An enhanced dual inception-attention-BiGRU-attention model integrating wavelet transform for wearable sensor-based human activity recognition.

With advancements in deep learning (DL), machine learning (ML) techniques, and data processing capabilities have improved the accuracy and efficiency of HAR systems. Current literature emphasises the effectiveness of hybrid complex models that incorporate advanced computing components. In this study, we present a hybrid model integrating Inception, attention mechanisms, and BiGRU, termed Dual Inception-Attention-BiGRU-Attention (Dual-INABA). Our model achieves a classification accuracy of 96.95% on the self-collected WSMC-HAR dataset and 98.15%, 98.89%, and 96.78% on the benchmark datasets PAMAP2, UCI-HAR, and WISDM, respectively. We conduct ablation experiments to investigate the individual and synergistic effects of the components in our model. The results reveal BiGRU's effectiveness in reducing loss rates, highlight the Inception model's contribution to enhancing overall accuracy, and the benefits of component collaboration. We examine the influence of processing methods on the hybrid model, demonstrating that signal-denoising techniques can enhance classification performance. Notably, the Daubechies (Db) wavelet transform method outperforms its counterparts in signal-to-noise ratio (SNR) and leads to high classification accuracy. What is more, we collect acceleration signals using the MPU6050 accelerometer sensor, resulting in a dataset comparable in quality to commonly used benchmark datasets.