%—substantially outperforming existing approaches.

0 citationsRead paper

A Multimodal Lightweight Approach to Fault Diagnosis of Induction Motors in High-Dimensional Dataset

Jan 07, 2025

To address data scarcity, high computational overhead, and fine-grained classification challenges in broken rotor bar (BRB) fault diagnosis for induction motors, this work introduces the first large-scale, real-world industrial dataset comprising 57,500 current–vibration dual-modal short-time Fourier transform (STFT) spectrograms. We propose a multimodal spectrogram joint modeling framework that enhances fault harmonic visibility via FFT-based spectral enhancement. Furthermore, we pioneer the integration of the lightweight ShuffleNetV2 architecture with transfer learning for BRB diagnosis. Evaluated on 10,000 test spectrograms, our model achieves 98.856% classification accuracy, supporting fine-grained identification of one to four broken bars. It significantly reduces parameter count and inference latency while maintaining robust performance under varying load and speed conditions—demonstrating strong suitability for high-accuracy, low-overhead, and robust industrial deployment.

0 citationsRead paper

Recent publications

Latest Papers

Autonomous AI Surveillance: Multimodal Deep Learning for Cognitive and Behavioral Monitoring

Jul 02, 2025

This study addresses the challenge of automated classroom student attention and behavioral assessment. We propose a multimodal end-to-end intelligent monitoring framework that integrates face identity recognition (LResNet + MTCNN), mobile phone usage detection, and drowsiness behavior analysis (both based on YOLOv8), deployed on an ESP32-CAM edge acquisition platform and a PHP-based web system for real-time processing and feedback. Our key contribution lies in synergistic multi-model modeling of student attentiveness, enabling simultaneous automated attendance tracking and fine-grained behavioral cognition analysis. Experimental results demonstrate strong performance: 97.42% mAP@50 for drowsiness detection, 86.45% accuracy for face recognition, and 85.89% mAP@50 for mobile phone usage detection. The system exhibits significant advantages in accuracy, real-time responsiveness, and scalability, offering a practical, deployable solution for intelligent classroom management.

0 citationsRead paper

AWARE-NET: adaptive weighted averaging for robust ensemble network in deepfake detection

Apr 01, 2025IET Conference Proceedings

To address weak cross-dataset and cross-manipulation-type generalization in deepfake detection, this paper proposes a two-level adaptive weighted ensemble framework. At the first level, three randomly initialized models of the same architecture (Xception, Res2Net101, EfficientNet-B7) are ensembled via mean fusion to reduce variance. At the second level, learnable backpropagation-based weighting dynamically assigns architecture-level weights according to each model’s reliability, enabling adaptive fusion. This work is the first to jointly integrate hierarchical weighting with diversity enhancement via random initialization. The method achieves state-of-the-art performance on FF++ and CelebDF-v2 (AUC = 100.00%, F1 = 99.95%). For cross-dataset generalization, it attains AUCs of 88.20% and 72.52%—substantially outperforming existing approaches.

0 citationsRead paper

A Multimodal Lightweight Approach to Fault Diagnosis of Induction Motors in High-Dimensional Dataset

Jan 07, 2025

To address data scarcity, high computational overhead, and fine-grained classification challenges in broken rotor bar (BRB) fault diagnosis for induction motors, this work introduces the first large-scale, real-world industrial dataset comprising 57,500 current–vibration dual-modal short-time Fourier transform (STFT) spectrograms. We propose a multimodal spectrogram joint modeling framework that enhances fault harmonic visibility via FFT-based spectral enhancement. Furthermore, we pioneer the integration of the lightweight ShuffleNetV2 architecture with transfer learning for BRB diagnosis. Evaluated on 10,000 test spectrograms, our model achieves 98.856% classification accuracy, supporting fine-grained identification of one to four broken bars. It significantly reduces parameter count and inference latency while maintaining robust performance under varying load and speed conditions—demonstrating strong suitability for high-accuracy, low-overhead, and robust industrial deployment.

0 citationsRead paper