Spatial-Temporal Fusion Network with Hybrid Attention for Energy Expenditure Prediction Based on Multi-Sensor

To reduce the measurement time of the metabolic energy expenditure in human-in-the-loop optimization, this paper proposes a spatial-temporal fusion network that combines convolutional neural network, multi-head attention and cross-attention mechanism in both temporal and frequency domains. We collected sEMG signals, inertial signals, metabolic energy expenditure signals, and heart rate signals. Four scenarios with different walking speeds and slopes are set up for the experiment. The results show that the prediction accuracy of the fusion network is better than that of traditional models. The fusion network performs the best in predicting the metabolic energy expenditure of the same subject under different walking conditions, with the mean absolute error and root-mean-squared error of 0.268 kcal/min and 0.342 kcal/min, respectively, and it also has a small error and strong anti-interference ability among different subjects.