B(C6F5)3-Catalyzed Reduction of Cyclic N-Sulfonyl Ketimines

Article abstract of DOI:10.1002/ejoc.201900706

A metal-free method for reduction of cyclic N-sulfonyl ketimines catalyzed by B(C₆F₅)₃, using commercially available methylphenylsilane as a reducing reagent under mild conditions has been developed. This reductive protoco

Full text of DOI:10.1002/ejoc.201900706

JournalofFoodEngineering263(2019)437–445
Contents lists available at ScienceDirect
Journal of Food Engineering
journal homepage: www.elsevier.com/locate/jfoodeng
A deep feature mining method of electronic nose sensor data for identifying
beer olfactory information
Yan Shia, Furong Gonga, Mingyang Wanga, Jingjing Liua, Yinong Wub, Hong Mena,*
a College of Automation Engineering, Northeast Electric Power University, 169 Changchun Road, Jilin, 132012, PR China
b College of Mathematics, Jilin University, 2699 Qianjin Street, Changchun, 130012, PR China
A R T I C L E I N F O
A B S T R A C T
Keywords:
In this work, a deep feature mining method for electronic nose (E-nose) sensor data based on the convolutional
neural network (CNN) was proposed in combination with a support vector machine (SVM) to identify beer
olfactory information. According to the characteristics of E-nose sensor data, the structure and parameters of the
CNN was designed. By means of convolution and pooling operations, the beer olfaction features were extracted
automatically. Meanwhile, the SVM replaced the full connection layer of the CNN to enhance the generalization
ability of the model, and two important parameters affecting the classification performance of the SVM were
optimized based on an improved particle swarm optimization (PSO). The results indicated that the CNN-SVM
model achieved deep feature automatic extraction of beer olfactory information, and a good classification
performance of 96.67% was obtained in the testing set. This study shows that the CNN-SVM can be used as an
effective tool for high precision intelligent identification of beer olfactory information.
Electronic nose
Feature mining
Convolutional neural network
Support vector machine
Beer
1. Introduction
rate and detect. Although there are clear requirements for the content
of additives in different brewing stages of beer, the comprehensive ef-
E-nose is an intelligence instrument consisting of a sensor array and
pattern recognition method, which is designed to simulate the human
olfactory system. The sensor array acquires the olfactory information of
the detected object, and the pattern recognition method processes de-
tection information and gives a decision. As a new sensing technology,
the E-nose has been used widely in the field of food engineering. Such
as food classification (Ciptohadijoyo et al., 2016; Jia et al., 2016;
Banerjee et al., 2019), quality assessment (Majchrzak et al., 2018; Ke
et al., 2017; Zhu et al., 2017), freshness prediction (Chen et al., 2017;
Han et al., 2013; Min et al., 2018), identification authenticity (Majcher
et al., 2015; Śliwińska et al., 2016; Men et al., 2014) and shelf life
evaluation (Buratti et al., 2018; Luo et al., 2016; Dipan et al., 2014) etc.
Beer is one of the most productive and consumed alcoholic bev-
erages in the world (Denke, 2000). The total output and per capita
consumption of beer in European and North American countries are
among the highest worldwide. The aroma of beer affects people's sen-
sory experience directly. There are more than 100 ingredients that af-
fect the beer aroma, mainly including alcohols, esters, acids and other
substances. Different ingredients play different roles in beer aroma
(Denke, 2000; Nardini and Ghiselli, 2004; Vanbeneden et al., 2006).
The Beer odor ingredients are complicated, and very difficult to sepa-
fects of various substances will still affect the overall olfactory in-
formation. A deviation in the overall olfactory information at a certain
brewing stage indicates that the brewing process or the material allo-
cation ratio at this stage does not meet industrial requirements, and the
brewing process at that stage will be effectively controlled. At present,
the main detection methods are chemical analysis, chromatography and
mass spectrometry (Castro and Ross, 2015; Jie et al., 2018). These
methods can only detect single substances, but do not reflect the overall
odor information of beer. The cross-sensitive sensor array of the E-nose
can detect the comprehensive odor information of beer and has the
advantages of easy operation and high precision.
There are three main steps for the E-nose to decide on the measured
object: data acquisition, feature mining and recognition decision.
Feature mining methods affect the decision result of intelligent algo-
rithm directly. At present, the features extracted from the original
sensor signal are mainly divided into time-domain features, frequency-
domain features and spatial-domain features. Time-domain features
include the maximum value (Men et al., 2018b., Wei et al., 2015),
average value (Xu et al., 2016), steady state value (Qin et al., 2014),
integral value (Yin et al., 2016), differential value (Yu et al., 2013), etc.
Frequency-domain features, including the maximum energy (Zhi et al.,
* Corresponding author.
E-mail address: menhong@neepu.edu.cn (H. Men).
https://doi.org/10.1016/j.jfoodeng.2019.07.023
Received 26 February 2019; Received in revised form 23 June 2019; Accepted 25 July 2019
Availableonline29July2019
0260-8774/©2019ElsevierLtd.Allrightsreserved.

Y. Shi, et al.
JournalofFoodEngineering263(2019)437–445
2017) and the average energy (Men et al., 2018c) of wavelet packet
decomposition (Yin et al., 2014), etc. Spatial-domain features are de-
scribed by a response curve, which is composed of the sensitivity and
sensitivity change rate of sensor, including characteristic parameters
extracted from the response curve (Zhang et al., 2008). A single feature
form cannot represent the overall olfactory information of beer to a
certain extent. It is often necessary to fuse multiple features to char-
acterize the overall olfactory information of beer, which causes diffi-
culty in the feature extraction process.
ensure the homogeneity of samples, each beer was produced in the
same batch and at the same origin.
2.2. Electronic nose and experiment
A PEN3 E-nose, developed by the Airsense Analytics Inc.(Schwerin,
Germany), was employed to collect beer olfactory information. PEN3
mainly includes a sensor array, cleaning and sampling channels, and a
signal collecting system. Fig. 1 shows the PEN3 schematic diagram.
There are 10 metal oxide sensors in the PEN3 sensor array chamber.
Table 2 shows the sensitivity characteristics of each sensor. The sensor
response value of PEN3 is the ratio of the conductivity G of the sensor
after contact with the sample volatile gas to the conductivity G0 of the
sensor in contact with the standard gas filtered by activated carbon. The
interaction between sensor and gas will produce a redox reaction,
which changes the conductivity of sensor active materials, then changes
the conductivity G/G0, and finally realizes the detection of cross-sen-
sitive odor information.
A convolutional neural network (CNN) is a feedforward neural
network, which includes deep structure and convolution computation.
The CNN is one of the representative algorithms of deep learning (Ren
et al., 2017). After the 21st century, with the development of deep
learning theory and the improvement of numerical calculation
methods, CNNs have been used in computer vision (Garea et al., 2018),
natural language processing (Hang, 2018), real-time object detection
(Ren et al., 2017), etc. In contrast to time-domain features, frequency-
domain features and space-domain features, the input data feature can
be extracted automatically by means of the convolution layer and
pooling layer in the structure of the CNN without pretreatment and
statistical analysis. Meanwhile, the CNN can freely transform the form
of input data, set up a reasonable convolution structure to auto-
matically extract the features, and send them to the classifier for pattern
recognition directly. Most importantly, the CNN achieves integration of
the feature extraction and recognition processes. Although CNN has
many advantages, its training process is similar to that of the traditional
BP neural network, which requires a large amount of training data and
has the problem of overfitting. However, a large sample data acquisi-
tion is not allowed in the process of industrial detection. According to
the principle of structural risk minimization, the support vector ma-
chine (SVM) has good pattern recognition ability for small sample sizes
of data (Liu et al., 2012; Wu et al., 2018). Therefore, this paper com-
bines a CNN and a SVM to automate extraction and recognition of beer
olfactory features.
The experimental environment temperature of the E-nose was
20
0.5 °C, and the humidity was 65
2% RH. The experimental
steps were as follows:
(1) 5 ml beer was placed in a 50 ml sampler for 10 min to ensure that
the gas was saturated at the top of the sealed bottle.
(2) Before testing began, the sensor chamber was cleaned and cali-
brated. Clean air was filtered by activated carbon, and entered the
sensor array chamber for 60 s, with a flow rate of 300 mL/min.
(3) The detection started after the calibration was finished. The de-
tection time of each sample was 100 s. Fig. 2 shows the sensor re-
sponse output curve.
(4) Steps (1)–(3) was repeated, without loss of generality for 18 parallel
samples of each beer. Ninety samples of data were obtained for five
beers.
In this paper, in order to propose an effective deep feature mining
method, and provide a useful way of analyzing for beer olfactory in-
formation, the CNN-SVM is applied to identify beer olfactory informa-
tion in the field of food engineering. Five different beers with similar
alcohol content, wort concentration and raw materials were used as
experimental samples. According to the characteristics of E-nose sensor
data, the structure and parameters of the CNN were designed. The
convolution and pooling operations were applied to achieve the deep
extraction of original olfactory data. Meanwhile, in the process of SVM
classification, the penalty factor c and kernel function parameter g af-
fect the classification performance. Therefore, an improved particle
swarm optimization (PSO) method was proposed to optimize the two
important parameters. The design process and recognition results of the
CNN-SVM are discussed in detail.
Each sample of data represented the overall olfactory information of
beer.
2.3. CNN model
The CNN is a typical feedforward neural network, which is com-
posed of an input layer, hidden layer and output layer. The hidden layer
consists of convolution layer, pooling layer and fully connected layer.
Convolution simulates the response of individual neurons to visual
stimuli. It uses the convolution layer to convolute input data, and then
transfers the results to the next layer.
The convolution layer is made up of a set of convolution kernels.
Although these kernels have smaller perceptual horizons, the kernels
extend to the full depth of the input data. The function of convolution
operation is to extract the deep features of input data. For example, a
single-layer convolution network can only extract surface features such
as center and edge, while a multilayer convolution network will extract
more deeper features.
2. Materials and methods
2.1. Samples preparation
In this work, five different beers with similar alcohol content, wort
concentration and raw materials were used as experimental samples.
Table 1 shows the detailed parameters of the five different beers. To
The mathematical definition of convolution is:
Table 1
Summarizes detailed information about the tested beer samples.
No.
Brand
Alcohol by Volume (% vol)
Wort Concentration (° P)
Raw Materials
Number of Samples
1
2
3
4
5
Baiwei
Harbin
Landai
Qingdao
Xuehua
≥3.6
≥3.6
≥4.3
≥4.3
≥3.3
9.7
9.1
11
11
9
Water, malt, wheat, hops
Water, malt, rice, hops
Water, malt, rice, hops
Water, malt, rice, hops
Water, malt, rice, hops
18
18
18
18
18
438

Y. Shi, et al.
JournalofFoodEngineering263(2019)437–445
Fig. 1. The PEN3 schematic diagram.
Table 2
Basic information of olfactory sensors.
No.
Sensor
Sensitive substance
Detectability (ppm)
1
W1C
W5S
W3C
W6S
W5C
W1S
W1W
W2S
W2W
W3S
Aromatic
10
1
2
Hydrocarbon
Ammonia and Aromatic
Hydrogen
3
10
100
1
4
5
Alkanes and Aromatics
Methane
6
100
1
7
Sulphide
8
Ethanol
100
1
9
Organic Sulfides
Alkane
10
10
Fig. 3. The convolutional mapping process.
template on the input matrix. The output of the convolution layer
usually needs to use the activation function for nonlinear mapping. This
paper chose the rectified linear units (ReLU) activation function
(Krizhevsky et al., 2012).
The pooling layer performs a downsampling operation on the con-
volution output. Pooling can reduce the dimension of the output, and
retain significant features. The commonly used pooling methods are
maximum pooling, average pooling and random pooling. In this paper,
the average pooling method is used for the downsampling operation.
Fig. 4 shows the average pooling process. Taking the ‘A’ region as an
example, the average pooling operation sums all elements in ‘A’ region,
then divides the number of elements in the region to get an average and
passes it to ‘a’.
The full connection layer is the “classifier” of the CNN. Each node in
the full connection layer is connected with all the nodes in the upper
layer. Meanwhile, the full connection layer integrates the features after
the convolution and pooling operations, and maps the final feature
information to the decision space.
Fig. 2. The sensor response curve.
2.4. SVM model
P
1 Q 1
0
0
i < P + M
j < Q + N
1
1
Z (i, j) =
X (m, n) × Y (i m, j n)
As a supervised learning algorithm, SVM can analyze data and the
classification decision. SVM was proposed by Cortes and Vapnik et al.
(Cortes and Vapnik, 1995) based on the statistical theory. Based on the
principle of structural risk minimization, SVM has many advantages for
pattern recognition problems, such as small sample requirements,
nonlinearity and high-dimensional feature spaces, etc. In the process of
m=0 n=0
(1)
where the dimension of input matrix X is (P, Q), and the kernel matrix Y
is (M, N). Fig. 3 shows the convolution mapping process. The con-
volution kernel calculates the data covered by moving the weight
439

Y. Shi, et al.
JournalofFoodEngineering263(2019)437–445
Fig. 4. The average pooling process.
n
n
n
min 12
y y ai aj K (xi, xj)
i=1 ai
pattern recognition, SVM maps low-dimensional data to high-dimen-
sional space by means of a kernel function. Previous studies have shown
that the RBF kernel function expressed a good classification perfor-
mance (Li et al., 2017; Qiu et al., 2015). Therefore, the RBF was used as
a kernel function for the SVM to map low-dimensional data.
The procedure of the SVM algorithm is as follows:
i
j
i=1
j=1
n
s. t
y ai = 0,0 ai
c
i
i=1
(8)
(9)
where:
K (xi, xj) = (k (xi)·k (xj))
Set the data set D = [(x1, y1), ,(xl, yl)], sample in n-dimensional
In this paper, RBF kernels can be expressed as:
space.
2
The general form of the decision function is (Men et al., 2018a):
K (xi, xj) = exp( g xi xj
)
(10)
f (x) = ·k (x) + b
(2)
where g is the kernel function parameter. It controls the radial action
range of the function. Therefore, the above optimization problem is
converted to:
where is the weight vector, b is the domain value and k (x) is a non-
linear mapping function. To minimize structural risk, the optimal
classification plane can classify all samples correctly. The following
conditions should be satisfied:
1
2
n
n
n
2
min
y y ai aj exp( g xi xj
)
i=1 ai
i
j
i=1
j=1
n
s. t
y ai = 0,0 ai
c
i
y ( T·k (xi) + b)
1
i=1
(11)
(3)
i
As seen from formula (11), the parameters c and g affect the clas-
sification performance for the SVM. Therefore, PSO was introduced to
calculate the parameters.
To achieve a certain balance between experiential risk and gen-
eralization performance, the existence of misclassified samples is al-
lowed by introducing nonnegative slack variable i. Therefore, the op-
timization problem is converted to:
PSO is an optimization algorithm based on swarm intelligence in the
field of computational science. Its basic concept originates from the
study of bird predation behavior (Fong et al., 2016; Messerschmidt and
Engelbrecht, 2004). In the process of iterative optimization, the parti-
cles keep track of each other's historical optimum accuracy and con-
stantly update their search direction and speed, so that the particles
converge toward the optimum direction.
n
min 12
+ c
i, c
0
2
i=1
s. t y ( T·k (xi) + b)
1
,
i i
0
i
(4)
where c is the penalty factor. It can control the degree of punishment
for misclassification samples. Here, the Lagrange multiplier algorithm is
introduced to:
The speed update formulae of traditional PSO algorithm are as
follows:
n
1
2
2
T
L( , b, ai) =
+
ai (1 y ( k (xi) + b)), ai = (a1, a2, ,an)
i
v (t + 1) = t + c1rand()(qbest (t) q(t)) + c2 rand()(pbest (t) q(t))
i=1
q(t + 1) = q(t) + v(t + 1)
(5)
(12)
Then, L( , b, ai) derives partial derivatives of and b respectively:
where v(t) is the velocity of particle at time t,
is the inertia weight,
m
=
ai y xi
q
best(t) is the optimal solution of particle at time t, q(t) is the solution of
i
i=1
m
particle at time t, pbest(t) is the global optimal solution for all particles at
time t, rand () is a random number in the range [0, 1], and c1 and c2 are
learning factors.
0 =
ai y
i
(6)
i=1
The Lagrange multiplier algorithm is converted to:
In the traditional PSO algorithm, the
describes the influence of
1
2
n
n
n
T
L( , b, ai) =
+
i=1 ai
ai y Txi
T
ai y b
i
i
i=1
i=1
the particle's previous generation velocity to the current generation
velocity. The larger the search range of particles is, the better the al-
gorithm can find global optimization and avoid falling into local op-
timal solutions. The smaller the search range of particles, the smaller
the search range will be, which will enhance the local search ability and
make the algorithm converge more quickly. In this paper, the balance
between global search and local optimal ability was adjusted. The
n
1
2
T
=
i=1 ai +
=
i=1 ai
(
ai y xi)T
aj y xj
j
n
1
n
n
i
i=1
j=1
2
n
1
2
n
n
=
i=1 ai
ai aj y yj xiT xj
i
i=1
j=1
(7)
Therefore, the optimization problem is converted to dual form:
formula for calculating the
can be defined as follow:
440

Y. Shi, et al.
JournalofFoodEngineering263(2019)437–445
max
min ·N
3.2. CNN structure
=
max
Nmax
(13)
The matrix form of the beer olfactory information was 100*10,
where 100 was the number of sampling points for each sensor and 10
was the number of sensors. In this paper, the first 90 sampling points
were selected for each sensor, and the sample matrix became 90*10,
which was converted to 30*30 as the input of the CNN.
where
is the maximum inertia weight, which is 0.9,
is the
minimum inertia weight, which is 0.3, Nmax is the maximummiitneration
max
algebra and N is the current iteration algebra. Formula (13) shows that
the
value is the largest at the beginning of the iteration, which en-
ables the particles to search globally in a wide range. As the number of
iterations increases, the particle gradually approaches the global op-
Fig. 6 shows the structure schematic diagram of the CNN. The
structure of the CNN consisted of 4 convolution layers, 3 pooling layers
and 2 full connection layers. After the last pooling operation, all feature
matrices were connected into a vector as input to the first full con-
nection layer. Table 3 shows the network parameters of the CNN. In all
convolution operations, the convolution kernel size was 3*3, the stride
was 1 and ReLU was selected as activation function. Padding was
‘same’, which meant that 0 was added to the periphery of matrix data to
preserve and extract edge features. In contrast, ‘valid’ did not add the
padding. In all downsampling operations, the stride was 2 and the filter
was 2*2. In the first full connection operation, ReLU was selected as the
activation function, and the number of neurons was 32 according to the
number of the pooling3 feature metrics. In the second full connection
operation, Sigmoid was selected as activation function, and the number
of neurons was 5 according to the number of categories. The design
process of each layer was as follows:
timal solution. Meanwhile, the
value decreases, which enables the
particle to search locally in a small range and ultimately achieve the
global optimal solution. The fitness of inertia weight
varies with the
number of iterations, so it is called adaptive inertia weight.
c1 reflects the information exchange between individual particles,
and c2 reflects the information exchange between the particle popula-
tion and the historical optimal trajectory. This paper introduced the
asynchronous learning formula to dynamically adjust c1 and c2. The
adjusted formulae can be defined as follows:
c
c
1max
c1 = c1max
1min ·N
2max ·N
N
2min
max
c
c
c2 = c2min
N
max
(14)
where c1max is the maximum of c1 learning factor, and its value is 2,
c
1min is the minimum of c1 learning factor, and its value is 1, c2max is the
maximum of c2 learning factor, and its value is 2, and c2min is the
minimum of c2 learning factor, and its value is 1.
(1) The original E-nose data input matrix was 90*10, which was con-
verted into 30*30. In principle, more features can be acquired by
means of convolution kernels, but too many features can lead to
overfitting of the recognition model. Therefore, 4 convolution
kernels were set to convolve the original data after adding padding
items. Here, 4 feature matrices were obtained in the same form, and
the matrix size of each feature was still 30*30.
It can be seen from Formula (14), with the increase of iterations, the
learning factor c1 value is the largest at the beginning of the iteration
and then decreases, while the c2 value is the smallest at the beginning of
the iteration and then increases. In this way, using the asynchronous
learning characteristics can exchange information between particles
effectively (Zhao and Fang, 2013).
(2) Eight convolution kernels were set to convolute the input matrices.
Here, 8 feature matrices were obtained in the same form, and each
feature matrix size was changed to 28*28.
To control the flying speed of particles effectively, the algorithm
achieved an effective balance between global detection and local
mining. In this paper, the compression factor was introduced and the
formula can be defined as follow:
(3) The data were compressed by means of pooling operation. In this
paper, the global average pooling operation was applied. Here, the
number of feature matrices remained constant, and each feature
matrix size was changed to 14*14.
2
=
2
c
c2 4c
(15)
(4) According to the parameters in Table 3, the calculation process of
No.4-No.7 were the same as that of (2)–(3). Finally, 32 feature
matrices were obtained in the same form, and each feature matrix
size was changed to 2*2.
where c = c1+c2.
Finally, the adaptive particle swarm optimization algorithm with
compression factor and asynchronous learning factor was proposed
(CAAPSO). The particle velocity position updated formula can be de-
fined as follows:
(5) Before the full connection operation, it converted 32 feature ma-
trices with sizes of 2*2 into a feature matrix as the input to the first
full connection layer.
v (t + 1) =
(
t + c1rand()(qbest (t) q(t)) + c2rand()(pbest (t) q(t)))
3.3. CNN performance evaluation
q(t + 1) = q(t) + v(t + 1)
(16)
The original 90 groups of beer data were divided into two groups
randomly: 2/3 were used to train the CNN as training set (containing
validation sets), and 1/3 were used as the testing set. Data were pro-
cessed based on section 3.2 CNN structure.
3. Results and discussion
The batch training mode was applied to train CNN. The initializa-
tion batch size was 20 based on the smaller beer samples. The BP al-
gorithm was used to train the CNN by means of the gradient descent
algorithm. In the iteration process of weights and biases, only the
learning rate needs to be set, which was set to 0.1. Xavier was applied to
make the information flow better in the network. The variance in the
output of each layer should be as equal as possible (Xavier and Yoshua,
2010). The connection weights between adjacent network layers and
the weight matrix of the convolution kernels were initialized according
to the following uniform distribution.
3.1. Data analysis
A radar plot was used to illustrate the relationships and trends of
sensors response data. To visualize the data, one sample was randomly
selected from the five different beer samples. Fig. 5 shows the radar plot
of sensors 90 s for five different beers. The radar response forms of five
beers were similar, which may mean that the distinction was difficult.
While the W5S, W1S, W1W, W2S, W2W, W3S responses were larger,
the W1C, W3C, W6S, W5C responses were smaller. However, for beer
identification, we are not sure whether a large response sensor is highly
important, or a small response sensor is less important (Men et al.,
2018a). Therefore, it is particularly important to deeply mine the im-
portant features within the sensor data.
6
6
W~U
,
nj + nj+1
nj + nj+1
(17)
441

Y. Shi, et al.
JournalofFoodEngineering263(2019)437–445
Fig. 5. The sensor response radar plots for five different beers: (a) Baiwei, (b) Harbin, (c) Landai, (d) Qingdao, (e) Xuehua.
Fig. 6. The CNN structure schematic diagram.
For the connection weight between the network layers, njand nj+1
represents the number of steps in the training of CNN for all the data of
the training set. Since the training set contained 60 samples, the
number of batch size was 20. Therefore, the total number of iterations
was 600. As the number of iterations increases, the training status no
longer changes when the MSE value changes less than 10−3. Fig. 7
shows the loss function curve based on MSE under the 5-fold cross-
validation of training set. The MSE of the training set and the validation
set decreased significantly before 300 iterations, while the MSE value of
represented the number of adjacent two layers of neurons, respectively,
and the bias was initialized to 0. For the weight matrix of convolution
kernels, nj and nj+1 represented the product of the number of adjacent
two layers feature metrics and the size of convolution kernels respec-
tively.
The MSE loss function was used for error calculation in CNN
learning. In the process of iteration calculation, epoch was 200, which
442

Y. Shi, et al.
JournalofFoodEngineering263(2019)437–445
Table 3
The network parameters of the CNN.
No.
Type
Kernel
Stride
Padding
Input size
Output size
Active function
1
Convolution1
Convolution2
Pooling1
3 × 3 (4)
3 × 3 (8)
2 × 2
3 × 3 (16)
2 × 2
3 × 3 (32)
2 × 2
–
1
1
2
1
2
1
2
–
–
–
same
32 × 32
30 × 30 (4)
28 × 28 (8)
14 × 14 (8)
12 × 12 (16)
6 × 6 (16)
4 × 4 (32)
2 × 2 (32)
128 × 1
ReLU
ReLU
–
2
valid
30 × 30 (4)
28 × 28 (8)
14 × 14 (8)
12 × 12 (16)
6 × 6 (16)
4 × 4 (32)
2 × 2 (32)
–
3
–
4
Convolution3
Pooling2
valid
ReLU
–
5
–
6
Convolution4
Pooling3
valid
ReLU
–
7
–
–
–
–
8
Feature vector
Full connected1
Full connected2
–
9
–
–
ReLU
Sigmoid
10
–
–
–
Note: The numbers in parentheses represent the number of convolution kernels and the number of inputs and outputs of the matrix.
fold cross-validation of training set. When the accuracy rate reached a
maximum and no longer increased, c and g were selected as the best
parameter. In this paper, c and g were selected in the range of (0, 1000).
In the testing phase, the trained SVM model replaced the full connec-
tion layer of the CNN. Meanwhile, the testing set samples were input to
the trained CNN (only the convolution layer and downsampling layer
were left at this time), to obtain the corresponding eigenvectors of each
testing set sample. The eigenvectors of the test samples were input into
the trained SVM model for pattern recognition. Fig. 9 (a) shows the
parameter optimization process with CAAPSO. The highest 5-fold cross-
validation accuracy was 98.3333%, the optimal parameter c was
7.3589 and g was 0.01. Fig. 9 (b) shows the classification results. One of
the fifth beers was misclassified into the fourth category, the final
classification accuracy was 96.67%.
4. Conclusions
In this study, a deep feature mining method was proposed to extract
the sensors data of E-nose. Meanwhile, the feature extraction and pat-
tern recognition process of E-nose sensor data were integrated. The
main conclusions are as follows:
Fig. 7. The loss function curve based on MSE.
the validation set changed less than 10−3 after 480 iterations. Mean-
while, the overall MSE value of the validation set was significantly
higher than that of the training set, which indicated that even under the
five-fold cross-validation, the training process of CNN was fitted. Fi-
nally, the training accuracy was 80%, and the validation accuracy was
70.77%. Clearly, such training effects cannot meet the actual applica-
tion requirements.
(1) A new structure of CNN was designed which included 4 convolution
layers, 3 pooling layers and 2 full connection layers. By setting the
CNN input form, convolution kernels, convolution stride, activation
function and other parameters, the features of E-nose sensor data
can be extracted automatically.
(2) From the training process of the CNN, it can be seen that the model
was overfitted, and the overall MSE value of the validation set was
higher than that of the training set. The classification accuracy of
the validation set was 70.77%, which is obviously not in line with
the actual application requirements. Therefore, SVM replaced the
full connection layer of the CNN to enhance the pattern recognition
ability.
3.4. CNN-SVM results
Fig. 8 shows the implementation process of the CNN-SVM. After
CNN training is completed, the features of training set samples will be
extracted automatically. The features were sent to CAAPSO-SVM for
training. In the initialization process of CAAPSO, the number of parti-
cles was 30, the number of iterations was 100 and each particle had the
same velocity. The fitness function had the highest accuracy under 5-
(3) The adaptive PSO algorithm with compression factor and asyn-
chronous learning factor was proposed to avoid the shortcomings of
particles prematurity and local optima due to unreasonable para-
meter setting. Based on CAAPSO, two important parameters
Fig. 8. The flow chart of the beer samples pattern recognition.
443

Y. Shi, et al.
JournalofFoodEngineering263(2019)437–445
Fig. 9. Decision process for testing set with the CNN-SVM. (a) The fitness curve in parameter optimization of CAAPSO; (b) The model decision result.
affecting the classification performance of SVM were optimized.
Finally, the highest classification accuracy of the validation set was
98.33%, and the recognition result of CNN-SVM was 96.67%.
of shelf-life of fried potato wedges using electronic noses: sensor screening by fuzzy
logic analysis. J. Food Eng. 133, 23–29. https://doi.org/10.1016/j.jfoodeng.2014.02.
009.
Fong, S., Wong, R., Vasilakos, A.V., 2016. Accelerated PSO swarm search feature selection
for data stream mining big data. IEEE Trans. Ser. Comput. 9 (1), 33–45. https://doi.
org/10.1109/tsc.2015.2439695.
This study shows that CNN can extract beer E-nose sensor features
effectively. SVM based on an improved PSO enhanced the classification
performance of the CNN. It can reduce the detection difficulty and
improve the detection efficiency with as little sample data as possible,
and obtain a better qualitative analysis result. Moreover, it also pro-
vided a new and effective method for beer quality control.
Garea, A.S., Heras, D.B., Argüello, F., 2018. Caffe CNN-based classification of hyper-
spectral images on GPU. J. Supercomput. (3), 1–13. https://doi.org/10.1007/
s11227-018-2300-2.
Han, F., Huang, X., Teye, E., Gu, F., Gu, H., 2013. Nondestructive detection of fish
freshness during its preservation by combining electronic nose and electronic tongue
techniques in conjunction with chemometric analysis. Anal. Methods 6 (2), 529–536.
https://doi.org/10.1039/C3AY41579A.
Hang, L., 2018. Deep learning for natural language processing:advantages and challenges.
Natl. Sci. Rev. 5 (1), 24–26. https://doi.org/10.1093/nsr/nwx110.
Jia, X.M., Meng, Q.H., Jing, Y.Q., Qi, P.F., Zeng, M., Ma, S.G., 2016. A new method
combining KECA-LDA with ELM for classification of Chinese liquors using electronic
nose. IEEE Sens. J. 16 (22), 8010–8017. https://doi.org/10.1109/jsen.2016.
2606163.
Acknowledge
This work was supported by the National Natural Science
Foundation of China [31772059, 31871882]; the Key Science and
Technology Project of Jilin Province [20170204004SF]; and the
Provincial Special Funds for Industrial Innovation of Jilin Province
[2018C034-8].
Jie, G., Chen, D., Zhao, Y., 2018. Determination of carbendazim and thiabendazole in
wine and beer by ultra high performance liquid chromatography high resolution mass
spectrometry coupled with dispersive micro solid-phase extraction. Chin. J.
Chromatogr. 36 (2), 143–149. https://doi.org/10.3724/SP.J.1123.2017.10018.
Ke, X., Jun, W., Zhen, Wei, Fan, D., Yong, W., Shao, C., 2017. An optimization of the MOS
electronic nose sensor array for the detection of Chinese pecan quality. J. Food Eng.
203, 25–31. https://doi.org/10.1016/j.jfoodeng.2017.01.023.
Appendix A. Supplementary data
Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. Imagenet classification with deep con-
volutional neural networks. Adv. Neural Inf. Process. Syst. 25 (2), 1097–1105.
https://doi.org/10.1145/3065386.
Supplementary data to this article can be found online at https://
doi.org/10.1016/j.jfoodeng.2019.07.023.
Li, Y., Zhang, J., Li, T., Liu, H., Li, J., Wang, Y., 2017. Geographical traceability of wild
Boletus edulis based on data fusion of FT-MIR and ICP-AES coupled with data mining
methods (SVM). Spectrochim. Acta Part A Molecular & Biomolecular Spectroscopy
177, 20–27. https://doi.org/10.1016/j.saa.2017.01.029.
References
Liu, J., Zeng, F., Guo, Z., 2012. Gas outburst risk analysis based on pattern recognition of
RSSVM model. Procedia Engineering (29), 170–173. https://doi.org/10.1016/j.
proeng.2011.12.689.
Banerjee, M.B., Runu, B.R., Bipan, T., Rajib, B., Nabarun, B., 2019. Black tea classification
employing feature fusion of E-Nose and E-Tongue responses. J. Food Eng. 24, 55–63.
https://doi.org/10.13982/j.mfst.1673-9078.2016.2.035.
Luo, F., Lu, X.X., Zhang, P., Chen, S.H., Li, J.K., 2016. Fruit quality and content of volatile
substances in summit cherry during shelf life. Modern Food Science & Technology 32
(2), 235–245. https://doi.org/10.13982/j.mfst.1673-9078.2016.2.035.
Majcher, M.A., Kaczmarek, A., Klensporf-Pawlik, D., Pikul, J., Jeleń, H.H., 2015. SPME-
MS-Based electronic nose as a tool for determination of authenticity of PDO cheese,
Oscypek. Food Analytical Methods 8 (9), 2211–2217. https://doi.org/10.1007/
s12161-015-0114-x.
Buratti, S., Malegori, C., Benedetti, S., Oliveri, P., Giovanelli, G., 2018. E-nose, e-tongue
and e-eye for edible olive oil characterisation and shelf life assessment: a powerful
data fusion approach. Talanta 182, 131–141. https://doi.org/10.1016/j.talanta.
2018.01.096.
Castro, L.F., Ross, C.F., 2015. Determination of flavour compounds in beer using stir-bar
sorptive extraction and solid-phase microextraction. J. Inst. Brew. 121 (2), 197–203.
https://doi.org/10.1002/jib.219.
Majchrzak, T., Wojnowski, W., Dymerski, T., Gębicki, J., Namieśnik, J., 2018. Electronic
noses in classification and quality control of edible oils: a review. Food Chem. 246,
192–201. https://doi.org/10.1016/j.foodchem.2017.11.013.
Chen, H.Z., Min, Z., Bhandari, B., Guo, Z., 2017. Evaluation of the freshness of fresh-cut
green bell pepper ( Capsicum annuum var. grossum ) using electronic nose. LWT -
Food Sci. Technol. (Lebensmittel-Wissenschaft -Technol.) 87, 77–84. https://doi.org/
10.1016/j.lwt.2017.08.052.
Men, H., Chen, D., Zhang, X., Liu, J., Ning, K., 2014. Data fusion of electronic nose and
electronic tongue for detection of Mixed Edible-Oil. Journal of Sensors, 2014 1–8.
https://doi.org/10.3390/s18010285.
Ciptohadijoyo, S., Litananda, W.S., Rivai, M., Purnomo, M.H., 2016. Electronic nose
based on partition column integrated with gas sensor for fruit identification and
classification. Comput. Electron. Agric. 121, 429–435. https://doi.org/10.1016/j.
compag.2015.11.013.
Men, H., Fu, S., Yang, J., Chen, M., Shi, Y., Liu, J., 2018c. Comparison of SVM, RF and
ELM on an electronic nose for the intelligent evaluation of paraffin samples. Sensors
18 (1), 285–292. https://doi.org/10.3390/s18010285.
Men, H., Jiao, Y., Shi, Y., Gong, F., Chen, Y., Fang, H., Liu, J., 2018b. Odor fingerprint
analysis using feature mining method based on olfactory sensory evaluation. Sensors
18 (10), 3387–3408. https://doi.org/10.3390/s18103387.
Cortes, C., Vapnik, V., 1995. Support vector network. Mach. Learn. 20 (3), 273–297.
https://doi.org/10.1023/A:1022627411411.
Denke, M.A., 2000. Nutritional and health benefits of beer. Am. J. Med. Sci. 320 (5),
320–326. https://doi.org/10.1097/00000441-200011000-00004.
Dipan, C., Paramita, B., Nabarun, B., 2014. Development of methodology for assessment
Men, H., Shi, Y., Jiao, Y., Fu, S., Liu, J., 2018a. Electronic nose sensors data feature
mining: a synergetic strategy for the classification of beer. Anal. Methods 10 (17),
444

Y. Shi, et al.
JournalofFoodEngineering263(2019)437–445
2016–2025. https://doi.org/10.1039/c8ay00280k.
recognition of knee motion using mechanomyography signals. J. Electromyogr.
Kinesiol. 42, 136–142. https://doi.org/10.1016/j.jelekin.2018.07.005.
Xavier, G., Yoshua, 2010. Understanding the difficulty of training deep feedforward
neural networks. J. Mach. Learn. Res. 9, 249–256.
Messerschmidt, L., Engelbrecht, A.P., 2004. Learning to play games using a PSO-based
competitive learning approach. IEEE Trans. Evol. Comput. 8 (3), 280–288. https://
doi.org/10.1109/TEVC.2004.826070.
Min, W., Fan, G., Qian, W., Zhang, J., Xue, Y., Hao, W., Ping, W., 2018. The real-time
assessment of food freshness in refrigerator based on miniaturized electronic nose.
Anal. Methods 10, 4147–4149. https://doi.org/10.1039/C8AY01242C.
Nardini, M., Ghiselli, A., 2004. Determination of free and bound phenolic acids in beer.
Food Chem. 84 (1), 137–143. https://doi.org/10.1016/s0308-8146(03)00257-7.
Qin, O., Zhao, J., Chen, Q., 2014. Instrumental intelligent test of food sensory quality as
mimic of human panel test combining multiple cross-perception sensors and data
fusion. Anal. Chim. Acta 841 (23), 68–76. https://doi.org/10.1016/j.aca.2014.06.
001.
Xu, L., Yu, X., Liu, L., Zhang, R., 2016. A novel method for qualitative analysis of edible
oil oxidation using an electronic nose. Food Chem. 202, 229–235. https://doi.org/10.
1016/j.foodchem.2016.01.144.
Yin, Y., Hao, Y., Yu, H., 2016. Identification method for different moldy degrees of maize
using electronic nose coupled with multi-features fusion. Trans. Chin. Soc. Agric. Eng.
32 (12), 254–260. https://doi.org/10.11975/j.issn.1002-6819.2016.12.036.
Yin, Y., Wu, W., Yu, H., 2014. Drift elimination method of electronic nose signals based
on independent component analysis coupled with wavelet energy threshold value.
Trans. Chin. Soc. Agric. Eng. 30 (24), 325–331. https://doi.org/10.3969/j.issn.1002-
6819.2014.24.040.
Qiu, S., Wang, J., Tang, C., Du, D., 2015. Comparison of ELM, RF, and SVM on E-nose and
E-tongue to trace the quality status of Mandarin (Citrus unshiu Marc.). J. Food Eng.
166, 193–203. https://doi.org/10.1016/j.jfoodeng.2015.06.007.
Yu, H., Chu, B., Yin, Y., 2013. Evaluation method of feature vector in vinegar identifi-
cation by electronic nose. Trans. Chin. Soc. Agric. Eng. 29 (3), 258–264. https://doi.
org/10.3969/j.issn.1002-6819.2013.03.034.
Ren, S., He, K., Girshick, R., Sun, J., 2017. Faster R-CNN: towards real-time object de-
tection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39 (6),
1137–1149. https://doi.org/10.1109/TPAMI.2016.2577031.
Zhang, S., Xie, C., Mulin, H.U., Huayao, L.I., Bai, Z., Zeng, D., 2008. An entire feature
extraction method of metal oxide gas sensors. Sens. Actuators B Chem. 132 (1),
81–89. https://doi.org/10.1016/j.snb.2008.01.015.
Śliwińska, M., Wiśniewska, P., Dymerski, T., Wardencki, W., Namieśnik, J., 2016.
Application of electronic nose based on fast GC for authenticity assessment of polish
homemade liqueurs called nalewka. Food Analytical Methods 9 (9), 2670–2681.
https://doi.org/10.1007/s12161-016-0448-z.
Zhao, Y., Fang, Z., 2013. Particle swarm optimization algorithm with weight function's
learning factor. J. Comput. Appl. 33 (8), 2265–2268. https://doi.org/10.3724/SP.J.
1087.2013.02265.
Vanbeneden, N., Delvaux, F., Delvaux, F.R., 2006. Determination of hydroxycinnamic
acids and volatile phenols in wort and beer by isocratic high-performance liquid
chromatography using electrochemical detection. J. Chromatogr. A 1136 (2),
237–242. https://doi.org/10.1016/j.chroma.2006.11.001.
Zhi, R., Zhao, L., Zhang, D., 2017. A framework for the multi-level fusion of electronic
nose and electronic tongue for tea quality assessment. Sensors 17 (5), 1007–1022.
https://doi.org/10.3390/s17051007.
Zhu, J., Chen, F., Wang, L., Niu, Y., Xiao, Z., 2017. Evaluation of the synergism among
volatile compounds in oolong tea infusion by odour threshold with sensory analysis
and E-nose. Food Chem. 221, 1484–1490. https://doi.org/10.1016/j.foodchem.2016.
11.002.
Wei, Z., Wang, J., Zhang, W., 2015. Detecting internal quality of peanuts during storage
using electronic nose responses combined with physicochemical methods. Food
Chem. 177, 89–96. https://doi.org/10.1016/j.foodchem.2014.12.100.
Wu, H., Huang, Q., Wang, D., Gao, L., 2018. A CNN-SVM combined model for pattern
445