PSR-based research of feature extraction from one-second EEG signals: a neural network study

Research Article PSR‑based research of feature extraction from one‑second EEG signals: a neural network study Aleksander Dawid1 Received: 1 August 2019 / Accepted: 25 October 2019 / Published online: 4 November 2019 © The Author(s) 2019  OPEN Abstract The speed and accuracy of signal classification are the most valuable parameters to create real-time systems for interaction between the brain and the computer system. In this work, we propose a schema of the extraction of features from one-second electroencephalographic (EEG) signals generated by facial muscle stress. We have tested here three sorts of EEG signals. The signals originate from different facial expressions. The phase-space reconstruction (PSR) method has been used to convert EEG signals from these three classes of facial muscle tension. For further processing, the data has been converted into a two-dimensional (2D) matrix and saved in the form of color images. The 2D convolutional neural network (CNN) served to determine the accuracy of the classifications of the previously unknown PSR generated images from the EEG signals. We have witnessed an improvement in the accuracy of the signal classification in the phase-space representation. We have found that the CNN network better classifies colored trajectories in the 2D phase-space graph. At the end of this work, we compared our results with the results obtained by a one-dimensional convolution neural network. Keywords Phase space reconstruction · Signal processing · Facial expressions · EEG · CNN 1 Introduction In 1924, Hans Berger developed the electroencephalography method of human brain study. [1]. The activity of the brain cortex, in this method, is visible in the change of electric potential on the scalp. Amplitude and frequency describe the oscillatory character of the signals obtained by EEG devices. Earlier research by Hans Berger has shown that there are frequency bands highly connected with the activity of the brain [2]. These frequency bands were named delta (< 4 Hz), theta (4–7 Hz), alpha (8–15 Hz), beta (16–31 Hz), and gamma (> 32 Hz). The original EEG signal is the compilation of all neuron’s activities. In this method, we are not able to distinguish the activity of single neurons or even a small group of neurons. What we get is preferably the general state of the brain. The second kind of signal is known as electromyographic (EMG) signal. The EMG is usually studied in the case of muscle disorders [3, 4], in controlling robotic limbs [5, 6] or face emotion recognition [7, 8]. As we see, the signal obtained by an EEG device represents the state of the brain and skull muscles. The signal must be processed if we want to extract information about the state of the brain or face muscles. Recently the phase space reconstruction method became the most popular in analyzing EEG data [9, 10]. The basic concept of this method is the multidimensional phase space [11]. The time-dependent EEG signal represents a single point in this space. The dimension of this space depends on the amount of data representing the signal. Nevertheless, analyzing data in such a multidimensional space could be quite challenging. The projection of this vector on a given surface creates a two-dimensional pattern that can * Aleksander Dawid, | 1Department of Transport and Computer Science, WSB University, 1c Cieplaka St, 41‑300 Dabrowa, Gornicza, Poland. SN Applied Sciences (2019) 1:1536 | https://doi.org/10.1007/s42452-019-1579-9 Vol.:(0123456789) Research Article SN Applied Sciences (2019) 1:1536 | https://doi.org/10.1007/s42452-019-1579-9 say more about the non-linear nature of the EEG signal and is easier to analyze [12, 13]. This method was adopted to process a short-timed EEG series [14]. These short-term signals are very significant in real-time brain-computer interface (BCI) development [15]. The PSR method of signal processing is widely used in transport control [16]. All those methods of signal processing are usually introducing before the classification process of the patterns. Nowadays, the neural network (NN) techniques getting more significant in expert systems. For example, in breast cancer detection [17] or medical diagnostic applications [18]. The accuracy of NN depends on its topology and the quality of input data. Currently, there are many results from multichannel EEG mapping, where spatial and temporal data of EEG signals are studied, mainly for applications in neurology [19–21]. One of the good examples may be the work of Jiao et al. In which the pattern recognition methods were used to classify EEG signals coming from working memory during cognitive tasks [22]. The installation of multi-electrode EEG devices usually requires the help of another person, which hinders the use of these devices in everyday applications. EEG devices that are currently gaining increasing popularity are mobile devices with only one active electrode and one reference electrode fastened on the ear. These systems, due to the user’s convenience, are potential candidates for use in remotely controlled computer devices. In this work, we want to propose a different approach to short-timed signal processing. As a scientific background, we will use the phase space theory. We will show that processed EEG signal, as an input to the neural network, can give better validation accuracy than raw EEG signal. Three kinds of signals will be analyzed here. The first one is the Fig. 1  Overview of project workflow Vol:.(1234567890) signal without any facial muscle tension. The second signal is coming from the jaw movement. The last signal registers the tightening of the mouth. The auto-mutual information theory will be used to show the separation of these three classes of signals. Finally, we will compare our results to that obtained from one-dimensional CNN. 2 Signals acquisition In this work, a device from NeuroSky Mindwave Mobile [23] was used to collect EEG signals. This device is an inexpensive wireless EEG headset with a single electrode powered by one 1.5 V AAA battery. The sampling frequency of this device is 512 samples per second. This device can transmit both raw data and mental state data [24]. It is also able, at the hardware level, to filter the basic frequency bands of brain activity, such as bands; delta, theta, alpha, beta, and gamma. The conversion of the signal of analog activity of neurons to a digital signal takes place inside the EEG headset. The Bluetooth (BT) wireless system connects the device and a computer. Receiving data from the device is handled by the ThinkGear library procedure written in the C# programming language. The author of this work is responsible for the development of this software. Figure 1 is showing the scheme of the EEG signal acquisition system. The time-dependent raw signal in the form of two columns as a text file is stored. The first column represents the time, and the second column represents the values of the EEG signals. This data then (...truncated)