An Architecture for Human Activity Recognition Using TCN-Bi-LSTM HAR based on Wearable Sensor

Human activity recognition (HAR) based on sensor data plays a vital role in various fields. Therefore, it is important to improve the recognition performance of different types of actions. In this work, we propose a Temporal Convolutional Network (TCN) combined with a Bidirectional Long Short-Term Memory (Bi-LSTM) architecture to address the issues of insufficient time-varying feature extraction and gradient explosion caused by too many network layers. This architecture effectively recognizes and emphasizes key feature information. The TCN enhances temporal feature extraction (TFE) with an appropriately sized receptive field, while the BiLSTM captures information from both past and future time steps, making it well-suited for tasks requiring bidirectional temporal context. This integration enables the architecture to learn and identify human activities more effectively. The performance of the proposed architecture is evaluated on three benchmark datasets: UCI-HAR, PAMAP2, and WISDM, achieving significant accuracies of 99.1%, 94.8%, and 98.3%, respectively, outperforming other state-of-the-art architectures.