<p>With the increasing popularity of human activity recognition (HAR) in wearable devices due to ease of use and user accessibility, this field has become one of the most popular research topics. The applications of HAR in security and monitoring, sports, human-computer interaction, and healthcare have gained special importance for long-term and remote monitoring of the elderly and patients. Despite progress, existing HAR models that integrate convolutional neural networks (CNNs), recurrent neural networks, and attention mechanisms struggle to accurately distinguish similar or complex activities. Another major challenge in HAR is noise in sensor signals, which can negatively affect recognition accuracy. Additionally, one-dimensional CNNs (1DCNNs) are limited in capturing spatial patterns. To address these challenges, we propose a deep learning architecture combining a two-dimensional CNN (2DCNN) for robust spatial feature extraction, a bidirectional long short-term memory (BiLSTM) network for enhanced temporal modeling, and a soft attention mechanism to prioritize critical features while mitigating noise. This model has been evaluated on three public datasets: WISDM, MHEALTH, and KU-HAR. The overall accuracy of the proposed method reached 99.10% in the WISDM dataset with six activities, 99.38% in the MHEALTH dataset with 13 activities, and 97.50% in the KU-HAR dataset with 18 activities. Experimental results reveal that the proposed method achieves higher accuracy than state-of-the-art deep learning models and can be effectively utilized in various fields, including smart health and the Internet of Things.</p>