Using deep learning for movement classification eeg/emg type time series

In this study, electromyography (EMG) data is classified according to hand movements using time series with deep learning. In signal processing classification results can be obtained through Long-Short Term Memory Networks (LSTM), time-frequency analysis and classification method. The data set used in the study was taken from the University of California machine learning warehouse. In the dataset used in the study; EMG data are obtained from 5 healthy individuals (2 males and 3 females of the same age (approximately 20-22)). In order to gather this information about muscle activation, an anterior surface EMG electrode is attached to the anterior and posterior parts of individuals' arms. Dataset 6 is taken while doing hand gesture. These hand gestures are: for holding spherical tools (Spherical), for holding small tools (Tip), for grasping with palm facing the object (Palmar), for holding thin/flat objects (Lateral), for holding cylindrical tools (Clyindrical) and for supporting a heavy load (Hook).Speed and force are left to the individual's desire while these movements are desired. The individuals performed each movement for 6 seconds and each process (movement) is repeated along 30 times. As a result, 180 6-second long 2-channel EMG data were recorded for each individual. In order to reduce data loss and distortion during recording of signals, adjacent windowing method is used. In adjacent windowing method used in this study, window sizes are as follows; selecting 0.05 sec (25 data points), 0.10 sec (50 data points), 0.15 sec (75 data points), 0.20 sec (100 data points), 0.25 sec (125 data points) and 0.30 sec (150 data points). the performance rates obtained were compared. While 80% of the data are used for training purposes, 10% are used for validation and the remaining 10% are used for testing purposes. As a result of the classification process, the best result was obtained from the data received with a window size of 150 data points (0.30 sec). As a result of classification of the data of the first channel; The success rate of the training data is 96.18%, the performance rate of the second channel data is 99.34%, and the performance rate of the data obtained by the average of the two channel data is 99.81%.