Lower Limb Analysis Based on Surface Electromyography (sEMG) Using Different Time-frequency Representation Techniques

Abstract

Using time-frequency representation techniques, projecting 1D sEMG signals onto a 2D image space can help diagnose several muscle activities. The acquired sEMG signal can provide valuable representative information about the muscle activity firing rates during muscle contraction. Different phases of muscle activity can be discernible via the sEMG signals by extracting discriminating features. The behavior of muscle activity was acquired in measurements of five muscles, i.e., RF, BF, VM, ST, and FX. Previous attempts to visualize lower limb analysis to extract sEMG features adopted One-dimensional (1D) sEMG segments. This work proposes a comparative experiment between three time-frequency representation techniques. The three time-frequency representation techniques, scalogram, spectrogram, and persistence spectrum, were used to map muscles' (1D) sEMG signal straightening the knee. The two-dimensional (2D) projected images are then fed into a convolutional neural network (CNN) model for detecting knee abnormality. The experiments are performed via 10-fold cross-validation. The number of kernels is incremented along with model layers. The fully connected layers were adjusted according to the loss value. Besides, tuning the hyper-parameters of the dropout parameters and the ReLU activation function to verify optimal performance. This research shows that the scalogram image representation gives significantly better performance than the spectrogram and persistence spectrum in recognizing knee abnormality. In addition, this study may help in guiding the diagnosis of several human muscle activities via the sEMG signal. A more diverse of muscles can be further investigated and can be useful for future work to enhance the diagnosis accuracy.