Related Experiment Video
Updated: Jan 8, 2026
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
5.2K
An enhanced dual inception-attention-BiGRU-attention model integrating wavelet transform for wearable sensor-based
Xinguo Zhang1,2, Yitao Wang1,2, Tao Deng3,4,5
1Key Laboratory of Linguistic and Cultural Computing of Ministry of Education, Chinese National Information Technology Research Institute, Northwest Minzu University, Lanzhou, 730030, China.
Scientific Reports
|December 16, 2025
Summary
This study introduces Dual-INABA, a hybrid deep learning model for human activity recognition (HAR). It achieves high accuracy on multiple datasets, demonstrating the effectiveness of combining Inception, attention, and BiGRU components.
Area of Science:
- Computer Science
- Artificial Intelligence
- Signal Processing
Background:
- Deep learning (DL) and machine learning (ML) have advanced human activity recognition (HAR) systems.
- Hybrid models integrating complex computing components show enhanced accuracy and efficiency in HAR.
- Existing HAR systems benefit from sophisticated data processing and feature extraction techniques.
Purpose of the Study:
- To present a novel hybrid deep learning model, Dual Inception-Attention-BiGRU-Attention (Dual-INABA), for improved HAR.
- To evaluate the performance of Dual-INABA on self-collected and benchmark HAR datasets.
- To analyze the contribution of individual components and signal processing methods to the model's effectiveness.
Main Methods:
- Developed a hybrid model (Dual-INABA) combining Inception, attention mechanisms, and Bidirectional Gated Recurrent Unit (BiGRU).
- Utilized ablation experiments to assess the impact of model components (Inception, attention, BiGRU).
- Investigated the influence of signal-denoising techniques, specifically the Daubechies (Db) wavelet transform, on classification performance.
Main Results:
- Dual-INABA achieved high classification accuracies: 96.95% (WSMC-HAR), 98.15% (PAMAP2), 98.89% (UCI-HAR), and 96.78% (WISDM).
- BiGRU effectively reduced loss rates, while the Inception model enhanced overall accuracy.
- The Daubechies (Db) wavelet transform significantly improved signal-to-noise ratio (SNR) and classification accuracy.
Conclusions:
- The proposed Dual-INABA model demonstrates superior performance in human activity recognition.
- Component collaboration within the hybrid model and advanced signal processing are crucial for high HAR accuracy.
- The MPU6050 accelerometer sensor can generate high-quality acceleration data for HAR applications.
