Contents I. Introduction
Biomedical signal processing is extensively researched in the scientific society due to its wide range of applications in health care and human machine interactions. In recent years several techniques are being developed to recognize the different human movements and transform them into meaningful commands to control the operation of machines. The use of biomedical signals such as ECG, EEG, EMG, etc. for human machine interactions has become very popular. Electromyography (EMG) is the electrical signals that result from muscle contractions in human body. They reflects information about the movements of a subject. Surface electromyography (sEMG) which are signals collected from the surface of muscles using electrodes, is widely applied for hand gesture classification, robot control, development of prosthetic and in many more applications [1]–[3] because of its user friendliness and low cost. Several researchers in the recent years have proposed deep learning techniques to classify sEMG signals and have proved that it produces promising results. Although sEMG gesture recognition using hand crafted features achieve good results, they have several drawbacks. It performs poorly under varying conditions in the dataset and hence is impacted by external influences. These hand crafted machine learning techniques cannot deal with the complexity of sEMG signals and require a lot of manual technique for feature extraction. As an alternative, deep learning techniques which are automatic feature learning algorithms are employed. Most EMG gesture classifications convert the information into 2 dimensional data using short time fourier transform (STFT), Wavelet, etc. Oh et al. [4] have converted the signal into two dimensional (2D) images by taking the short time Fourier transform (STFT) and wavelet transform (WT). The transforms are performed channel wise to obtain the results. This image is then given to a simple CNN model for classification. Wang et al. [5] have converted the signal into the time- frequency domain by using STFT. The designed structure has five convolution layers followed by fully connected layers for classification. Wang et al. [6] have proposed a novel split spectrogram approach where the spectrogram is first split along the frequency for each channel and them merged along the channels. This image is then given to a CNN model. While 2D CNNs are mostly used for feature extraction in images, 1D CNNs are found to perform well on physiological signals [7]. Kim et al. [8] proposed a 1D CNN model that can recognize the respiratory patterns obtained from a ultra wide band radar and predict the health condition of a person. The CNN model consisted of 3 convolution layers, 4 dense layers and a dropout layer along with the dense layers to reduce over fitting. A novel 1D CNN architecture that can detect autism spectrum disorder from EEG signals is proposed by Mohi et al. [9]. The authors have developed a CNN model with 6 1D convolution layers and 4 fully connected layers that has a recognition accuracy of 92%. It was observed that as more layers are added, computation time as well as efficiency of the model increases. Cheikhrouhou et al. in [10] has proposed a 1D CNN model that analyses ECG signals and detect arrhythmia cardiovascular diseases with an accuracy of 99.46%. The proposed model consisted of two 1D convolution layers and two dense layers. Dropouts are applied during the training to avoid over fitting. From the literature survey, we can conclude that even though EMG signals are one dimensional signals, 1D CNN architecture has not been applied in EMG signal classification. Since 1D CNNs has been used in various other applications with several other biological signals and since it requires low computation due to its compact configuration, it is fit to apply in real time and low expense applications.
