A strategy based on paraconsistent random forest for sEMG gesture recognition systems robust to contaminated data

Applying machine learning algorithms to physical signals is always challenging since undesirable events can occur when signals are acquired outside a controlled environment. Among several applications, movement recognition through sEMG signals is especially complicated, since they are subject to several types of contaminants that can degrade the signal. These degradations alter the characteristics of myoelectric signals, hindering the ability of pattern recognition algorithms to discriminate movement classes. In this context, this work presents the Paraconsistent Random Forest method, which combines the advantages of hybrid classifiers, including low susceptibility to noise using a Random Forest approach and the ability of Paraconsistent Logic to deal with non-ideal data. Furthermore, this hybridization of techniques increases the representative power of Decision Trees and their applicability in vague or contradictory contexts. Several experimental procedures were used to analyze the viability and robustness of the method regarding contaminants typical of the surface electromyography field, such as movement artifacts, thermal noise, and loss of electrode-skin contact. The Paraconsistent Random Forest method proved promising for use in contexts where input data degradation occurs, presenting a decrease of less than 20 % in movement prediction compared to traditional methods that showed, in the same situation, decreases of up to 90 %, invalidating the model. All experiments were statistically validated.