Methods of collecting and analyzing human muscle activity data for exoskeleton control

The analysis of modifications of modern noninvasive methods of collecting muscle activity data for controlling robotic prostheses and exoskeletons is carried out. The main focus is on non-invasive control methods based on electromyography (EMG) and electroencephalography, as well as the use of artificial neural networks for signal classification. Key issues are considered, in particular, the presence of low-frequency noise in EMG data, the high computational complexity of neural network training, and the formation of a sufficiently large dataset for neural network training. To eliminate noise, digital filtering and wavelet transform methods are proposed, the use of which eliminates most of the noise, improving the accuracy of gesture recognition by up to 94%. Data storage formats (XDF, WFDB) and existing devices for recording EMG signals, such as hardware and software complexes using SVM and neural networks, are analyzed. An analysis of the existing dataset, which was formed by eight EMG sensors, was carried out, which made it possible to study the structure of EMG signals. The features, target variables, and their maximum and minimum values for the training dataset are defined. The analysis of the muscles of the upper extremities made it possible to conditionally group them into functional groups and select in each group the surface muscles that make the greatest contribution to the movement of the limb, which made it possible to determine the attachment of EMG electrodes. A prototype device based on an Arduino Nano microcontroller, EMG sensors from BiTronics Lab and MPU6050 6DOF sensors designed for synchronous recording of muscle biopotentials and spatial position of limbs has been developed. The device generates a dataset containing bioelectric muscle signals, acceleration, and angular velocities along three axes for each 6DOF sensor. A preliminary training dataset has been formed, which will be processed by the proposed filtering methods and used in the future for the development and training of a neural network.