Neurocomputing 404 (2020) 108–121

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

A novel end-to-end 1D-ResCNN model to remove artifact from EEG

signals
Weitong Sun a,b, Yuping Su a,b, Xia Wu a,b, Xiaojun Wu a,b,∗
a
Key Laboratory of Modern Teaching Technology, Ministry of Education Shaanxi Normal University, Xi’an 710062, China
b
School of Computer Science, Shaanxi Normal University, Xi’an 710062, China

a r t i c l e i n f o a b s t r a c t

Article history: Electroencephalography (EEG) signals are an important tool in the field of clinical medicine, brain re-
Received 18 February 2019 search and the study of neurological diseases. EEG is very susceptible to a variety of physiological sig-
Revised 5 November 2019
nals, which brings great difficulties to the research and analysis of EEG signals. Therefore, removing noise
Accepted 8 April 2020
from EEG signals is a key prerequisite for analyzing EEG signals. In this paper, a one-dimensional residual
Available online 23 April 2020
Convolutional Neural Networks (1D-ResCNN) model for raw waveform-based EEG denoising is proposed
Communicated by Dr. Nianyin Zeng to solve the above problem. An end-to-end (i.e. waveform in and waveform out) manner is used to map
a noisy EEG signal to a clean EEG signal. In the training stage, an objective function is often adopted to
Keywords:
Electroencephalogram (EEG) optimize the model parameters and in the test stage, the trained 1D-ResCNN model is used as a filter
Artifacts removal to automatically remove noise from the contaminated EEG signal. The proposed model is evaluated on
Deep learning the EEG signal from the CHB-MIT Scalp EEG Database, and the added noise signals are obtained from the
End-to-end database. We compared the proposed model with the independent of the composite analysis (ICA), the
One-dimensional residual convolutional fast independent composite analysis (FICA),Recursive least squares(RLS) filter,Wavelet Transform (WT) and
neural networks model (1D-ResCNN) Deep neural network(DNN) models. Experimental Results show that the proposed model can yield cleaner
waveforms and achieve significant improvement in SNR and RMSE.Meanwhile, the proposed model can
also preserve the nonlinear characteristics of EEG signals.
© 2020 Published by Elsevier B.V.

1. Introduction acquisition process and often mask the waveform characteristics of

EEG, which makes the reading of EEG signals more difficult and
Electroencephalography (EEG) is the electrical response of brain brings great difficulties to the subsequent research and application
cells in the cerebral cortex. It is typically collected by an electrode of EEG signal [4,5]. Therefore, it is of great theoretical and practical
collection system placed on the head of the brain [1]. Through significance to develop relevant methods to remove artifacts from
the analysis of EEG, we can obtain rich physiological, psycholog- EEG signals and remain useful information.
ical and pathological information. The EEG signal not only shows The most common method of denoising EEG is the artifact
brain function but also shows the state of all body systems. In elimination method, which is to identify and remove artifacts in
addition, the EEG signal plays an important role in the detection the brain signal and to completely preserve the neurological fea-
and treatment of brain diseases such as epilepsy and brain tu- tures and phenomena of the original signal [6,7]. Artifacts reduc-
mors and can be used to diagnose brain death [2,3]. However, EEG tion algorithms are mainly done in two ways: The first is through
is a highly random nonlinear non-stationary signal, which con- regression and filtering methods; the second is by separating or
tains very complex components, and the signal amplitude is mi- decomposing EEG data and noise data into other domains [8].
crovolts, the intensity is very weak, and it is very susceptible to The regression model uses a function to fit the data to smooth
other physiological signals of the human body, such as ElectroOcu- the data. Using a regression analysis method, observe the multi-
logram(EOG), Electrocardiogram (ECG), muscle artifacts(EMG), or modal linear model or nonlinear model between each EEG signal
interference from non-physiological signals such as spatial elec- channel and between the EEG signal channel and the reference sig-
tromagnetic noise. These artifacts exist in almost the entire EEG nal EOG, EMG or ECG channel, and find the mathematical equa-
tion suiTable for the data to eliminate noise [9–11]. However, this
method only works for reference channels that are available.
∗
Corresponding author at: Key Laboratory of Modern Teaching Technology, Min-
In digital signal processing, the filter is an important unit. Adap-
istry of Education Shaanxi Normal University, Xi’an 710062, China.
E-mail address: [email protected] (X. Wu).
tive filters can be divided into linear and nonlinear adaptive filters.

https://doi.org/10.1016/j.neucom.2020.04.029
0925-2312/© 2020 Published by Elsevier B.V.
 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 109

Nonlinear adaptive filters further include Volterra filters and neural structure in the input EEG signal, with weight sharing and defor-
network based adaptive filters [12–17]. Nonlinear adaptive filters mation robustness.Although they have inherent advantages in EEG
have stronger signal processing capabilities and complex calcula- analysis, the CNN model has not received sufficient attention in
tion [17,18]. Linear adaptive filters are too sensitive and unsTable EEG noise reduction.
to adjust the parameters. Most importantly, artifacts overlap most EEG signals are usually long one-dimensional complex sig-
of the clean EEG signals [19,20]. Therefore, filters may eliminate nals.Due to the time-varying and complexity of EEG signals, more
useful EEG signals during artifact deletion. complex nonlinear features should be extracted for EEG artifact
Theempirical mode decomposition(EMD) proposed by Huang removal.Therefore,in order to overcome the shortcomings of the
[21] decomposes the input signals into multiple empirical modes above traditional methods,and considering the nonlinear character-
according to the inherent mode (IMFs) function, which is benefi- istics of EEG time-varying signals and the advantages of CNN fea-
cial to the analysis of multi-component signals.EMD is an empirical ture extraction, this paper proposes a new one-dimensional resid-
and data-driven method for dealing with non-stationary, nonlinear, ual convolutional neural network model (1D-ResCNN) based on
stochastic processes, so it is well suited for EEG signal analysis and multi-scale kernel to remove noise from the EEG signal.The pro-
processing.However,EMD is computationally complex and may not posed model can automatically learn the nonlinear and discrimi-
be suiTable for online applications [22,23]. native deep features of the noisy EEG data and true EEG data.Then,
Blind source separation (BSS) is one of the most popular artifact these features are used to distinguish them and automatically re-
removal methods [24–27]. Independent Component Analysis (ICA) construct to obtain a clean EEG signal.
is a multi-dimensional signal processing method developed from The main contributions of this paper are as follows:
BBS, which can separate the ideal signal and noise included in the
(1) In the absence of sufficient prior knowledge, a new 1D-
EEG signal as independent components to achieve EEG signal de-
ResCNN based brain signal denoising model is proposed,
noising [28,29]. Many BSS algorithms require human intervention
which is the first application of CNN in EEG denoising;
to identify artifact components. This makes it subjective and time
(2) The proposed model operates directly on the raw EEG sig-
consuming [28,29].
nal without pre-processing or manual feature extraction.Mo-
The wavelet transform(WT) maps the signal to the wavelet do-
reover, the EEG nonlinear characteristics and the details of
main. According to the wavelet coefficients of signal and noise,
the waveform are also kept;
they have different properties and mechanisms at different scales,
(3) Combining the residual blocks of different scales, the model
eliminating the wavelet coefficients generated by noise and maxi-
obtains more abundant features and increases the nonlinear
mally retaining the coefficients of real signals [30–32].
expression ability of the convolutional neural network.
EEG signal is a complex chaotic signal with nonlinear character-
istics, and the preservation of nonlinear features is of great signif- The structure of this paper is as follows. Section 2 describes the
icance for EEG analysis and classification. Therefore, an EEG noise structure of the proposed 1D-ResCNN; Section 3 gives the experi-
reduction method is needed to make the denoised EEG signal still mental settings and datasets; Section 4 gives the experimental re-
maintains non-linear characteristics. On the other hand, with the sults; Finally, Section 5 shows the conclusion.
advancement of technology, the collection of EEG data has be-
come more convenient, providing a strong support for the imple- 2. Proposed 1D-ResCNN model
mentation of deep learning. In recent years, deep learning tech-
niques have been able to learn high-level and hierarchical repre- In this section, a deep network structure including residual
sentations directly in massive raw data, thus achieving a series of Convolutional Neural Network is designed for long duration seg-
breakthroughs in signal processing. Specifically, Xu et al. [33] ap- ment denoising of EEG signals.The designed denoising system has
plied Deep Neural Networks(DNN) to speech enhancement, and a complete end-to-end structure that does not require feature ex-
proposed a DNN-based minimum mean square error regression fit- traction of the signal at any stage.The input of the network struc-
ting speech enhancement algorithm based on logarithmic power ture are long segments of original EEG signal.The reconstruction of
spectrum of the complex relationship between noisy speech and the noisy signal has been provided at the network output.
clean speech. Xu et al. [33] and Rodrigues and Couto [34] intro-
duces an Restricted Boltzmann Machine(RBM)-based ECG denois- 2.1. Application of CNN in EEG artifact removal
ing method, and [35] proposes an improved denoising automatic
encoder (DAE) improved by wavelet transform (WT) for ECG de- The brain is a complex nonlinear system.EEG is a complex elec-
noising.Yang [36] introduced a Deep Learning Networks(DLN) EEG trophysiological signal that usually has nonlinearities,diversity and
denoising method, which subtly utilizes the structural features of uncertainty that are difficult to process by linear methods.There-
deep learning and powerful learning capabilities to improve EEG fore,nonlinear denoising for EEG signal is imperative.
denoising due to EOG artifact. Since the EEG signal is a one-dimensional discrete time se-
Convolutional Neural Networks(CNN) is a subset of deep learn- ries,we propose a 1D-ResCNN model that extracts the EEG signal
ing, which has attracted a lot of attention in recent years [37,38]. characteristics using 1D convolutional layer. Fig. 1 shows the over-
It has been used in other fields for raw continuous signals, and all proposed 1D-ResCNN structure for EEG signal waveform denois-
it works well, starting with image applications, followed by many ing, due to the small amount of EEG data, the data is enhanced
other fields, such as natural language processing, speech processing with a fixed-size sliding window [43]. The data is enhanced by
[39]. In contrast, CNN networks have not yet been used in remove dividing the given full length EEG signal into sub-signals using a
artifacts from EEG signals. However, CNN networks have recently fixed-size sliding window.We use a window size of 400 with a
found applications in studies focusing on other areas of EEG time stride of 380 (%95 of 400); each record is divided into multiple
series analysis. In works [40–42] CNN is implemented on EEG sig- equal EEG sub-signals, and the last sample is discarded. The sub-
nals, and the effectiveness of CNN algorithm in signal analysis is signal is used as the input to the 1D-ResCNN model, 1D-ResCNN
studied. In general, CNN has several natural advantages.First, CNN model uses the M convolutional layers, each of which has a 1 × N
is superior to traditional methods not only in accuracy but also in convolution kernel to learn the noisy EEG features,where Filter-
speed.Second,CNN can automatically extract and learn the best fea- M − N represents the Nth filter in layer M. Each filter is con-
tures from the original signal to achieve adaptive design.The most volved with all generated waveforms from the previous layer and
important point is that CNN is good at mining the spatio-temporal produces a further filtered waveform EEG.Finally, these features
110 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121

Fig. 1. The architecture of the training phase of the EEG denoising model based on the one-dimensional residual convolutional neural network (1D-ResCNN). First, a fixed-
size sliding window is used to divide the given full-length EEG signal into small segments to enhance the data. The colored frame overlap represents the signal overlap
rate, and each small signal is used as a separate example of learning the 1D-ResCNN model. The 1D-ResCNN model uses the M convolutional layers, each of which has a
1 × N convolution kernel to learn the noisy EEG features. Finally, these features that can distinguish EEG signals and noise are automatically reconstructed by minimizing
the objective function(MSE) to get the clean EEG signal.

Fig. 2. 1D-ResCNN network structure. Resblock1, Resblock2, and Resblock3 represent the residual blocks of the convolution kernels of 3∗ 1, 5∗ 1, and 7∗ 1, respectively, and the
three residual blocks are connected in series twice, and finally the outputs of the three multi-scale residual blocks are merged. And through a convolution layer, the final
reconstruction signal is obtained.

that can distinguish EEG signals and noise are automatically re- main layers: batch normalized layers(BN), convolutional layers, and
constructed by minimizing the objective function(MSE) to get the Inception-ResNet.
clean EEG signal.This deep network is end-to-end (noisy EEG signal BN layer: During training, changes in network parameters can
waveform and clean EEG signal waveform) mode,providing auto- cause changes in the distribution of input data later, which greatly
matic noise reduction and inputting EEG waveform segments with- reduces the generalization ability and training speed of the net-
out any manual feature extraction or post-processing. work, and makes saturation nonlinearity more difficult. This phe-
nomenon is called Internal Covariate Shift, Ioffe and Szegedy
[44] proposed batch normalization (BN) to solve this problem. For
2.2. 1D-ResCNN network architecture each batch, the activation output of the layer is normalized using
a batch normalization layer, which helps to avoid special initializa-
The structure of the 1D-ResCNN module in this paper is divided tion of parameters, speed up training, optimize results, and provide
into three sub-modules. The first is the data enhancement module, faster convergence. In order to optimize the training process of 1D-
which uses the sliding window method to divide its EEG signal ResCNN, our proposed model performs BN after the convolutional
data into sub-signals [43] using a fixed-size overlapping window. layer.
Then, its 1D-ResCNN extracts features by convolving each sub- 1D-Conv +ReLU layer: In view of the time-varying characteris-
signal. The convolution operation of our proposed method con- tics of EEG, 1D-ResCNN model is constructed considering the ob-
sists of two convolutional layers and an Inception module, which servation of the local features of the convolution kernel. The in-
has both vertical extension and horizontal expansion. The tradi- put feature in 1D-ResCNN model is a one-dimensional vector, the
tional CNN only has a vertical extension. Finally, the features ex- one-dimensional convolution kernel can be regarded as a sliding
tracted by the convolution operation are convolved by the network window on a time series, and the short-term features between
to generate a clean EEG signal. The overall model of 1D-ResCNN is sequences are extracted to achieve the effect of processing time-
shown in Fig. 2. In this study, the 1D-ResCNN is made up of 3 con- varying features. In the first layer of our proposed 1D-ResCNN
volutional layers,3 batch normalized layers (BN) and 1 Inception- model, a 1D convolution is performed on the input EEG raw signal,
ResNet.In the rest of this section, we will briefly introduce the the number of kernels is 16, and receptive field for each kernel is
 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 111

Table 1
Detailed parameters used for all the layers of
proposed CNN model.

Layer name Kernel unit Output shapes

Input – (None,400,1)
Conv1D 1 × 5,32 (None,400,32)
BN – (None,400,32)
ResBlock1 1 × 3 (None,400,32)
ResBlock1 1 × 3 (None,400,32)
add – –
ResBlock2 1 × 5 (None,400,32)
ResBlock2 1 × 3 (None,400,32)
ResBlock3 1 × 5 (None,400,32)
ResBlock3 1 × 3 (None,400,32)
merge – (None,400,96)
Conv1D 1 × 1,64 (None,400,1)
BN – (None,400,1)
Dense 1 (None,400,1)

Fig. 3. The Inception-ResNet structure of the one-dimensional-ResCNN denoising

model uses 1∗ 3, 1∗ 5, 1∗ 7 convolution kernels of different scales to extract features
5 (ie 1 × 5). The number of kernels for Conv-2 is 32, and receptive of different sizes and orientations. Residual learning is embedded in the multi-scale
field of each kernel is 5 (i.e.1 × 5);the number of kernels for Conv- kernel CNN to avoid performance degradation and build deeper networks.

3 is 1 and receptive field of each kernel is 5(i.e. 1 × 5).Network

specific parameters are displayed in the Table 1. In order to learn the features from the EEG signals more ef-
In the convolutional layer, the convolution kernel convolves ficiently, we combine the residual learning [46] and the Network
the feature vector output from the previous layer, and constructs in Network [48] idea of the Inception architecture [49] to build
the output feature vector using the nonlinear activation function. 1D-ResCNN, which are both vertical and horizontal extensions. The
Given the input signal sequence xt , t = 1, . . . , n, and filter wt , t = residual network uses the Inception structure to increase the net-
1, . . . , m, the filter in turn performs a partial convolution operation work width by paralleling the residual blocks of different scale
on the input features of the previous layer. Typically, the length convolution kernels, and obtain various scale features in the EEG
of the filter, m, is much smaller than the length of the signal se- signal, thereby improving the network feature extraction capability.
quence, n.The convolved output yt is: The structure of the Inception-ResNet module used in this paper is

m shown in Fig. 3.
yt = wk × xt−k+1 (1) In this paper, the Inception-ResNet module is designed with
k=1 three parallel residual block branches, which use residual blocks
In the convolutional layer, each neuron in the lth layer is only of 1 × 3, 1 × 5, 1 × 7 convolution kernels of different scales,
connected to neurons in a local window of the (l − 1 )th layer to and residual blocks of each branch are repeatedly stacked twice
form a local connection network. The convolutional layer requires to extract features of different sizes and different orientations,
an activation function f(x) for nonlinear feature mapping. In this so that the features of different convolution kernels are com-
paper, a modified linear unit with faster convergence speed is se- bined. Through the connection operation, it forms a tensor of three
lected as the activation function: branches and expands the feature mapping space to achieve the
fusion of multi-scale features.
f (x ) = max(0, x ) (2) Each Res-block consists of three one-dimensional convolutional
Then the input of the ith neurons of the lth layer is defined as: layers, three BN layers and one shortcut connection. The size of
  the convolution kernel of the three one-dimensional convolutional

m layers in Res-block1 is [1 × 3].The first and third layers uses 32
ali = f wlj × al−1
i− j+m
l
+b = f (wl × al−1
(i+m−1 ):i + b )
i
(3) filters, and the second layer uses 16 filters. Unlike Res-block1, the
j=1 convolution kernels of the three one-dimensional convolution lay-
ers in Res-block2 and Res-block3 are [1 × 5] and [1 × 7] respec-
Then the input of the ith neurons of the lth layer is defined
tively, the number of filters is the same. The internal structure of
as: Where wl ∈ Rm is the m dimension filter, wl ∈ Rm al(i+m−1 ):i =
each Res-block is shown in the Fig. 2. Signals can be transmitted
[al(i+m−1 ) , . . . , ali ]T , bi is offset Parameters, i = 1, . . . , n. wl is the directly from one unit to other units to prevent gradients from dis-
same for all neurons in the convolutional layer. appearing and simplify training.
Res-block1, Res-block2, and Res-block3 also use Relu as the ac-
2.3. Inception-ResNet structure tivation function, and the BN and Relu pre-processed sequential
connections following the convolutional layer are used to scale and
There are three most important factors when building a net- normalize the output values of the convolutional layer.
work: depth, width, and filter size [45]. The traditional CNN im-
provement method only emphasizes the improvement of network 2.4. Artifacts removal by using the proposed 1D-ResCNN model
feature processing ability by deepening the network layer. Due to
the gradient explosion problem and degradation phenomenon, it is Assume that the noisy EEG signal yi consists of pure EEG signals
difficult for the very deep network to train to obtain the desired xi and noise ni .
result. Therefore, ResNet is introduced to solve this problem [46].
yi = xi + ni (5)
Zagoruyko and Komodakis [47] proves that a wide and shallow
network model can outperform a thin and deep network model, The purpose of EEG denoising is to obtain an estimate x˜i of xi
such as ResNet. Moreover, when selecting the filter size, the single- under the condition of known yi , i is the sample index number.
scale convolution kernel cannot achieve the utilization of multi- The basic idea of denoising based on deep learning EEG can be de-
scale features. scribed as: constructing a highly complex nonlinear function f()
112 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121

Fig. 4. The flow chart of the 1D-ResCNN denoising model mainly includes training and testing. In the training phase, the network automatically learns the nonlinear
and discriminative deep features in the noisy EEG data collected. Finally, the network utilizes These features that distinguish the EEG signal and noise are automatically
reconstructed to obtain a clean EEG signal. In the test phase, the test focused noisy EEG signal is input into the trained 1D-ResCNN network to obtain a clean EEG signal.

Fig. 5. A typical contaminated EEG signal for 8s.The signal is divided into 10 windows in the same time period, and the excess is removed. Noise mainly exists in window
1, window 3, window 5, window 7 and window 9, while the EEG segments in the remaining windows can be considered to contain no noise.

by training network parameter sets to minimize the following er- 3. Dataset

ror function.
The 1D-ResCNN model aims to learn the mapping function All datasets used in this paper are contained within the
f (y) to reconstruct the EEG signal. The loss function of this pa- CHB-MIT database, which can be downloaded from the Phys-
per uses the traditional mean square error (MSE) method to train ioNet website (http://www.physionet.org). Our paper uses 20
1D-ResCNN by small batch gradient descent method and minimize EEG datasets from 23 active electrodes (FP1-F7,F7-T7,T7-P7,P7-
the loss function by Adam optimization method. O1,FP1-F3,F3-C3,C3-P3,P3-O1,FP2-F4,F4-C4,C4-P4,P4-O2,FP2-F8,F8-
T8,T8-P8,P8-O2,FZ-CZ,CZ-PZ,P7-T7,T7-FT9,FT9-FT10,FT10-T8,T8-P8)
1
n
L() =  f (yi ) − xi 22 (6) recorded, collected at Boston Children’s Hospital, including pae-
n diatric patients from intracTable seizures EEG recording. Subjects
i=1
were monitored for several days after stopping anti-epileptic
where n is the number of training samples, i.e. the number of sam- drugs to characterize their seizures and assess their eligibility for
ples per training. The loss function can be used to learn the train- surgical intervention. All signals are sampled at 256 samples per
able parameter . second with a resolution of 16 bits.This paper presents the exper-
In practice, we developed a 5-step 1D-ResCNN model training imental results of 20 EEG data, respectively, chb01-chb21 (except
process, as shown in Algorithm 1, and the overall flow chart is chb19).
shown in Fig. 4. The noise dataset is derived from EMG, ECG, and EOG three
datasets, which can be downloaded from the PhysioNet website
Algorithm 1: Artifacts removal by using the proposed 1D- (http://www.physionet.org). The clean segmentation was first per-
ResCNN. formed on the EEG data of 20 patients, and then the clean seg-
ment and the three types of noise were synthesized respectively
1 The sliding window method is used for segmentation
using seven different SNRs of 5 dB, 3 dB, 1 dB, 0 dB, -3 dB, and -
preprocessing;
5 dB.The specific synthesis method of each segment of noisy brain
2 Normalize the noisy EEG, the normalization formula is as
electricity is as follows: Each time one segment is randomly se-
follows:
xi − μ lected from pure EEG, and one class is randomly selected from the
xˆi = (4 )
σ three types of noise, then the random intercepted segment of the
noise is selected from the seven SNR.One of the randomly selected
where μ is the mean of all sample data, σ is the standard
ones is mixed into the EEG to build training database and test-
deviation of all sample data,and i is the number of samples ;
ing database consisting of pairs of clean EEG and noisy EEG sig-
3 Construct a one-dimensional convolutional neural network;
nals.Each patients data set contains 921600 × 23samples, 23 is the
4 Network model training process. The parameters of the
EEG electrode channel, and we adopt the sliding window method
1D-ResCNN model are adjusted through the training process
and adopt the window size of 400, the stride is 20 (%5 of 400), and
and trained to minimize the loss function of Equation 6;
each sample becomes 400 × 256 × 23, which will generate many
5 The EEG data of the test set is input into the trained
instances from one record. This method increases the amount of
1D-ResCNN model, and after normalization according to
data and meet the requirements of deep learning. The entire data
Equation 4, the denoised EEG signal is output;
set is divided into a training set (%80 data) and a test set (%20
data).
 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 113

Fig. 7. Comparison of spectrograms: (a) the noisy EEG signal of the ECG noise at
0 dB, and the speech enhanced by (b) 1D-ResCNN, (c) DNN, (d) WT, (e) FICA, (f)
ICA, (g) RLS filter and (h) the clean EEG signal.

is divided into 10 windows in the same time period, and the ex-
cess is removed. Fig. 5 shows that noise mainly exists in window 1,
window 3, window 5, window 7 and window 9, while the EEG seg-
ments in the remaining windows can be considered to contain no
noise. Calculate the kurtosis of the EEG segment in each window
and obtain the kurtosis value kw (w indicates the serial number of
the window). Then based on a certain amount of observation and
empirical analysis, set the threshold K to pick up the EEG segment.
If the EEG segment kurtosis kw exceeds the threshold K, this seg-
ment is eliminated. The remaining EEG segments are considered
uncontaminated EEG segments. In this way, we can get enough
clean EEG.

3.2. Noisy EEG sample

It is achieved by using a common stacking technique that

addsnoiseto the EEG to maintain SNR at 5 dB, 3 dB, 1 dB,
0 dB,1 dB,3 dB and 5 dB. The obtained pure EEG data (EEGc ) is
contaminated by additional ECG noise, and the amount of noise
is a part given by Eq. (7). Therefore, in order to generate a noisy
EEG (EEGt ), the fraction λ is adjusted such that the obtained SNR
according to Eq. (8) is a desired value.
E E Gn = E E Gc + λnoise (7)

RMS(λnoise )
SNR = (8)
RMS(E E Gc )
Where RMS is defined as the root mean square energy of the sig-
nal.
These synthetically contaminated EEG signals are further used
to correct the evaluation of the algorithm and change the intensity
of the contamination to verify the efficiency of the algorithm.

Fig. 6. An example of an experimental result waveform for eliminating noise from 3.2.1. Unknown noise test dataset
an EEG signal containing ECG noise of SNR = 0 dB. From top to bottom, the re-
In order to test the generalization ability of the EEG noise re-
sults are denoised by ICA, FICA, RLS filter, WT, DNN and the method proposed in
this paper. The dotted line represents the noisy EEG waveform, and the solid line duction model, an unknown noise test set is constructed. The two
represents the EEG waveform after using different denoising methods. types of noise in the unknown noise test set are unknown noises
completely different from the training set noise, namely Gaussian
White Noise (WGN), and electrode motion(EM). In order to make
3.1. Intercepting clean EEG training samples the experimental conditions more realistic, the noise type and SNR
level of the training and test sets do not match. The EEG signal of
In this paper, we use the EEG samples without noise to train the test set and the randomly intercepted segment of the 2 types
the 1D-ResCNN model. Inspired by [36], We used a clean EEG sig- of noise are mixed according to the global SNR of −10 dB, 0 dB
nal based on kurtosis interception. Fig. 5 is a typical contaminated and 10 dB, respectively, to obtain a test set of noisy brain electric-
EEG signal within 8 s (with sampling frequency 256 Hz).The signal ity with unknown noise.
114 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121

Fig. 9. Comparison of spectrograms: (a) the noisy EEG signal of the EMG noise at
0 dB, and the speech enhanced by (b) 1D-ResCNN, (c) DNN, (d) WT, (e) FICA, (f)
ICA, (g) RLS filter and (h) the clean EEG signal.

evaluation and nonlinear characteristics are used. In this paper, a

subjective performance analysis of the proposed technique is per-
formed by visual inspection.

4.1.1. Objective evaluation

This section describes the quantitative performance analysis
of the described techniques. Consider two performance metrics,
signal-to-noise ratio (SNR) and mean square error (MSE) for eval-
uation.The SNR defines the signal energy relative to the energy
of the associated noise. SNR is a related performance index for
quantifying the effect of noise reduction to reduce the influence
of background noise. It is defined as:
n
x (i )
2
SNR = 10 log n−1 i=1 (9)
i=0 [ (x (i ) − x (i )]
ˆ 2

where x(i) is the original EEG signal, then x(ˆi ) represents the de-
noised EEG signal.
The RMSE defines the energy of the error signal during noise
reduction. A lower value for RMSE means better estimation of the
original signal and better preservation of signal details.

n 2
i=1 [(x(i ) − x(ˆi )]
RMSE = (10)
n

4.1.2. Influence of denoising on nonlinear characteristics of signals

Our paper selects the nonlinear dynamics parameters of time
series complexity and statistical quantificationâpower spectral den-
sity (PSD), approximate entropy (ApEn) and autocorrelation func-
tion as nonlinear characteristics of EEG signals are used to mea-
sure the influence of this method on the nonlinear characteristics
of reconstructed EEG signals.

Fig. 8. An example of an experimental result waveform for eliminating noise from 4.2. Results and discussion
an EEG signal containing EMG noise of SNR = 0 dB. From top to bottom, the re-
sults are denoised by ICA, FICA, RLS filter, WT, DNN and the method proposed in
this paper. The dotted line represents the noisy EEG waveform, and the solid line To further verify the denoising performance of the proposed de-
represents the EEG waveform after using different denoising methods. noising method, it is compared with ICA [28], FICA [50], RLS filter
[51], WT [52] and DNN.

4. Experiments 4.2.1. ECG artifact elimination effect

Fig. 6 shows the experimental results of removing ECG noise
4.1. Performance indicators in SNR = 0dB condition from the EEG signal. The dashed line indi-
cates the EEG signal with ECG noise, and the solid line indicates
In order to evaluate the performance of the proposed method, the EEG signal after denoising. From top to bottom, the denoising
the following evaluation indicators, subjective evaluation, objective results using the proposed method (1D-ResCNN), DNN, WT, ICA,
 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 115

Fig. 11. Comparison of spectrograms: (a) the noisy EEG signal of the EOG noise at
0 dB, and the speech enhanced by (b) 1D-ResCNN, (c) DNN, (d) WT, (e) FICA, (f)
ICA, (g) RLS filter and (h) the clean EEG signal.

has restored to the normal waveform. The WT denoising method

eliminates ECG noise, but the corrected EEG waveform is deformed
compared to the original EEG. The removal of ECG artifacts af-
ter denoising by ICA, FICA, and RLS filter methods is incomplete.
In addition, the corrected EEG details are severely distorted com-
pared to the original EEG. In the spectrum analysis view, the ab-
scissa is time, the ordinate is frequency, and the coordinate point
is the energy of the EEG signal data. Since the two-dimensional
plane is used to represent three-dimensional information, the mag-
nitude of the energy value is represented by color, and the color is
dark, indicating that the energy of the point is stronger. Fig. 7((a)–
(h)) present the spectrograms of noisy EEG signal mixed with ECG
noise at 0 dB,and EEG denoising by the 1D-ResCNN, DNN, WT,
FICA,ICA, RLS filter methods and clean EEG signal, respectively.It
is obvious that FICA, ICA, and RLS filter could not effectively re-
move the noise components. After the WT method noise compo-
nents can still be observed. In contrast, the DNN model and our
proposed model in this paper effectively suppressed most of the
noise components, and the EEG spectrum after the proposed model
denoising is closest to the clean EEG spectrum.

4.2.2. EMG Artifact elimination effect

Fig. 8 shows the experimental results of removing EMG noise
from the EEG signal. The dashed line indicates the EEG signal
with EMG noise, the SNR is 0 dB, and the solid line indicates
the EEG signal after denoising. The denoising results of the pro-
posed method (1D-ResCNN), DNN, WT, ICA, FICA and RLS filter
are shown in turn. It has been observed that 1D-ResCNN basically
eliminates EMG noise and is better, and some low-frequency detail
features with small amplitudes are well restored and maintained.
The waveform details are preserved, and the DNN eliminates most
of the EMG noise, but the noise-reduced EEG waveform has a cer-
Fig. 10. An example of an experimental result waveform for eliminating noise from tain up and down float compared to the original EEG. After the
an EEG signal containing EOG noise of SNR = 0 dB. From top to bottom, the re- noise is reduced by the WT, ICA, FICA and RLS filter methods, the
sults are denoised by ICA, FICA, RLS filter, WT, DNN and the method proposed in
corrected EEG morphology changes.
this paper. The dotted line represents the noisy EEG waveform, and the solid line
represents the EEG waveform after using different denoising methods.
Fig. 9 is a spectrum analysis diagram before and after noise re-
duction for various noise reduction methods for EMG noise with a
signal-to-noise ratio of 0 dB. Fig. 9((a)–(h)) gives EEG with SNR =
FICA, and RLS filter methods are shown in the figure. As can be 0dBEMG noise, the EEG signal denoised by (b)1D-ResCNN, (c) DNN,
seen from Fig. 6, although the DNN method is similar to the 1D- (d) WT, (e) FICA, (f) ICA, (g)RLS filter method and (h) spectrogram
ResCNN, the corrected EEG waveform after DNN is floating com- of clean EEG. The DNN model and the model proposed in this pa-
pared to the original EEG, and the 1D-ResCNN performs better in per effectively suppress most of the noise components, and the
reconstructing the signal details. Especially from the local details of EEG spectrum after denoising of the 1D-ResCNN model is closest
the noise-reduced EEG signal, it can be found that the EEG signal to the clean EEG spectrum.
116 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121

tion, the corrected EEG details are not preserved by these methods
compared to the original EEG.
Fig. 10 is the spectrum analysis of EOG noise removal for SNR =
0dB using various denoising methods. As the results show, the pro-
posed method and the noise-reduced DNN method are closer to
the clean EEG signal spectrum.
By removing ECG, EMG and EOG different noise waveform com-
parison graphs Figs. 6, 8 and 10 and spectrum analysis graphs
Figs. 7, 9, and 11, we can see that the proposed method performs
well for all three noise elimination, and DNN performs better for
the elimination of EOG noise.

4.3. Performance analysis

4.3.1. Performance analysis based on signal to noise ratio

In this section, the performance of the proposed model is ana-
lyzed based on the SNR.These noises for each of the different SNRs
were tested on 20 patients, and each test sample was repeated 10
times and averaged.
Table 2, 3 and 4 respectively show the use of the proposed
method (1D-ResCNN), DNN, WT, ICA, FICA And RLS filter meth-
ods remove SNR values for ECG noise, EMG noise, and EOG noise.
Fig. 12(a)–(c) show the SNR values of three noise denoising meth-
ods for ECG, EMG and EOG with different signal-to-noise ratios us-
ing several different noise reduction methods.The WT, ICA, FICA,
and RLS filter methods were found to provide minor improvements
in SNR, and the DNN approach provides further enhancements. For
the three kinds of noise cancellation, the SNR value of the pro-
posed model is significantly improved compared with other meth-
ods, which indicates that the proposed algorithm can filter noise
more effectively.

4.3.2. Performance analysis based on root mean square error (RMSE)

value
Table 5, 6 and 7 illustrate the RMSE analysis of EEG signals
with various artifacts (ECG, EMG and EOG). Fig. 13(a)–(c) shows
the RMSE values with different SNR ECG, EMG and EOG noise de-
noised using several different methods of noise reduction. Usu-
ally, The minimum of RMSE can produce better performance. As
can be seen, compared to other existing models such as ICA, FICA,
WT, and RLS filter, the proposed model achieves a minimum RMSE
value compared to DNN.

4.3.3. Influence of denoising on nonlinear characteristics of signals

Fig. 14 shows the PSD of contaminated EEG and denoising EEG
treated by ICA, FICA,RLS filter, WT, DNN and the methods de-
scribed herein, respectively. Fig. 14(a)–(c) are PSDs for EEG for var-
ious methods, including ECG noise, EMG noise, and EOG noise, re-
spectively. As can be seen from the three figures, the PSD value
of the EEG signal after noise reduction by the six methods is de-
creasing in the concentrated frequency range of the noise. And the
performance of this method and DNN is slightly better than ICA,
FICA, RLS filter and WT.
The smaller the difference between the ApEn values of the
EEG before and after 1D-ResCNN processing, the better the perfor-
Fig. 12. SNR results of remove noise from EEG signal.
mance of the denoising model. The Table 9 shows that among the
six methods, 1D-ResCNN provides the minimum value. This indi-
4.2.3. EOG artifact elimination effect cates that the artifact removal process of 1D-ResCNN preserves the
Fig. 10 shows the waveforms of the method (1D-ResCNN), DNN, characteristics of the EEG.
WT, ICA, FICA and RLS filter methods proposed by removing the The results in Table 8 show that our approach produces
SNR noise of SNR = 0dB. The dashed line indicates the EEG signal the greatest cross-correlation between cleaning EEG and noise-
with EMG noise, the SNR is 0 dB, and the solid line indicates the reducing EEG. Our approach significantly reduces the correlation
EEG signal after denoising. It was observed that both 1D-ResCNN between artifacts and EEG signals. This reduction indicates that
and DNN recovered the morphological characteristics of the orig- EEG artifacts in the noisy EEG signal are substantially eliminated.
inal signal well. After noise reduction by the WT, ICA, FICA and The denoising algorithm 1D-ResCNN has a significant effect in re-
RLS filter methods, EOG noise removal is incomplete and, in addi- moving EEG artifacts. This means that almost all of the original
 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 117

Table 2
SNR experimental results of ECG noise with different SNR.

Evaluation index SNR −5dB −3dB −1dB 0dB 1dB 3dB 5dB

ICA 4.789 6.025 6.842 7.072 7.173 7.925 8.817

FICA 8.844 11.337 12.078 12.437 12.98 13.14 13.334
RLS filter 9.894 11.763 13.296 13.529 13.618 13.63 13.737
WT 12.083 14.094 15.852 16.66 16.853 16.792 17.024
DNN 17.933 19.378 19.957 20.394 21.079 22.319 23.393
Proposed method 19.826 22.863 23.526 23.957 24.017 24.337 24.675

Table 3
SNR experimental results of EMG noise with different SNR.

Evaluation index SNR −5dB −3dB −1dB 0dB 1dB 3dB 5dB

ICA 4.872 4.483 5.47 5.93 6.204 6.9110 7.864

FICA 9.049 9.232 9.508 10.047 11.348 11.906 12.856
RLS filter 9.973 10.094 10.25 11.163 11.793 12.355 12.593
WT 13.021 13.116 13.62 14.725 14.852 15.147 15.454
DNN 17.203 18.527 18.935 19.577 19.729 20.728 22.602
Proposed method 19.338 21.743 23.837 24.484 25.065 25.711 26.472

Table 4
SNR experimental results of EOG noise with different SNR.

Evaluation index SNR −5dB −3dB −1dB 0dB 1dB 3dB 5dB

ICA 5.784 6.097 6.669 6.7113 6.885 7.273 7.464

FICA 9.433 9.976 10.097 10.502 10.976 11.137 11.397
RLS filter 10.078 10.667 11.796 12.293 13.078 13.667 13.792
WT 14.042 14.649 14.904 15.974 16.769 16.986 16.982
DNN 20.03 20.59 21.21 21.941 22.43 22.559 22.95
Proposed method 20.63 20.96 21.46 21.964 22.47 22.57 23.03

Table 5
Add RMSE experimental results of ECG noise.

Evaluation index RMSE −5dB −3dB −1dB 0dB 1dB 3dB 5dB

ICA 0.6083 0.5870 0.5253 0.4889 0.4851 0.4775 0.4283

FICA 0.3738 0.3706 0.3675 0.3608 0.3537 0.3502 0.3485
RLS filter 0.3818 0.3794 0.3717 0.3663 0.3606 0.3578 0.3146
WT 0.3486 0.3399 0.3306 0.3292 0.3239 0.3094 0.2992
DNN 0.1749 0.1727 0.1686 0.1605 0.1578 0.1463 0.1392
Proposed method 0.1118 0.0946 0.0874 0.08194 0.08021 0.07869 0.0754

Table 6
Add RMSE experimental results of EMG noise.

Evaluation index RMSE −5dB −3dB −1dB 0dB 1dB 3dB 5dB

ICA 0.5812 0.6071 0.6213 0.6194 0.6401 0.5982 0.5797

FICA 0.4085 0.4276 0.4406 0.4369 0.4613 0.4276 0.4089
RLS filter 0.3697 0.3895 0.4078 0.3913 0.4223 0.4102 0.3687
WT 0.3762 0.3916 0.3796 0.3667 0.3998 0.3951 0.3368
DNN 0.2033 0.2239 0.2072 0.1954 0.2265 0.2775 0.2315
Proposed method 0.1092 0.0985 0.0903 0.0896 0.0851 0.0841 0.0796

Table 7
Add RMSE experimental results of EOG noise.

Evaluation index RMSE −5dB −3dB −1dB 0dB 1dB 3dB 5dB

ICA 0.5857 0.5779 0.5673 0.5549 0.5471 0.5373 0.5324

FICA 0.5009 0.4963 0.4935 0.4838 0.4803 0.4773 0.4756
RLS filter 0.4677 0.4518 0.4477 0.4419 0.4335 0.4206 0.4153
WT 0.3942 0.3857 0.3842 0.3770 0.3784 0.3511 0.3581
DNN 0.1005 0.0914 0.0856 0.0764 0.0756 0.0710 0.0687
Proposed method 0.0948 0.0852 0.0815 0.0775 0.0761 0.0699 0.0680

EEG information remains in the filtered EEG and these artifacts are U test [53] was used in this study. It assumes that the two
almost completely eliminated. samples are from the two identical populations except the
population mean, if the significance level is less than 0.05,
4.3.4. Hypothesis test then assume (two Overall equal) will be rejected (p value
In order to determine whether there is a significant differ- < 0.05). Tables 10–12 respectively show RMSE, SNR and ApEn
ence between the results of noise reduction by these meth- of Mann–Whitney U test result with different noise denoised
ods, the non-parametric statistical hypothesis test Mann–Whitney signals.
118 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121

Table 8
Correlation between EEG signal and true EEG signal.

Evaluation index correlation ICA FICA RLS filter WT DNN Proposed method

Before removing ECG 0.006 0.006 0.006 0.006 0.006 0.006

After removing ECG 0.1153 0.2290 0.4388 0.5839 0.7897 0.89240
Before removing EMG 0.1437 0.1437 0.1437 0.1437 0.1437 0.1437
After removing EMG 0.3861 0.4116 0.4920 0.5189 0.8324 0.9117
Before removing EOG −0.05 −0.05 −0.05 −0.05 −0.05 −0.05
After removing EOG 0.1336 0.3747 0.4694 0.6418 0.9231 0.9567

Table 9
The ApEn comparison between Clean EEG and removed artifact from EEG signal.

Evaluation index ApEn ICA FICA RLS filter WT DNN Proposed method

Before removing ECG 0.59 0.59 0.59 0.59 0.59 0.59

After removing ECG 0.75( ± %27.1) 0.73(+ %23.7) 0.71(+%16.4) 0.69(+%16.9) 0.62(+%5.1) 0.60(+%1.7)
Before removing EMG 0.56 0.56 0.56 0.56 0.56 0.56
After removing EMG 0.77(+%37.5) 0.73(+%30.3) 0.71(+%26.8) 0.67(+%19.4) 0.60(+%7.14) 0.58(+%3.57)
Before removing EOG 0.67 0.67 0.67 0.67 0.67 0.67
After removing EOG 0.91(+%35.8) 0.83(+%23.9) 0.78(+%16.4) 0.72(+%7.5) 0.70(+%4.8) 0.68(+%1.4)

Table 10
Hypothesis test result of ECG noise removel.

RMSE SNR ApEn

MODEL
P-values Null hypothesis P-values Null hypothesis P-values Null hypothesis

DNN 0.00108 Reject 0.0107 Reject 0.01079 Reject

1D-ResCNN
WT 0.00016 Reject 0.00022 Reject 0.00609 Reject
1D-ResCNN
RLS filter 0.00015 Reject 0.00021 Reject 0.00609 Reject
1D-ResCNN
FICA 0.00012 Reject 0.00020 Reject 0.00609 Reject
1D-ResCNN
ICA 0.00010 Reject 0.00020 Reject 0.00609 Reject
1D-ResCNN

Table 11
Hypothesis test result of EMG noise removel.

RMSE SNR ApEn

MODEL
P-values Null hypothesis P-values Null hypothesis P-values Null hypothesis

DNN 0.00759 Reject 0.02026 Reject 0.01063 Reject

1D-ResCNN
WT 0.00108 Reject 0.00179 Reject 0.007013 Reject
1D-ResCNN
RLS filter 0.00108 Reject 0.00179 Reject 0.007012 Reject
1D-ResCNN
FICA 0.00108 Reject 0.00179 Reject 0.007012 Reject
1D-ResCNN
ICA 0.00108 Reject 0.00179 Reject 0.007013 Reject
1D-ResCNN

Table 12
Hypothesis test result of EOG noise removel.

RMSE SNR ApEn

MODEL
P-values Null hypothesis P-values Null hypothesis P-values Null hypothesis

DNN 0.10365 Accept 0.07062 Accept 0.01779 Reject

1D-ResCNN
WT 0.00647 Reject 0.00578 Reject 0.00760 Reject
1D-ResCNN
RLS filter 0.00647 Reject 0.00578 Reject 0.00760 Reject
1D-ResCNN
FICA 0.00647 Reject 0.00578 Reject 0.00760 Reject
1D-ResCNN
ICA 0.00647 Reject 0.00578 Reject 0.00760 Reject
1D-ResCNN
 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 119

Fig. 14. The PSD result of eliminating ECG noise from the EEG signal, from the top
and bottom, by the ICA, FICA, RLS filter, WT, DNN and the algorithm proposed in
this paper, the EEG contaminated by the ECG, the EEG contaminated by the EMG,
by the EOG Contaminated EEG and corrected ESG PSD.

Fig. 15. Run-time bar graphs for different denoising methods ICA, FICA, RLS filter,
Fig. 13. RMSE results of remove noise from EEG signal. WT, DNN and the methods proposed in this paper.

Hypothesis test results show that there is a significant differ-

ence between 1D-ResCNN and other methods for remove EMG and 4.3.5. Time consumption
ECG noise. There is no significant difference between the DNN Time consumption is an important feature to test the effective-
method and the proposed method in terms of the RMSE and SNR ness of the algorithm. Fig. 15 gives the processing time for 200 tri-
of the denoising performance metrics for removing EOG noise, but als in different methods (each test contains 23 EEG channels, each
the test result of nonlinear characteristic index ApEn shown that of which contains 400 samples of EEG data). Fig. 15 shows that the
there are significant differences between the proposed method and method proposed in this article saves more time than ICA, FICA,
the DNN method. It is shown the proposed method can retain RLS filter, and WT. Moreover, the DNN method saves a little time
more nonlinear features than DNN. compared to the method of this paper.
120 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121

Table 13 5. Conclusion
The average SNR for DNN and 1D-ResCNN methods.

Noise type SNR(dB) Noisy EEG DNN 1D-ResCNN In order to further improve the EEG denoising performance un-
GWN −10 −11.90 −0.69 −0.05 der unknown noise, CNN has better local feature expression abil-
0 −4.50 0.34 0.95 ity and can better utilize the correlation between EEG signal and
10 4.43 6.24 6.80 noise signal. In this paper, a reasonable network structure suiTable
EM −10 −12.21 −1.40 −0.96 for EEG denoising is proposed. An EEG denoising method based
0 −4.90 −0.39 0.11
on deep one-dimensional residual convolutional neural network is
10 3.97 5.53 7.03
proposed. The deep convolutional neural network is used to estab-
lish a regression model to express the complex nonlinear relation-
Table 14
ship between noisy EEG signals and pure EEG signals.Convolutional
The average RMSE for DNN and 1D-ResCNN methods.
networks filter residual data, use large data sets to improve sig-
Noise type SNR(dB) DNN 1D-ResCNN nal denoising, and help capture the main features of EEG sig-
GWN −10 0.4992 0.3847 nals. A large number of experimental results show that the one-
0 0.4820 0.3824 dimensional convolutional neural network proposed in this paper
10 0.4624 0.3718 achieves a smaller RMSE and better signal-to-noise ratio on the
EM −10 0.5503 0.5116
test set, and has better noise suppression ability than other de-
0 0.5416 0.5020
10 0.5371 0.4814 noising methods. Under various noise and various signal-to-noise
ratios, the nonlinear characteristics of EEG after denoising are sig-
nificantly maintained, and the EEG denoising performance under
unknown noise is further improved.

Declaration of Competing Interest

The authors declare that they have no known competing finan-

cial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Acknowledgments

This work is partially supported by the National Key Research

and Development Program of China (No. 2017YFB1402102), the
National Natural Science Foundation of China (Nos. 11772178,
11502133, 11872036, 61701291),the Fundamental Research Fund
for the Central Universities (Nos. 2018CBLY007, GK201801004),
the China Postdoctoral Science Foundation funded project (No.
2017M613053) and the Shaanxi Natural Science Foundation Project
under grant (No. 2018JQ6089).
Fig. 16. An example of an experimental result waveform that eliminates noise from
an EEG signal that includes GWN noise with SNR = −10 dB. In the first figure, the
References
dotted line indicates the noisy EEG waveform, and the solid line indicates the clean
EEG waveform. The second and third figures show the denoising of noisy brain [1] A. Turnip, E. Junaidi, Removal artifacts from eeg signal using independent com-
power using DNN and the method proposed in this paper. The dotted line indi- ponent analysis and principal component analysis, in: International Conference
on Technology, Informatics, Management, Engineering, and Environment, IEEE,
cates the noisy EEG waveform, and the solid line indicates the EEG waveform after
2014, pp. 296–302.
using different denoising methods.
[2] L.J. Hirsch, R.P. Brenner, Atlas of EEG in critical care, John Wiley & Sons, 2011.
[3] U. Seneviratne, A. Mohamed, M. Cook, W. D’Souza, The utility of ambulatory
electroencephalography in routine clinical practice: a critical review., Epilepsy
4.3.6. Unknown noise removl result Res. 105 (1–2) (2013) 1–12.
In order to further compare the noise reduction performance [4] M.K.I. Molla, M.R. Islam, T. Tanaka, T.M. Rutkowski, Artifact suppression from
eeg signals using data adaptive time domain filtering, Neurocomputing 97 (97)
of the unknown noise of DNN and 1D-ResCNN, we compare the (2012) 297–308.
EEG obtained by the noise reduction of the unkonwn test dataset [5] C.K. Kovach, T. Naotsugu, K. Hiroto, O. Hiroyuki, M.A. Howard, A. Ralph, Man-
by two methods. Tables 13 and 14 respectively show the SNR and ifestation of ocular-muscle EMG contamination in human intracranial record-
ings, Neuroimage 54 (1) (2011) 213–233.
RMSE of EEG after denoised, and give the SNR of unprocessed [6] L. Sörnmo, P. Laguna, Bioelectrical Signal Processing in Cardiac and Neurologi-
noisy EEG as a comparison. It can be seen that by using multi- cal Applications, Academic Press, 2005.
ple types of noise for training, for two kinds of The two types [7] Y. Yu, Z. Zhou, E. Yin, J. Jiang, J. Tang, Y. Liu, D. Hu, Toward brain-actuated car
applications: self-paced control with a motor imagery-based brain-computer
of trained noise types can effectively remove EEG artifacts, and interface, Comput. Biol. Med. 77 (2016) 148–155.
among the two different indicators, 1D-ResCNN achieves better re- [8] M.K. Islam, A. Rastegarnia, Z. Yang, Methods for artifact detection and re-
sults than DNN under different noise types and different SNR. moval from scalp eeg: a review, Neurophysiol. Clin.-Clin. Neurophysiol. 46
(4–5) (2016) 287–305.
In order to more intuitively compare the EEG noise reduction
[9] M.A. Klados, C. Papadelis, C. Braun, P.D. Bamidis, REG-ICA: a hybrid methodol-
performance of the two methods, we use two methods to perform ogy combining blind source separation and regression techniques for the rejec-
EEG noise reduction on a noisy EEG with Gaussian white noise SNR tion of ocular artifacts, Biomed. Signal Process. Control 6 (3) (2011) 291–300.
[10] A. Schlögl, C. Keinrath, D. Zimmermann, A fully automated correction method
of −10dB, and then compare EEG waveform after noise reduction.
of EOG artifacts in eeg recordings., Clin. Neurophysiol. 118 (1) (2007) 98–104.
Fig. 16 respectively show the waveforms of the noisy EEG and its [11] G.L. Wallstrom, R.E. Kass, A. Miller, J.F. Cohn, N.A. Fox, Automatic correction of
corresponding EEG after DNN and 1D-ResCNN noise reduction. It ocular artifacts in the eeg: a comparison of regression-based and componen-
can be seen that 1D-ResCNN reduces the residual noise component t-based methods, Int. J. Psychophysiol. 53 (2) (2004) 105–119.
[12] A.G. Correa, E. Laciar, H. Patiño, M. Valentinuzzi, Artifact removal from EEG
of EEG signals after noise reduction, and its waveform is closer to signals using adaptive filters in cascade, in: Journal of Physics: Conference Se-
the waveform of pure EEG. ries, 90, IOP Publishing, 2007, p. 12081.
 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 121

[13] A. Jafarifarmand, M.A. Badamchizadeh, Artifacts removal in eeg signal using a [41] Y.R. Tabar, U. Halici, A novel deep learning approach for classification of EEG
new neural network enhanced adaptive filter, Neurocomputing 103 (2) (2013) motor imagery signals, J. Neural Eng. 14 (1) (2016) 16003.
222–231. [42] T. Wen, Z. Zhang, Deep convolution neural network and autoencoder-
[14] J.E. Meng, Y. Gao, An adaptive RBFN-based filter for adaptive noise cancellation, s-based unsupervised feature learning of eeg signals, IEEE Access 6 (2018)
in: IEEE Conference on Decision and Control, 2003. Proceedings, 6, IEEE, 2003, 25399–25410.
pp. 6175–6180. [43] I. Ullah, M. Hussain, E.U.H. Qazi, H. Aboalsamh, An automated system for
[15] L. Fan, E.M. Joo, Learning algorithm for constructing fuzzy neural networks epilepsy detection using eeg brain signals based on deep learning approach,
with application to regression problems, in: International Conference on In- Expert Syst. Appl. 107 (2018) 61–71.
formation Science and Technology, 2011, pp. 318–322. [44] S. Ioffe, C. Szegedy, in: Batch Normalization: Accelerating Deep Network Train-
[16] C.K.S. Vijilal, P. Kanagasabapathy, S. Johnson, V. Ewards, Artifacts removal in ing by Reducing Internal Covariate Shift, 2015, pp. 448–456.
EEG signal using adaptive neuro fuzzy inference system, in: International [45] K. He, S. Jian, Convolutional neural networks at constrained time cost, in: Com-
Conference on Signal Processing, Communications and NETWORKING, 2007, puter Vision & Pattern Recognition, 2015.
pp. 589–591. [46] K. He, X. Zhang, S. Ren, J. Sun, in: Deep Residual Learning for Image Recog-
[17] J. Hu, C.S. Wang, M. Wu, Y.X. Du, Y. He, J. She, Removal of EOG and EMG arti- nition, In Proceedings of the IEEE conference on computer vision and pattern
facts from eeg using combination of functional link neural network and adap- recognition, 2016, pp. 770–778.
tive neural fuzzy inference system? Neurocomputing 151 (2015) 278–287. [47] S. Zagoruyko, N. Komodakis, in: Wide Residual Networks, 2016.
[18] M.K. Ahirwal, A. Kumar, G.K. Singh, Adaptive filtering of EEG/ERP through [48] Y. Pang, M. Sun, X. Jiang, X. Li, Convolution in convolution for network in net-
noise cancellers using an improved PSO algorithm, Swarm Evolut. Comput. 14 work, IEEE Trans. Neural Netw. Learn.Syst. 29 (5) (2017) 1587–1597.
(2014) 76–91. [49] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception
[19] H.A.T. Nguyen, J. Musson, F. Li, W. Wang, G. Zhang, R. Xu, C. Richey, T. Schnell, architecture for computer vision, in: Computer Vision & Pattern Recognition,
F.D. Mckenzie, J. Li, EOG artifact removal using a wavelet neural network, Neu- 2016, pp. 2818–2826.
rocomputing 97 (1) (2012) 374–389. [50] A. Hyvarinen, Fast and robust fixed-point algorithms for independent compo-
[20] B. Noureddin, P.D. Lawrence, G.E. Birch, Time-frequency analysis of eye blinks nent analysis, IEEE Trans. Neural Netw. 10 (3) (1999) 626–634.
and saccades in EOG for EEG artifact removal, in: International IEEE/EMBS Con- [51] D.P. He, G. Wilson, C. Russell, Removal of ocular artifacts from electro-en-
ference on Neural Engineering, 2007, pp. 564–567. cephalogram by adaptive filtering, Med. Biol. Eng. Comput. 42 (3) (2004)
[21] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.-C. Yen, 407–412.
C.C. Tung, H.H. Liu, The empirical mode decomposition and the hilbert spec- [52] V. Krishnaveni, S. Jayaraman, L. Anitha, K. Ramadoss, Removal of ocular arti-
trum for nonlinear and non-stationary time series analysis, in: Proceedings of facts from eeg using adaptive thresholding of wavelet coefficients, J. Neural
the Royal Society of London A: Mathematical, Physical and Engineering Sci- Eng. 3 (4) (2006) 338.
ences, 454, The Royal Society, 1998, pp. 903–995. [53] M.P. Fay, M.A. Proschan, Wilcoxon-Mann–Whitney or t-test? On assumptions
[22] K.T. Sweeney, S.F. Mcloone, T.E. Ward, The use of ensemble empirical mode for hypothesis tests and multiple interpretations of decision rules, Stat. Surv. 4
decomposition with canonical correlation analysis as a novel artifact removal (2010) 1–39.
technique, IEEE Trans. Biomed. Eng. 60 (1) (2012) 97–105.
[23] A. Egambaram, N. Badruddin, V.S. Asirvadam, T. Begum, Comparison of enve- Weitong Sun received B.S. degree in 2014. Now, she is
lope interpolation techniques in empirical mode decomposition (EMD) for eye- pursuing Ph.D. Degree in Computer Software and Theory
blink artifact removal from EEG, in: Biomedical Engineering and Sciences, IEEE, from School of Computer Science, Shaanxi Normal Univer-
2016, pp. 590–595. sity. Her research interests include nonlinear dynamics,
[24] J. Li, C. Li, N. Thakor, A. Cichocki, A. Bezerianos, J. Li, C. Li, N. Thakor, chaotic systems, signal processing and machine learning.
A. Cichocki, A. Bezerianos, A new perspective of noise removal from EEG,
in: International IEEE/EMBS Conference on Neural Engineering, IEEE, 2017,
pp. 501–504.
[25] N. Siew-Cheok, R. Paramesran, Enhanced ${\mu} $ rhythm extraction using
blind source separation and wavelet transform, IEEE Trans. Biomed. Eng. 56
(8) (2009) 2024–2034.
[26] L. Shoker, S. Sanei, M.A. Latif, Removal of eye blinking artifacts from eeg in-
corporating a new constrained BSS algorithm, in: Sensor Array & Multichannel
Signal Processing Workshop, IEEE, 2004, pp. 177–181. Yuping Su, received the Ph.D. degree from the school
[27] K. Ting, P. Fung, C. Chang, F. Chan, Automatic correction of artifact from sin- of Communication Engeering, Xi’dian University, Xi’an,
gle-trial event-related potentials by blind source separation using second order China, in 2015.She is currently a lecturer of the School of
statistics only, Med. Eng. Phys. 28 (8) (2006) 780–794. Computer and Science, Shaanxi Normal University, Xi’an,
[28] Y. Li, Z. Ma, W. Lu, Y. Li, Automatic removal of the eye blink artifact from eeg China.Her research interests include Data mining and Ma-
using an ICA-based template matching approach, Physiol Meas 27 (4) (2006) chine Leaning.
425–436.
[29] T.-P. Jung, S. Makeig, A.J. Bell, T.J. Sejnowski, Independent component analysis
of electroencephalographic and event-related potential data, Adv. Neural Inf.
Process. Syst. 8 (8) (1996) 145–151.
[30] Y. Chen, Q. Zhao, B. Hu, J. Li, H. Jiang, W. Lin, Y. Li, S. Zhou, H. Peng, A
method of removing ocular artifacts from eeg using discrete wavelet transform
and Kalman filtering, in: IEEE International Conference on Bioinformatics and
Biomedicine, IEEE, 2016, pp. 1485–1492. Xia Wu received B.S. degree in Mathematics and Applied
[31] C. Burger, D.J.V.D. Heever, Removal of eog artefacts by combining wavelet Mathematics from college of Science, Xidian University
neural network and independent component analysis, Biomed. Signal Process. and M.S. degree from Key Laboratory of Modern Teach-
Control 15 (15) (2015) 67–79. ing technology, Ministry of Education, Shaanxi Normal
[32] P.S. Kumar, R. Arumuganathan, K. Sivakumar, C. Vimal, Removal of artifacts University. Now, she is pursuing Ph.D. Degree in Com-
from EEG signals using adaptive filter through wavelet transform, in: Inter- puter Software and Theory from School of Computer Sci-
national Conference on Signal Processing, IEEE, 2008, pp. 2138–2141. ence, Shaanxi Normal University. Her research interests
[33] Y. Xu, J. Du, L.R. Dai, C.H. Lee, An experimental study on speech enhancement include nonlinear dynamics, chaotic systems, signal pro-
based on deep neural networks, IEEE Signal Process. Lett. 21 (1) (2013) 65–68. cessing and machine learning.
[34] R. Rodrigues, P. Couto, A neural network approach to ECG denoising, Comput.
Sci. (2012).
[35] P. Xiong, H. Wang, M. Liu, S. Zhou, Z. Hou, X. Liu, ECG signal enhancement
based on improved denoising auto-encoder, Eng. Appl. Artif. Intell. 52 (C)
(2016) 194–202.
[36] B. Yang, K. Duan, C. Fan, C. Hu, J. Wang, Automatic ocular artifacts removal in WU Xiaojun (corresponding author)is a professor in
eeg using deep learning, Biomed. Signal Process. Control 43 (2018) 148–158. Shaanxi Normal University, Xi’an, China. His research in-
[37] Z. Zhang, F. Xing, H. Su, X. Shi, L. Yang, in: Recent Advances in the Applications terests include pattern recognition, intelligent system and
of Convolutional Neural Networks to Medical Image Contour Detection, 2017. system complexity.
[38] S. Bartunov, A. Santoro, B.A. Richards, G.E. Hinton, T. Lillicrap, in: Assess-
ing the Scalability of Biologically-motivated Deep Learning Algorithms and
Architectures, In Advances in Neural Information Processing Systems, 2018,
pp. 9368–9378.
[39] T. Jiang, Y. Zhang, X. Wu, G. Lu, F. Hao, Y. Zhang, Single image super resolution
via a refined densely connected inception network, in: 2018 25th IEEE Inter-
national Conference on Image Processing (ICIP), IEEE, 2018, pp. 3588–3592.
[40] P. Bashivan, I. Rish, M. Yeasin, N. Codella, Learning representations from EEG
with deep recurrent-convolutional neural networks, Comput. Sci. (2015).