Real-time processing of electromyography signals using artificial neural networks

The processing of electromyography (EMG) signals is complex due to the stochastic nature of the signal itself. Artificial neural networks (ANN) computationally implement a type of processing similar to that of the human brain, which is appropriate for these signals, and are one of the valid techniques used for their processing. Using supervised learning algorithms; with ANNs it is possible to identify patterns in EMG signal recordings. With this, it is possible to build an interface that allows us to interact with technological devices. In this work, a trained ANN is implemented using the records of 3 types of movement. The network must be able to identify: twisting of the wrist, extension of the fingers of the hand and contraction of the arm. Data acquisition and network implementation are performed using a microcontroller for signal conversion and a Python computational environment. A sequential network structure was established, whose output indicates the probability that the input corresponds to one of the patterns to be classified. A database was created for the training and validation of the network. Analyzing the results obtained by it, we see that a precision level of 88% was reached on the training set and 84% on the validation set, which shows the viability of this type of processing for pattern classification.