Human action recognition has emerged as a pivotal area of contemporary research, due to its immense potential across various fields.Yet, traditional machine learning methods struggle with complex recognition tasks amidst rapid advancements in sensor technology and deep learning. To resolve this challenge, our research introduces MHCAGE, a novel dual-stream neural network model with a multi-head attention mechanism based on the CAGE model. This model seamlessly integrates accelerometer and gyroscope signal characteristics, boosting recognition accuracy and robustness. Furthermore, to further enhance the feature extraction module, we append a batch normalization layer subsequent to each convolutional layer’s output, aiming to bolster the model’s stability and reliability during training.Experiments on UCI_HAR and mHealth datasets confirm MHCAGE’s superior recognition performance, validating our method’s efficacy and feasibility.