THE SIGNIFICANCE OF THE
CONVOLUTIONAL DEEP LEARNING MODEL
IN THE INTELLIGENT COLLABORATIVE
CORRECTION OF ENGLISH WRITING
Hong Wang*
Department of Social Work, Formerly the Ministry of Civil Affairs Management
Cadre Institute, Beijing College of Social Administration, Beijing, 102628, China
hongw_200466@163.com
Reception: 18/11/2022 Acceptance: 07/01/2023 Publication: 26/01/2023
Suggested citation:
W., Hong. (2023). The significance of the convolutional deep learning
model in the intelligent collaborative correction of English writing. 3C
Tecnología. Glosas de innovación aplicada a la pyme, 12(1), 142-157. https://
doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
142
THE SIGNIFICANCE OF THE
CONVOLUTIONAL DEEP LEARNING MODEL
IN THE INTELLIGENT COLLABORATIVE
CORRECTION OF ENGLISH WRITING
Hong Wang*
Department of Social Work, Formerly the Ministry of Civil Affairs Management
Cadre Institute, Beijing College of Social Administration, Beijing, 102628, China
hongw_200466@163.com
Reception: 18/11/2022 Acceptance: 07/01/2023 Publication: 26/01/2023
Suggested citation:
W., Hong. (2023). The significance of the convolutional deep learning
model in the intelligent collaborative correction of English writing. 3C
Tecnología. Glosas de innovación aplicada a la pyme, 12(1), 142-157. https://
doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
ABSTRACT
To ensure the normal operation of English composition grammar correction and avoid
inaccurate detection caused by faults, it is of great significance to detect abnormal
working conditions in time and diagnose them accurately. Aiming at the complexity of
grammar correction, this paper proposes a PLSTM-CNN model for fault detection in
the grammar correction process. The model effectively combines the global feature
extraction ability of LSTM for time series data and the ability of the CNN model to
extract local features, which reduces the loss of feature information and achieves a
higher fault detection rate. A one-dimensional dense CNN is used as the main body of
the CNN, and the LSTM network is sensitive to changes in sequence information to
avoid model overfitting while building a deeper network. The maximum mutual
information coefficient (MMIC) data preprocessing method is adopted to improve the
local correlation of the data and improve the efficiency of the PLSTM-CNN model to
detect faults from different initial conditions. The research results show that the
parallel PLSTM-CNN has better prediction performance than the serial PLSTM-CNN,
and its FDR and FPR are 90.5% and 0.051, respectively. It shows that the use of
convolutional deep learning models for the prediction of writing grammar correction
faults has strong application prospects.
KEYWORDS
PLSTM-CNN; Fault detection; English writing; grammar; Deep learning.
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
143
PAPER INDEX
ABSTRACT
KEYWORDS
1. INTRODUCTION
2. CORRELATION MODEL THEORY
2.1. 1D-CNN
2.2. DCNN
2.3. LSTM Long Short-Term Memory Network
2.4. Model Structure Diagram
3. EXPERIMENT AND RESULT ANALYSIS
3.1. Evaluation Results
3.2. Comparative Experiment
3.2.1. Comparison of Average Failure Detection Rates
3.2.2. MODEL inference and inference time comparison
3.2.3. Comparison of Average Fault Detection Rates for Small Samples
4. CONCLUSION
5. CONFLICT OF INTEREST
REFERENCES
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
144
PAPER INDEX
ABSTRACT
KEYWORDS
1. INTRODUCTION
2. CORRELATION MODEL THEORY
2.1. 1D-CNN
2.2. DCNN
2.3. LSTM Long Short-Term Memory Network
2.4. Model Structure Diagram
3. EXPERIMENT AND RESULT ANALYSIS
3.1. Evaluation Results
3.2. Comparative Experiment
3.2.1. Comparison of Average Failure Detection Rates
3.2.2. MODEL inference and inference time comparison
3.2.3. Comparison of Average Fault Detection Rates for Small Samples
4. CONCLUSION
5. CONFLICT OF INTEREST
REFERENCES
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
1. INTRODUCTION
Writing ability is an important basis for measuring students' English learning and
practical ability, and plays an significant role in promoting the overall development of
language skills. In recent years, the rapid development of computers and networks
has laid the foundation for the reform of college English writing [1]. Due to the
limitations of the current technology of the automatic evaluation system, more
feedback is given to students at the vocabulary level, but there are still many
problems in the evaluation of syntax, text structure, logic, and coherence. Therefore,
relying solely on machine correction and feedback is of limited help to students [2]. To
solve this problem, manual intervention and feedback are required [3]. Therefore, the
need for intelligent collaboration is extremely necessary.
The research on English marking begins with writing feedback. Research on writing
feedback began in the 1950s. Before this, writing feedback was conducted by
teachers. During this period, some scholars conducted comparative studies on
teachers' correction of students' compositions and students' peer evaluation under the
guidance of teachers and found that peer feedback was better than teacher feedback
[4]. In the following decades, more and more researchers from China and abroad
began to pay attention to the application of peer feedback in practical teaching and
research, and their research results provided a lot of guidance for our teaching and
research [5]. However, some problems and difficulties with peer feedback have also
been found in some studies. For example, the correctness, fairness, and effectiveness
of feedback are often questioned, and the operability in writing classrooms also needs
to be verified. In the 1960s, Professor Ellis Page of Duke University in the United
States developed the PEG automatic composition scoring system, and the automatic
writing evaluation system (AWES) gradually developed [6]. For decades, especially
with the development of artificial intelligence technology, the development of foreign
writing automatic evaluation systems has made great progress, such as Criterion, My
Access, and Writing Roadmap [7]. Then there is the study of translation. According to
the basic principle of deep reinforcement learning algorithm, some researchers
designed a neural machine translation model, introduced the evaluation mechanism to
the level of the sentence to be translated, predicted the convergence of the
translation, and used the deep reinforcement learning algorithm as a guiding strategy
for translation. Optimize the word sequence of the translation target, integrate the
monolingual corpus into the training of deep reinforcement learning, and alleviate the
data-sparse problem of translation sentences. After experimental tests, it is found that
this model can improve the overall performance of machine translation. Compared
with other translation models, Whether it is Chinese-Korean or Korean-Chinese, the
BLEU value has been significantly improved [8], but the performance in other aspects
is poor. Considering the problems existing in traditional machine translation, some
researchers have designed a neural machine translation model based on the basic
principle of quality estimation, and used the quality estimation method to score the
pseudo-parallel data generated by reverse translation, and use the data with higher
scores as the basis for quality estimation. Design of CNN's English Machine
Translation Minor Error Detection System [9]. Some scholars use the input of the
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
145
neural network to control the quality of the pseudo-parallel data generated by the
reverse translation, and provide a rich training network as the output for the neural
network model. After experimental tests, it is found that compared with the traditional
model, the model has no effect on the forward translation or reverse translation. For
translation, the BLEU value has been improved, but the function cannot meet the
design requirements [10].
One-dimensional CNN [11] divides the input data along a single dimension without
windowing operation, which is easier to train and has less computational complexity.
Although 1D convolution is the current popular deep learning method, it still has some
limitations. Although the CNN can extract the local features of the data, its ability to
extract the global features of the data is weak. To obtain beneficial features such as
global and local use of data at the same time, this paper proposes an LSTM-CNN
structure based on a parallel structure, which combines the local features extracted by
the CNN and the global features extracted by LSTM to make full use of the data.
features to improve the accuracy of the model, thereby reducing the translation
accident rate [12].
2. CORRELATION MODEL THEORY
2.1. 1D-CNN
A CNN is one of the widely studied deep learning algorithms, which has the
characteristics of local connection, weight sharing, and downsampling [13]. The
difference between 1D-CNN and classical CNN is the dimension of the convolution
kernel, which has been widely used in time series feature extraction in recent years
[14]. The one-dimensional convolution operation is shown in Figure 1, and its
mathematical model is shown in formula (1):
Among them, Hi is the input feature quantity of the ith layer; Wi and bi represent the
weight and corresponding bias of the convolution kernel of the ith layer respectively; f
represents the activation function, here is the Relu activation function, which has a
good nonlinear expression ability.
Pooling layers are also known as subsampling layers. The sub-sampling layer
downsamples the feature map according to the rules and reduces the dimension of
the convolutional feature to reduce the parameters and calculation amount inside the
CNN, and at the same time suppress the network overfitting.
Suppose Hlt is the jth feature map of the l-th sub-sampling layer, and its sampling
process is as shown in formula (2), which represents the next sampling function. Each
output feature map corresponds to its own multiplicative bias Blj and an additive bias
blj [15].
(1)
1
( )
i i i i
H f H W b
−
=⊗+
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
146
neural network to control the quality of the pseudo-parallel data generated by the
reverse translation, and provide a rich training network as the output for the neural
network model. After experimental tests, it is found that compared with the traditional
model, the model has no effect on the forward translation or reverse translation. For
translation, the BLEU value has been improved, but the function cannot meet the
design requirements [10].
One-dimensional CNN [11] divides the input data along a single dimension without
windowing operation, which is easier to train and has less computational complexity.
Although 1D convolution is the current popular deep learning method, it still has some
limitations. Although the CNN can extract the local features of the data, its ability to
extract the global features of the data is weak. To obtain beneficial features such as
global and local use of data at the same time, this paper proposes an LSTM-CNN
structure based on a parallel structure, which combines the local features extracted by
the CNN and the global features extracted by LSTM to make full use of the data.
features to improve the accuracy of the model, thereby reducing the translation
accident rate [12].
2. CORRELATION MODEL THEORY
2.1. 1D-CNN
A CNN is one of the widely studied deep learning algorithms, which has the
characteristics of local connection, weight sharing, and downsampling [13]. The
difference between 1D-CNN and classical CNN is the dimension of the convolution
kernel, which has been widely used in time series feature extraction in recent years
[14]. The one-dimensional convolution operation is shown in Figure 1, and its
mathematical model is shown in formula (1):
Among them, Hi is the input feature quantity of the ith layer; Wi and bi represent the
weight and corresponding bias of the convolution kernel of the ith layer respectively; f
represents the activation function, here is the Relu activation function, which has a
good nonlinear expression ability.
Pooling layers are also known as subsampling layers. The sub-sampling layer
downsamples the feature map according to the rules and reduces the dimension of
the convolutional feature to reduce the parameters and calculation amount inside the
CNN, and at the same time suppress the network overfitting.
Suppose Hlt is the jth feature map of the l-th sub-sampling layer, and its sampling
process is as shown in formula (2), which represents the next sampling function. Each
output feature map corresponds to its own multiplicative bias Blj and an additive bias
blj [15].
(1)
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
2.2. DCNN
To alleviate the gradient explosion problem caused by the network depth, the
dense CNN [16] (Dense CNN, DCNN) further connects each sub-layer based on the
residual network structure, so that the output of each layer of the network is used as
The input of one layer ensures maximum feature reuse, alleviates the gradient
disappearance and gradient explosion problems caused by the increase in the
number of network layers and makes the network information flow more smoothly.
Assuming that the number of network layers is, the DCNN contains a total of N/
(N+1)/2 connections. Through the sequence x0
passed through the convolutional
layer, the input of the nth layer is the feature map of all previous layers, as shown in
the following formula (3):
Among them, [x0,x1, ... ,xn-1
] represents the feature map in the 0,1...,n-1 layer; Hl( )
represents the normalized linear correction unit, Relu activation function, pooling
operation and volume Product equivalence transformation. The output of the
convolutional layer and the pooling method formula is as follows:
2.3. LSTM LONG SHORT-TERM MEMORY NETWORK
A Long Short-Term Memory Network (LSTM) is a temporal recurrent neural network
that is an optimization of a Recurrent Neural Network (RNN). RNN is often used to
analyze and predict time series, but it is used for short time series, and it is not
suitable for long-distance and long period time series. LSTM perfectly solves the
shortcomings of RNN. It significantly improves the model's ability to analyze and
predict long sequences by designing hidden layers without changing the original
model structure. The operation state of LSTM is almost linear, and the entire operation
mode is also chain operation, and there will be no problems such as gradient
expansion and disappearance in the RNN training process, which improves the
prediction effect and accuracy. The LSTM model is mainly composed of a forgetting
gate, input gate, and output gate. After the unit gate enters the forgetting gate, the
forgetting gate is responsible for screening out the unit state that can be retained to
the current moment at the previous moment, and the input gate is responsible for
(2)
( )
( )
1l l l l
j j j j
H f down H bβ−
= +
(3)
(9)
(10)
input
output S
λ
λ⎡ ⎤
=⎢ ⎥
⎢ ⎥
1
input
output
F
S
λ
λ−+
⎡ ⎤
=⎢ ⎥
⎢ ⎥
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
147
screening out a certain number of cells. The input at the current moment is the unit
state at the current moment, and the output gate is responsible for the output of the
unit state [17].
Information is mainly selected through three gate structures: input gate, forget gate,
and output gate. Taking the t-th sample as an example is the input data at the current
moment. Equation (13)~(18) is the process performed by the LSTM unit [20]：
Where wf, wi, wc, and wo
are the weights of the corresponding forgetting gate, input
gate, and output gate respectively [18].
2.4. MODEL STRUCTURE DIAGRAM
According to the above model theory, a joint prediction and correction model of
DCNN and LSTM is established, combined with the advantages of the LSTM
algorithm, an LSTM-CNN hybrid model is formed to perform model training on the
data set, and the short text of unknown category is predicted by the trained model. As
shown in Fig.1 [19], this parallel network structure avoids feature loss to the greatest
extent and preserves the global and local information of data features; the residual
structure enhances CNN stability and reduces resource occupancy; batch
normalization layer (Batch normalization, BN) to speed up network training and
suppress network overfitting; the global average pooling layer on the left splices the
dimensionality-reduced feature map with the network feature map on the right, and
finally sends it to the classification layer [20].
(11)
(12)
(13)
[ ]
( )
1,
f f
t t t
f w h x bσ−
=⋅+
[ ]
( )
1
,
i i
t t t
i w h x bσ−
=⋅+
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
148
screening out a certain number of cells. The input at the current moment is the unit
state at the current moment, and the output gate is responsible for the output of the
unit state [17].
Information is mainly selected through three gate structures: input gate, forget gate,
and output gate. Taking the t-th sample as an example is the input data at the current
moment. Equation (13)~(18) is the process performed by the LSTM unit [20]：
Where wf, wi, wc, and wo are the weights of the corresponding forgetting gate, input
gate, and output gate respectively [18].
2.4. MODEL STRUCTURE DIAGRAM
According to the above model theory, a joint prediction and correction model of
DCNN and LSTM is established, combined with the advantages of the LSTM
algorithm, an LSTM-CNN hybrid model is formed to perform model training on the
data set, and the short text of unknown category is predicted by the trained model. As
shown in Fig.1 [19], this parallel network structure avoids feature loss to the greatest
extent and preserves the global and local information of data features; the residual
structure enhances CNN stability and reduces resource occupancy; batch
normalization layer (Batch normalization, BN) to speed up network training and
suppress network overfitting; the global average pooling layer on the left splices the
dimensionality-reduced feature map with the network feature map on the right, and
finally sends it to the classification layer [20].
(11)
(12)
(13)
[ ]
( )
1,
f f
t t t
f w h x bσ−
=⋅+
[ ]
( )
1
,
i i
t t t
i w h x bσ−
=⋅+
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
Figure 1. PLSTM-CNN network model
3. EXPERIMENT AND RESULT ANALYSIS
3.1. EVALUATION RESULTS
To show the fault checking results and evaluate the model performance, two
metrics, Fault Diagnosis Rate (FDR) and False Positive Rate (FPR), were used to
evaluate the model performance. FDR and FPR are defined in [21].
Among them, TP (True Positives) represents the number of instances that are
positive classes and are predicted to be positive classes, FN (False Negative)
represents the number of positive classes that are predicted to be negative, and FP
(False Positives) that instances are negative classes that are predicted to be positive
The number of classes, TN (True Negative) represents the number of instances that
are predicted to be negative classes.
To verify the performance of the proposed method, the diagnostic results of the test
data based on the PLSTM CNN model are compared with the results of the two-
dimensional CNN model-based method and the LSTM model-based method on the
test data, and the results are shown in the table., the average fault diagnosis rate
based on the PLSTM-CNN model in the table is 91.4% [22].
Input
Conv1D (32)
BN + Re1u
Conv1D (32)
BN + Re1u
Conv1D (32)
BN + Re1u
Global pooling
Dimension shuffle
LSTM (16)
Dropout
(0.8)
Concat
Softmax
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
149
Table 1. The comparison of fault results
We find that the classification accuracy of different faults varies widely. The
detection rate of faults 3 and 9 is less than 90%, and the detection rate of faults 15
and 16 is less than 50%. In addition, the detection rates of the remaining 16 faults are
all higher than 90%, of which the detection rate of faults 1, 2, 4, 5, 6, 7, 8, 17, 19, and
20 is 100%. Therefore, PLSTM-CNN can effectively isolate most of the faults, and
only a few faults perform poorly. Through the analysis of the data in the table, it is
found that the lower accuracy of faults 3, 9, 15, and 16 is due to the higher degree of
confusion among them, which is consistent with the previous mutual information
calculation results [23]. The PLSTM-CNN model proposed in this paper makes full use
of the local and global features of the original data, and the fault detection rate and
false-positive rate are better than 2D-CNN and LSTM models in faults 3, 9, 15, and 16
with a high degree of confusion. Because both fault 3 and fault 9 are related to the
Fault type FDR FPR
2D-CNN LSTM PLSTM-CNN 2D-CNN LSTM PLSTM-CNN
Normal 0.91 1.0 1.0 0.08 0.03 0
Fault1 1.0 1.0 1.0 0 0 0
Fault2 1.0 1.0 0.81 0 0 0
Fault3 0.48 1.0 0.92 0.24 0.14 0.07
Fault4 1.0 0.81 0.34 0 0 0
Fault5 1.0 0.92 1.0 0 0 0
Fault6 1.0 0.34 1.0 0 0 0
Fault7 1.0 1.0 1.0 0 0 0
Fault8 1.0 1.0 1.0 0.1 0 0
Fault9 1.0 1.0 1.0 0.58 0.21 0.17
Fault10 1.0 1.0 0.37 0.06 0.01 0.02
Fault11 0.81 1.0 0.33 0.03 0.02 0.01
Fault12 0.92 0.48 0.98 0.05 0.04 0.01
Fault13 0.34 1.0 0.92 0.16 0.05 0.04
Fault14 0.82 1.0 0.89 0.15 0.61 0.05
Fault15 0.96 1.0 1.0 0.79 0.61 0.51
Fault16 0.84 1.0 0.98 0.69 0.62 0.48
Fault17 0.09 1.0 1.0 0.05 0.02 0.02
Fault18 0.96 1.0 1.0 0 0 0
Fault19 1.0 0.48 1.0 0 0 0
Fault20 1.0 1.0 1.0 0 0 0
Average 0.86 0.91 0.88 0.14 0.11 0.07
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
150
Table 1. The comparison of fault results
We find that the classification accuracy of different faults varies widely. The
detection rate of faults 3 and 9 is less than 90%, and the detection rate of faults 15
and 16 is less than 50%. In addition, the detection rates of the remaining 16 faults are
all higher than 90%, of which the detection rate of faults 1, 2, 4, 5, 6, 7, 8, 17, 19, and
20 is 100%. Therefore, PLSTM-CNN can effectively isolate most of the faults, and
only a few faults perform poorly. Through the analysis of the data in the table, it is
found that the lower accuracy of faults 3, 9, 15, and 16 is due to the higher degree of
confusion among them, which is consistent with the previous mutual information
calculation results [23]. The PLSTM-CNN model proposed in this paper makes full use
of the local and global features of the original data, and the fault detection rate and
false-positive rate are better than 2D-CNN and LSTM models in faults 3, 9, 15, and 16
with a high degree of confusion. Because both fault 3 and fault 9 are related to the
Fault type
FDR
FPR
2D-CNN
LSTM
PLSTM-CNN
2D-CNN
LSTM
PLSTM-CNN
Normal
0.91
1.0
1.0
0.08
0.03
0
Fault1
1.0
1.0
1.0
0
0
0
Fault2
1.0
1.0
0.81
0
0
0
Fault3
0.48
1.0
0.92
0.24
0.14
0.07
Fault4
1.0
0.81
0.34
0
0
0
Fault5
1.0
0.92
1.0
0
0
0
Fault6
1.0
0.34
1.0
0
0
0
Fault7
1.0
1.0
1.0
0
0
0
Fault8
1.0
1.0
1.0
0.1
0
0
Fault9
1.0
1.0
1.0
0.58
0.21
0.17
Fault10
1.0
1.0
0.37
0.06
0.01
0.02
Fault11
0.81
1.0
0.33
0.03
0.02
0.01
Fault12
0.92
0.48
0.98
0.05
0.04
0.01
Fault13
0.34
1.0
0.92
0.16
0.05
0.04
Fault14
0.82
1.0
0.89
0.15
0.61
0.05
Fault15
0.96
1.0
1.0
0.79
0.61
0.51
Fault16
0.84
1.0
0.98
0.69
0.62
0.48
Fault17
0.09
1.0
1.0
0.05
0.02
0.02
Fault18
0.96
1.0
1.0
0
0
0
Fault19
1.0
0.48
1.0
0
0
0
Fault20
1.0
1.0
1.0
0
0
0
Average
0.86
0.91
0.88
0.14
0.11
0.07
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
change in the initial number of articles, this paper rearranges the variables according
to the method of the maximum mutual information coefficient based on the initial
number of articles, which further improves the fault detection rate of faults 3 and 9.
The degree of confusion of faults 15 and 16 is too high, and there is no effective fault
detection method [24].
Figure 2. Recognition accuracy of learning rate self-enhancement algorithm
It can be seen from Fig.2 that when the number of iterations is 25, the accuracy
rate is the highest, which is 97.3%, which shows that the use of PLSTM-CNN for
English composition grammar detection has strong applicability. In addition, it can be
seen from Figure 4 that as the iteration continues, the recognition accuracy of the
DCNN model continues to rise. After reaching a certain level, the accuracy does not
change and the network model converges. Therefore, the DCNN model used in this
study has a good effect., it starts to converge after reaching a certain number of
iterations. In this study, this algorithm is used to optimize DCNN with a certain
preprocessing effect [25-27].
3.2. COMPARATIVE EXPERIMENT
To further illustrate the advantages of the parallel neural network structure, we did
the following experiments on average fault detection rate, model training and testing
time, and model stability under small sample conditions.
3.2.1. COMPARISON OF AVERAGE FAILURE DETECTION RATES
To further compare the fault detection capabilities of serial and parallel network
models, based on the data set above, we designed a traditional serial network
structure, as shown in Fig.3. The experimental results are shown in Fig.4. The
average F1 scores on the PLSTM-CNN, tandem LSTM-CNN, LSTM, 1D-CNN, and
2D-CNN models are 92.13%, 89.54%, 84.08%, 84.80%, and 85.78%, respectively.
This shows that parallel CLSTM-CNN has better fault detection performance than
serial LSTM-CNN [28].
0 5 10 15 20 25 30
0.80
0.82
0.84
0.86
0.88
0.90
0.92
0.94
0.96
0.98
1.00
Accuracy
epoch
train
test
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
151
Figure 3. Serial LSTM-CNN network structure
The experimental results are shown in Fig.4. The average F1 scores on the
PLSTM-CNN, tandem LSTM-CNN, LSTM, 1D-CNN, and 2D-CNN models are
92.13%, 89.54%, 84.08%, 84.80%, and 85.78%, respectively. This shows that parallel
CLSTM-CNN has better fault detection performance than serial LSTM-CNN.
Figure 4. Mean failure detection rate
Input
LSTM (16)
Dropout
(0.8)
Conv1D (32)
BN+Re1u
Conv1D (32)
BN+Re1u
Conv1D (32)
BN+Re1u
Global pooling
Softmax
0 10 20 30 40 50
0
20
40
60
80
100
Detection rate
epoch
LSTM-CNN parallel connection
LSTM-CNN series connection
LSTM
1D-CNN
2D-CNN
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
152
Figure 3. Serial LSTM-CNN network structure
The experimental results are shown in Fig.4. The average F1 scores on the
PLSTM-CNN, tandem LSTM-CNN, LSTM, 1D-CNN, and 2D-CNN models are
92.13%, 89.54%, 84.08%, 84.80%, and 85.78%, respectively. This shows that parallel
CLSTM-CNN has better fault detection performance than serial LSTM-CNN.
Figure 4. Mean failure detection rate
Input
LSTM (16)
Dropout
(0.8)
Conv1D (32)
BN+Re1u
Conv1D (32)
BN+Re1u
Conv1D (32)
BN+Re1u
Global pooling
Softmax
0 10 20 30 40 50
0
20
40
60
80
100
Detection rate
epoch
LSTM-CNN parallel connection
LSTM-CNN series connection
LSTM
1D-CNN
2D-CNN
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3.2.2. MODEL INFERENCE AND INFERENCE TIME
COMPARISON
The PLSTM-CNN model takes 4.2 seconds to train for each epoch and 7 minutes
to train for 100 epochs. The main reasons for its faster training speed are: considering
the real-time nature of fault monitoring, the convolution layer in this paper is selected
as one-dimensional convolution, which has fewer parameters. Under the conditions of
the same network and hyperparameters, the calculation time is shorter, but Some
accuracy will be lost.
Meanwhile, training deep 2D convolutional networks usually require special
hardware devices, such as cloud computing or GPU acceleration. But 1D-CNN can be
implemented on the CPU of ordinary computers, and its low computational
requirements and compact structure are very suitable for real-time monitoring and
low-cost applications.
Table 2. Comparison of training and inference time
3.2.3. COMPARISON OF AVERAGE FAULT DETECTION RATES
FOR SMALL SAMPLES
Considering that the actual English composition grammar detection fault samples
are scarce, the experiment will reduce the number of each type of fault sample. Set
the sampling time to 3 minutes, run for 10 hours under normal conditions, and collect
2000 normal samples. In the simulation of 20 kinds of faults, the simulator runs
normally for 1 hour, then introduces the corresponding faults, and then continues to
run for 1 hour. Thus, 1 hour of failure data (200 failure samples) was collected per
simulation. The simulations for each failure type were repeated ten times with ten
different initial states. The simulation platform collects a total of 4200 sample data,
including 2000 normal samples and 2000 samples for each fault. Choose 70% of the
data for training, 20% for testing, and 10% for validation. The experimental results are
shown in Table 3:
Model Training time for
one epoch (s)
Reasoning time for
one epoch (ms)
1D-CNN 2.54 10
2D-CNN 65 200
LSTM 3 12
PC LSTM-CNN 3.8 20
SC-LSTM-CNN 4.2 25
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
153
Table 3. Average failure detection rate of small samples
The study found that SC-LSTM-CNN has a high FDR and a small value of FPR,
which shows that the parallel LSTM-CNN has extremely high prediction accuracy and
stability. Although PC-LSTM-CNN has a high FDR, its FPR value is high, indicating
that its stability is poor and the prediction accuracy is average. The above results
confirm that the parallel LSTM-CNN can still maintain high accuracy on small sample
datasets, and its network structure is more stable than the serial network [29]. 2D-
CNN requires a large number of training samples to guarantee the accuracy, while
1D-CNN still performs well on small-sample datasets[30-31].
4. CONCLUSION
This paper proposes a grammar writing check fault detection method based on the
PLSTM-CNN network model. The model is constructed by LSTM, one-dimensional
dense convolutional layer, one-dimensional global pooling layer, and Dropout layer,
which can effectively extract the local and Global features; after data analysis and
variable reordering based on the maximum information coefficient method, the data
distribution is made more regular and easy to train. The study compares the fault
detection results of PLSTM-CNN, tandem LSTM-CNN, LSTM, and 2D-CNN. The
experimental results show that: (1) the fault detection accuracy and false positive rate
of PLSTM-CNN are significantly better than other methods; (2) for the difficult-to-
detect faults 3 and 9, the PLSTM-CNN model still performs well; (3) parallel
Compared with the serial structure, the PLSTM-CNN structure has better accuracy
and stability, and its FDR and FPR are 90.5% and 0.051, respectively.
5. CONFLICT OF INTEREST
The authors declared that there is no conflict of interest. REFERENCES
REFERENCES
(1) Tao, Y., Shi, H., Song, B., & Tan, S. (2020). A Novel Dynamic Weight Principal
Component Analysis Method and Hierarchical Monitoring Strategy for
Process Fault Detection and Diagnosis. IEEE Transactions on Industrial
Electronics, (99), 1-1. https://doi.org/10.1109/TIE.2019.2942560
Model FDR FPR
1D-CNN 83.4% 0.1
2D-CNN 78.5% 0.08
LSTM 84.6% 0.09
PC LSTM-CNN 84.8% 0.12
SC-LSTM-CNN 90.5% 0.051
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
154
Table 3. Average failure detection rate of small samples
The study found that SC-LSTM-CNN has a high FDR and a small value of FPR,
which shows that the parallel LSTM-CNN has extremely high prediction accuracy and
stability. Although PC-LSTM-CNN has a high FDR, its FPR value is high, indicating
that its stability is poor and the prediction accuracy is average. The above results
confirm that the parallel LSTM-CNN can still maintain high accuracy on small sample
datasets, and its network structure is more stable than the serial network [29]. 2D-
CNN requires a large number of training samples to guarantee the accuracy, while
1D-CNN still performs well on small-sample datasets[30-31].
4. CONCLUSION
This paper proposes a grammar writing check fault detection method based on the
PLSTM-CNN network model. The model is constructed by LSTM, one-dimensional
dense convolutional layer, one-dimensional global pooling layer, and Dropout layer,
which can effectively extract the local and Global features; after data analysis and
variable reordering based on the maximum information coefficient method, the data
distribution is made more regular and easy to train. The study compares the fault
detection results of PLSTM-CNN, tandem LSTM-CNN, LSTM, and 2D-CNN. The
experimental results show that: (1) the fault detection accuracy and false positive rate
of PLSTM-CNN are significantly better than other methods; (2) for the difficult-to-
detect faults 3 and 9, the PLSTM-CNN model still performs well; (3) parallel
Compared with the serial structure, the PLSTM-CNN structure has better accuracy
and stability, and its FDR and FPR are 90.5% and 0.051, respectively.
5. CONFLICT OF INTEREST
The authors declared that there is no conflict of interest. REFERENCES
REFERENCES
(1) Tao, Y., Shi, H., Song, B., & Tan, S. (2020). A Novel Dynamic Weight Principal
Component Analysis Method and Hierarchical Monitoring Strategy for
Process Fault Detection and Diagnosis. IEEE Transactions on Industrial
Electronics, (99), 1-1. https://doi.org/10.1109/TIE.2019.2942560
Model
FDR
FPR
1D-CNN
83.4%
0.1
2D-CNN
78.5%
0.08
LSTM
84.6%
0.09
PC LSTM-CNN
84.8%
0.12
SC-LSTM-CNN
90.5%
0.051
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
(2) A, C. W., A, X. W., Jz, A., Liang, Z. A., Xiao, B. A., Xin, N. B. Ehd, A. (2021).
Uncertainty Estimation for Stereo Matching Based on Evidential Deep
Learning. https://doi.org/10.1016/j.patcog.2021.108498
(3) Cai, W., Zhai, B., Liu, Y., Liu, R., & Ning, X. (2021). Quadratic polynomial
guided fuzzy C-means and dual attention mechanism for medical image
segmentation. Displays, 70, 102106. https://doi.org/10.1016/
j.displa.2021.102106
(4) Ning, X., Duan, P., Li, W., & Zhang, S. (2020). Real-time 3D face alignment
using an encoder-decoder network with an efficient deconvolution layer.
IEEE Signal Processing Letters, 27, 1944-1948. https://doi.org/10.1109/
LSP.2020.3032277
(5) Miao, J., Wang, Z., Ning, X., Xiao, N., Cai, W., & Liu, R. (2022). Practical and
secure multifactor authentication protocol for autonomous vehicles in 5G.
Software: Practice and Experience. https://doi.org/10.1002/SPE.3087
(6) Beruvides, G., Casta?O, F., Quiza, R., & Haber, R. E. (2016). Surface
roughness modeling and optimization of tungsten–copper alloys in micro-
milling processes. Measurement, 246-252. https://doi.org/10.1016/
j.measurement.2016.03.002
(7) Chen, Y., Wang, L., Hu, J., & Ye, M. (2020). Vision-Based Fall Event Detection
in Complex Background Using Attention Guided Bi-directional LSTM.
https://doi.org/10.1109/ACCESS.2020.3021795
(8) Shan, W. (2022). Digital streaming media distribution and transmission
process optimisation based on adaptive recurrent neural network.
Connection Science, 34(1), 1169-1180. https://doi.org/
10.1080/09540091.2022.2052264
(9) Yan, C., Pang, G., Bai, X., Liu, C., Xin, N., Gu, L., & Zhou, J. (2021). Beyond
triplet loss: person re-identification with fine-grained difference-aware
pairwise loss. IEEE Transactions on Multimedia. https://doi.org/10.1109/
TMM.2021.3069562
(10) Hu, X., Liu, T., Hao, X., & Lin, C. (2022). Attention-based Conv-LSTM and Bi-
LSTM networks for large-scale traffic speed prediction. The Journal of
Supercomputing, 1-24. https://doi.org/10.1007/s11227-022-04386-7
(11) Huang, Z., Wei, X., & Kai, Y. (2015). Bidirectional LSTM-CRF Models for
Sequence Tagging. Computer Science. https://doi.org/10.48550/
arXiv.1508.01991
(12) Jin, C., Shi, Z., Li, W., & Guo, Y. (2021). Bidirectional LSTM-CRF Attention-
based Model for Chinese Word Segmentation. https://doi.org/10.48550/
arXiv.2105.09681
(13) Ying, L., Nan, Z. Q., Ping, W. F., Kiang, C. T., Pang, L. K., Chang, Z. H. Nam, L.
(2021). Adaptive weights learning in CNN feature fusion for crime scene
investigation image classification. Connection Science. https://doi.org/
10.1080/09540091.2021.1875987
(14) Li, D., & Lasenby, J. (2021). Spatiotemporal Attention-Based Graph
Convolution Network for Segment-Level Traffic Prediction. IEEE
Transactions on Intelligent Transportation Systems, (99), 1-9. https://doi.org/
10.1109/TITS.2021.3078187
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
155
(15) Li, M., Liu, X., & Xiong, A. (2002). Prediction of the mechanical properties of
forged TC11 titanium alloy by ANN. Journal of Materials Processing
Technology, 121(1), 1-4. https://doi.org/10.1016/S0924-0136(01)01006-8
(16) Liu, Z., Zhou, W., & Li, H. (2019). AB-LSTM: Attention-based Bidirectional
LSTM Model for Scene Text Detection. ACM Transactions on Multimedia
Computing Communications and Applications, 15(4), 1-23. https://doi.org/
10.1145/3356728
(17) Olave, M., Sagartzazu, X., Damian, J., & Serna, A. (2010). Design of Four
Contact-Point Slewing Bearing With a New Load Distribution Procedure to
Account for Structural Stiffness. Journal of Mechanical Design, 132(2),
021006. https://doi.org/10.1115/1.4000834
(18) Tang, D., Wei, F., Nan, Y., Ming, Z., & Bing, Q. (2014). Learning Sentiment-
Specific Word Embedding for Twitter Sentiment Classification. Paper
presented at the Proceedings of the 52nd Annual Meeting of the Association for
Computational Linguistics, 1.
(19) Gao, Y., & Yu, D. (2020). Total variation on horizontal visibility graph and its
application to rolling bearing fault diagnosis. Mechanism and Machine
Theory, 147, 103768. https://doi.org/10.1016/j.mechmachtheory.2019.103768
(20) Nguyen, T. (2019). Spatiotemporal Tile-based Attention-guided LSTMs for
Traffic Video Prediction. https://doi.org/10.48550/arXiv.1910.11030
(21) Sagnika, S., Mishra, B., & Meher, S. K. An attention-based CNN-LSTM model
for subjectivity detection in opinion-mining. Neural Computing and
Applications, 1-14. https://doi.org/10.1007/s00521-021-06328-5
(22) Shan, X., Wang, Y., Dong, M., & Xia, J. (2021). Application Research and
Analysis of Geographic Information System in Intelligent City Surveying
and Mappinge. Journal of Physics: Conference Series, 1881(4), 042071. https://
doi.org/10.1088/1742-6596/1881/4/042071
(23) Shi, X., & Wang, B. (2021). Application of New Surveying and Mapping
Technology in the Construction of Smart City. E3S Web of Conferences
, 236,
04031. https://doi.org/10.1051/e3sconf/202123604031
(24) Shi, Z. L., Gong, Y., Cao, M., & Xiao, S. (2010). Discussion on the Application
of Surveying and Mapping Technology in the Internet of Things Times.
Modern Surveying and Mapping, 65(4), 503-515. https://doi.org/10.1016/
j.neuron.2010.01.035
(25) Andrejic, M., Bojovic, N., & Kilibarda, M. (2016). A framework for measuring
transport efficiency in distribution centers. Transport Policy, 45(JAN.),
99-106. https://doi.org/10.1016/j.tranpol.2015.09.013
(26) Bergstrom, J. C., Braden, J. B., & Kolstad, C. D. (1991). Measuring the
demand for environmental quality. American Journal of Agricultural
Economics, 75(1), 244. https://doi.org/10.2307/1242975
(27) Brock, W. A., & Taylor, M. S. (2005). Economic Growth and The Environment:
A Review of Theory and Empirics. Handbook of Economic Growth. https://
doi.org/10.1016/S1574-0684(05)01028-2
(28) Wang, M., Zhou, J., Gao, J., Li, Z., & Li, E. (2020). Milling Tool Wear
Prediction Method Based on Deep Learning under Variable Working
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
156
(15) Li, M., Liu, X., & Xiong, A. (2002). Prediction of the mechanical properties of
forged TC11 titanium alloy by ANN. Journal of Materials Processing
Technology, 121(1), 1-4. https://doi.org/10.1016/S0924-0136(01)01006-8
(16) Liu, Z., Zhou, W., & Li, H. (2019). AB-LSTM: Attention-based Bidirectional
LSTM Model for Scene Text Detection. ACM Transactions on Multimedia
Computing Communications and Applications, 15(4), 1-23. https://doi.org/
10.1145/3356728
(17) Olave, M., Sagartzazu, X., Damian, J., & Serna, A. (2010). Design of Four
Contact-Point Slewing Bearing With a New Load Distribution Procedure to
Account for Structural Stiffness. Journal of Mechanical Design, 132(2),
021006. https://doi.org/10.1115/1.4000834
(18) Tang, D., Wei, F., Nan, Y., Ming, Z., & Bing, Q. (2014). Learning Sentiment-
Specific Word Embedding for Twitter Sentiment Classification. Paper
presented at the Proceedings of the 52nd Annual Meeting of the Association for
Computational Linguistics, 1.
(19) Gao, Y., & Yu, D. (2020). Total variation on horizontal visibility graph and its
application to rolling bearing fault diagnosis. Mechanism and Machine
Theory, 147, 103768. https://doi.org/10.1016/j.mechmachtheory.2019.103768
(20) Nguyen, T. (2019). Spatiotemporal Tile-based Attention-guided LSTMs for
Traffic Video Prediction. https://doi.org/10.48550/arXiv.1910.11030
(21) Sagnika, S., Mishra, B., & Meher, S. K. An attention-based CNN-LSTM model
for subjectivity detection in opinion-mining. Neural Computing and
Applications, 1-14. https://doi.org/10.1007/s00521-021-06328-5
(22) Shan, X., Wang, Y., Dong, M., & Xia, J. (2021). Application Research and
Analysis of Geographic Information System in Intelligent City Surveying
and Mappinge. Journal of Physics: Conference Series, 1881(4), 042071. https://
doi.org/10.1088/1742-6596/1881/4/042071
(23) Shi, X., & Wang, B. (2021). Application of New Surveying and Mapping
Technology in the Construction of Smart City. E3S Web of Conferences, 236,
04031. https://doi.org/10.1051/e3sconf/202123604031
(24) Shi, Z. L., Gong, Y., Cao, M., & Xiao, S. (2010). Discussion on the Application
of Surveying and Mapping Technology in the Internet of Things Times.
Modern Surveying and Mapping, 65(4), 503-515. https://doi.org/10.1016/
j.neuron.2010.01.035
(25) Andrejic, M., Bojovic, N., & Kilibarda, M. (2016). A framework for measuring
transport efficiency in distribution centers. Transport Policy, 45(JAN.),
99-106. https://doi.org/10.1016/j.tranpol.2015.09.013
(26) Bergstrom, J. C., Braden, J. B., & Kolstad, C. D. (1991). Measuring the
demand for environmental quality. American Journal of Agricultural
Economics, 75(1), 244. https://doi.org/10.2307/1242975
(27) Brock, W. A., & Taylor, M. S. (2005). Economic Growth and The Environment:
A Review of Theory and Empirics. Handbook of Economic Growth. https://
doi.org/10.1016/S1574-0684(05)01028-2
(28) Wang, M., Zhou, J., Gao, J., Li, Z., & Li, E. (2020). Milling Tool Wear
Prediction Method Based on Deep Learning under Variable Working
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
Conditions. IEEE Access, 99, 1-1. https://doi.org/10.1109/
ACCESS.2020.3010378
(29) Amin, T., Khan, F., Ahmed, S., & Imtiaz, S. (2020). A novel data-driven
methodology for fault detection and dynamic risk assessment. The
Canadian Journal of Chemical Engineering. https://doi.org/10.1002/cjce.23760
(30) Horani M. O., Najeeb, M., y Saeed, A. (2021). Model electric car with wireless
charging using solar energy. 3C Tecnología. Glosas de innovación aplicadas
a la pyme, 10(4), 89-101. https://doi.org/10.17993/3ctecno/
2021.v10n4e40.89-101
(31) Chang Jingying,Lan Weibin & Lan Wenhao. (2021). Higher education
innovation and reform model based on hierarchical probit. Applied
Mathematics and Nonlinear Sciences, 7(1), 175-182. https://doi.org/10.2478/
AMNS.2021.2.00154
https://doi.org/10.17993/3ctecno.2023.v12n1e43.142-157
3C Tecnología. Glosas de innovación aplicadas a la pyme. ISSN: 2254-4143
Ed.43 | Iss.12 | N.1 January - March 2023
157