Three-Position Vehicle Control Based on Neural Interface Using Machine Learning

Abstract

Listen

Translate article

The brain-computer interface is a complex system that allows you to control external electronic devices using brain activity. This system includes several elements – a device for reading brain activity signals, a hardware and software complex that processes and analyzes these signals, and a control object. The main challenge here is the development of methods and algorithms that can correctly recognize and predict the intentions of the person who uses this interface to provide solutions to control problems. This paper describes the mathematical formulation of the equipment control problem. Methods for preprocessing EEG signals, analyzing them, and making decisions about issuing a control signal are described; the structure of the software implementation of these methods is described, as well as a plan for experimental testing of the performance of the entire system that forms the brain-computer interface. For classification of EEG signals the methods of machine learning are used. A modification of the k-nearest neighbors method is proposed – the so-called fuzzy almost nearest neighbors method. An algorithm for the adaptive classification of EEG taking into account the drift of the parameters of the subject's model based on the method of recurrent objective inequalities (ROI) has also been developed. The control algorithm was implemented in the Python programming language. A remote-controlled wheelchair is considered as a control object, and turning the chair to the right or left is considered as a control task. To experimentally test the performance of the developed model and algorithms, more than 15 tests were carried out with five subjects in total. The approach developed and described in this article and its software implementation during testing demonstrated its effectiveness in the tasks of controlling the rotation of a wheelchair. Special attention was also paid to the resource intensity of the software implementation. Methods and algorithms were implemented taking into account the requirements that arise when performing calculations on low-performance devices with a limited amount of memory.