Page 52 - IJEEE-2023-Vol19-ISSUE-1

48 |
minimize the cost of data processing by recognizing the most
relevant feature components embedded in the signal.
In this respect, the feature extraction process can be
conducted in multiple signal processing domains, such as
spatial domain, time domain, frequency domain, and time-
frequency domain. Particularly, time-frequency features of
MI EEG signals are widely used for classification in BCI
applications, whereas it describes the density and intensity of
energy of signal at a different time and frequency by
designing a joint function of time and frequency [1].
Principally, EEG signal analysis in the time-frequency
domain based on wavelet transform (WT) das proved their
ability and usefulness in handling brain signal characteristics
compared to other methods such as short-time Fourier
transform (STFT), autoregressive model (ARM), and
wavelet transform (WT) [8]. To date, great attention is given
to the WT in the field of biomedical signal processing
because of its efficiency in the diagnostic as well as in the
pattern recognition [1].
Deep neural networks (DNNs) have recently
demonstrated impressive categorization capabilities in a
variety of applications, including computer vision, video
processing, and speech recognition. Several academics were
inspired by its enormous success to investigate how deep
neural networks may be used to categorize EEG signals [9,
10]. Researchers have begun using deep learning in their BCI
applications, including seizure detection, memory retrieval,
and MI categorization [11]. The convolution neural network
(CNN) has demonstrated that it is capable of extracting
spatial and temporal characteristics from magnetic induction
(MI) data. It has been shown that CNN can extract excellent
features using both shallow and deep models, indicating that
significant features may be retrieved at different levels [4].
However, one of the most significant challenges in the
categorization of MI EEG characteristics using deep learning
algorithms is the limited amount of data available due to the
exhaustion of patients throughout the experiments [7]. There
are also significant individual variations between various
subjects, making it hard to directly utilize the labeled data
from other subjects to train the classifier that would be used
to identify the target individuals [12]. Meanwhile, the
collection of EEG data is extremely costly, and a sufficient
quantity of labeled samples is difficult to come by [13].
When it comes to combining data from domains with
different distributions, transfer learning has emerged as a
viable approach. Incorporated within transfer learning are
methods that are designed to transfer representations and
information from one domain to another [14]. In other
words, the approach enables researchers to easily integrate
fresh datasets into a machine learning model that has already
been trained. Having this functionality can be especially
useful in a BCI system since the amount of data supplied is
frequently insufficient to ensure the appropriate training of a
machine learning model [15]. In the BCI studies, it was
discovered that the CNN-based subject-transfer technique
outperformed the others. Subject-transfer strategies are
based on the idea that the typical patterns of the target subject
and other subjects may be comparable when doing the same
activity [16].
Al-Qaysi1, Al-Saegh, Hussein & Ahmed
It is critical to select the ideal mother wavelet along the
course of using WT with deep transfer learning. The
selection of a mother wavelet (MWT) function, on the other
hand, has been reported in the literature as an important step
and component of wavelet analysis to demonstrate the
advantages of WT in denoising, component separation,
coefficient reconstruction, and feature extraction from
signals in the time and frequency domains [17]. The need for
this phase arises from the fact that no unique MWT basis
functions have yet been identified that cater to all EEG
channels.
Specifically, the research proposes an approach for
feature extraction and classification that is based on a
continuous wavelet transform (CWT) in conjunction with
deep learning-based transfer for feature extraction and
artificial neural network (ANN) for classification.
The following is the structure of the reminder for this
paper: WTNN development, evaluation, and validation were
all carried out under a methodological framework defined in
Section 2. Sections 3 and 4, which summarise the findings
and commentary for the experimental component,
respectively, are based on two separate datasets, namely the
BCI Competition dataset IV/2b and the Emotive EPOC
dataset, and are divided into two sections. After that, the final
section of this study displays the findings of this research.
II. LITERATURE REVIEW
Al-Qazzaz reported in [18] that even a 0.1%
intensification in the classification accuracy in medical
research fields is considered vital due to the high complexity
of their signals. Therefore, many methods are proposed to
attain the highest classification accuracy.
Xu et. al. [19] used a transference CNN framework based
on VGG-16 and time-frequency spectrum images generated
by STFT for analyzing MI EEG signals. They obtained
71.2% of BCI competition classification accuracy for the IV
2b dataset. Using a continuous wavelet transform (CWT)
filter bank to classify four MI tasks (left hand, right hand,
feet, and tongue).
Mahamune and Laskar [20] proposed a framework for
developing two-dimensional (2D) images for CNNs that uses
a continuous wavelet transform (CWT) filter bank to classify
four MI tasks. On the BCI competition IV 2a dataset, they
achieved an accuracy of 71.25 percent in classification
accuracy. Using a common spatial pattern (CSP) and
filtering in conjunction with the empirical mode
decomposition (EMD).
Alvarez-Meza et al. [21] were able to distinguish the mu
and beta rhythms from one another. The support vector
machine (SVM) classifier was utilized by the researchers.
They attained classification accuracy of 92.86 percent on the
BCIC IV-I dataset and 72.30 percent on the BCIC IV 2b
dataset, respectively, on the two datasets.
Wang et. al. [22] proposed a method based on CSP as
preprocessing while feature extraction was done using
autoregressive and log-variance; the Kullback-Leibler
divergence was for feature selection and time segment
selection. Th classification is achieved using the linear
discriminate analysis technique. They achieved a