A Simple Reshaping Method of sEMG Training Data for Faster Convergence in CNN-Based HAR Applications

Convolutional neural networks (CNNs) have demonstrated excellent image recognition performance. CNNs have also been successfully extended to human activity recognition (HAR) applications, which aim to recognize the intent of human actions or diagnoses for clinical purposes using sEMG collected from the human body. It has been observed in the literature that using visual image training data to train a CNN is prone to having a square matrix in terms of shape in image recognition, while using sEMG training data to train CNN may inherently have a rectangular matrix with a small number of rows and a vast number of columns in terms of shape in HAR applications. This leads to the assumption that CNN may converge much faster in its learning if sEMG training data are reshaped to have a square matrix in terms of shape without losing any information from the original training data in the shape of a rectangular matrix. This study proposes a simple but very effective reshaping method to reshape sEMG training data in terms of shape from a rectangular matrix to a square matrix without losing any of the original information. Empirical studies confirm that the proposed reshaping method enhances CNN learning such that it converges much faster regardless of optimizers and CNN models considered in the study. Our findings strongly recommend the use of CNN learning in sEMG-based HAR application.