Gated Recurrent Unit Based On Feature Attention Mechanism For Physical Behavior Recognition Analysis

In order to overcome the problem that traditional machine learning methods rely heavily on artificial feature selection and have low recognition accuracy in the field of human behavior recognition, a deep learning model based on multi-layer recurrent neural network (RNN) and feature attention mechanism is proposed. The feature of sensor data is automatically extracted to realize physical motion recognition. Feature attention mechanism is used to analyze the correlation between historical information and input features, and extract important features. Temporal attention mechanism independently selects historical information of Gated Recurrent Unit (GRU) network at key time points to improve the stability of long-term prediction effect. This model uses multi-scale convolutional neural network and GRU to extract features from sensor data. The feature matrix is spliced in the matrix dimension and then the feature classification is completed by Softmax. Experimental results show that the accuracy of human physical behavior recognition based on public human behavior recognition (HAR) data set is 97.87%. The proposed model achieves better accuracy and avoids complex signal preprocessing stage.