Extraction of appropriate features for EEG voluntary movement signals detection

Dissertation for Master's Degree

Artificial Intelligence

Abstract

In this thesis, we intend to perform classification on brain signals by presenting a suitable feature. For this purpose, first the noise of the recording device is removed from the brain signals, then the features are extracted from these signals using Walsh transform and entropy. After feature extraction, classification is done based on these features.

The first pre-processing to classify brain signals is to remove noise from these signals. In this thesis, two classic noise removal methods and two proposed noise removal methods are examined. First, noise is removed from the signal by using the classic ICA method, wavelet transform and the two proposed methods of Walsh transform and the combined method of Walsh and ICA. To have an evaluation of these several methods, the results of these four methods are evaluated using three criteria, signal-to-noise ratio (SNR), mean square error (MSE) and root mean square difference (percentage) (PRD). The evaluation results using these criteria showed that the combined method of Walsh and ICA and Walsh transformation has the lowest mean square error. Also, these two methods have the highest signal-to-noise ratio and the root mean square difference (percentage).

After removing the noise from the signal, the discussion of feature extraction from the signals and their classification is discussed. The extracted features are a small number of features and a feature vector has 22 components. These features are related to entropy of Walsh transformation of signal channels, entropy of Walsh transformation of the whole signal, Walsh transformation power of signal channels and Walsh transformation power of the whole signal. To evaluate the efficiency of these features, the same features are also extracted using wavelet and Fourier transform, and the classification process based on the extracted features of these three methods is done separately. After extracting the feature, based on the extracted features, the signals are classified using the SVM classifier and the nearest neighbor. The results show that the classification using the extracted features of the Walsh transformation is far better than the classification based on the features of the other two transformations. The detection rate using the proposed method and svm is 42.5% and the nearest neighbor method is 39.0%.

In another comparison, the results are compared with the results implemented on this data set, in the fourth BCI competition. The results showed that the classification method using Walsh transformation is better than all methods except the first one. But the advantage that the proposed method has over all methods is that in terms of time, this method has a small amount of testing and training time. This time is 52 seconds, which is far better than the first method, which is 403 and 640 seconds.

Key words: Walsh transform, brain signals, signal-to-noise ratio (SNR), mean square error (MSE) and root mean square difference (percentage) (PRD

Introduction

1-1- Introduction

Human interaction With the HCI, this is the study of the intersection of the computer and the interaction of the computer. Important about him The field includes branches from both parties involved, for example, computer graphics, operating systems, programming languages, communication theory and industrial design for the computer part of linguistics, psychology and human performance for its human part. This field is divided into many branches, one of which is the brain-computer interface (BCI) [2]. Telepathic payment. One of the methods of recording these signals is EEG (3). By recording these signals, human efforts to use these signals for various applications began.Now, the most common uses of these signals are in medical diagnosis and helping physically and mentally disabled people [1]. In the early days of recording these signals, it was difficult to work on and extract useful information from them because of the chaotic and noisy nature of these signals.

In the early days of discovering brain signals, due to the lack of suitable recording devices, it was believed that human communication with the surrounding environment is difficult and impossible. But with the advances made in the field of computers and electronics and with the invention of suitable tools to record brain signals, this connection is not far from reach. Today, BCI is the science that establishes this connection.

The brain-computer interface consists of a set of sensors and signal processing components that directly convert a person's brain activity into a series of communication or control signals. In this system, brain waves must first be recorded using brain wave recording devices, which usually use EEG to record brain waves due to high time accuracy and cheapness as well as ease of use. EEG electrodes are placed on the surface of the scalp and measure the electric field resulting from the activity of neurons [5]. In the next step, these waves are checked and the desired features are extracted, and from these features, it can be guessed what activity the user has in mind. In figure (-11), we see the processing units of the BCI system.

Due to the low signal-to-noise ratio in this system, first a pre-processing and noise removal operation is performed on these signals. The next step is the feature extraction stage, in the following chapters we will talk about the types of features and feature extraction methods. Finally, we will perform the classification process using the extracted features.

The brain-computer interface may have a fixed structure or it may be adaptive and adapt itself to the characteristic or characteristics of the signal. It is also possible to give feedback [6] to the tested person from the output of the system. This method is known as biofeedback.

In the first international conference held in June 1999, a common definition for BCI was presented as follows [2]: (a brain-computer interface is a communication system that does not depend on the normal output paths of the peripheral nervous system and muscles) The electrical signals of the brain overlap with some other vital signals in terms of amplitude and frequency, so the definition of BCI emphasizes the independence of signals from other nerve and muscle signals.

1-2- History of BCI

The first efforts in the field of human-computer interaction started with the discovery of EEG signals and scientists tried to make a connection between these signals and brain activities[1]. But due to the fact that initially these signals were very chaotic and noisy, these signals were only used in medicine and only expert doctors could use these signals due to their experience. But little by little, with the production of new devices and the ability to record these signals with better quality, more research was done in this field.

In 1969, Elul [3] made the first attempt. He worked on the signal of mathematical operation and showed that if a person does not perform a certain mental operation, 66% of the brain signal is Gaussian distribution, and if a person performs mathematical operation, 32% of the brain signal has a Gaussian distribution, and through the brain signal, he was able to identify what mental operation the person is doing.

At the University of Colorado, two researchers, Keirn and Aunon, began their research in this field to categorize five different mental activities [4]. They recorded the EEG signal while performing five specific mental activities simultaneously from several channels. Then, with the help of a Bayes separator [7], they used the power of different frequency bands as features to separate these mental activities. In their work, they proposed the idea that different mental activities can be used as an alphabet to communicate directly between the brain and the outside world; So that a person can convey his intention to the outside world by combining and choosing a sequence of certain activities. Several years later, Dr. Anderson and his colleagues [5,6] continued the work of these two researchers. This group used the same five mental activities in most of their work. They estimated statistical parameters such as coefficients (AR) [8] and using this coefficient, they extracted features to classify and distinguish these five practices.