https://doi.org/10.31449/inf.v46i3.3961                                                                                      Informatica 46 (2022) 343-354     343 
Application and Study of Artificial Intelligence in Railway Signal 
Interlocking Fault  
 
Hongwei Liang
1
*, Xiuxuan Wang
1
, Anjali Sharma
2
, Mohd Asif Shah
3
 
1
Zhengzhou Railway Vocational & Technical College, Zhengzhou, Henan ,451460, China 
2
School of Biological and Environmental Sciences, Shoolini University of Biotechnology and Management Sciences, 
Solan 173229 (H.P.), India 
3
Bakhtar University, Kabul, Afganistan 
Emails: hongweiliang7@163.com, xiuxuanwang9@126.com,  anjalisharmaas8749347@gmail.com, 
ohaasif@bakhtar.edu.af 
 
Keywords: Railway signal equipment; ADASYN data synthesis; Deep learning; Integrated learning; Fault diagnosis. 
 
Received: February 2, 2022 
 
The rapid development of railway transportation towards high speed, high density and heavy load has 
led to even higher requirements for the safety of railway signal equipment. The safety of railway signal 
equipment is an important part of ensuring railway traffic safety, thus, it is very necessary to study a 
system that can diagnose the fault of railway signal equipment according to the actual situation. This 
article utilizes the deep learning algorithm of artificial intelligence for investigating the interlocking 
faults in the railway transportation. This paper uses ADASYN data synthesis method to synthesize few 
category samples, uses TF-IDF to extract features and transform vectors, and proposes a deep learning 
integration method based on combined weight. The results show that BiGRU has better overall 
classification performance when evaluated on the index of primary and secondary fault classification 
accuracy. The classification accuracy improvement of 5% is achieved for primary fault classification 
and the comprehensive evaluation index of secondary fault classification is improved by about 9%. It 
was revealed that when compared with ADASYN + BiLSTM neural network, the comprehensive 
evaluation index of primary fault classification accuracy is improved by about 6%, and the 
comprehensive evaluation index of secondary fault classification is improved by about 10%. It is 
demonstrated that deep learning integration is an effective method to improve the classification 
performance of turnout fault diagnosis model. 
 
Povzetek: Za železniški sistem je bila uporabljena metodologija globokih nevronskih mrež za iskanje 
napak v signalih.
  
1 Introduction  
With the gradual increase of railway traffic density 
and operation speed in China, it is difficult to avoid 
various faults of railway signal equipment. If the faults 
cannot be handled in a short time, they will have a great 
impact on traffic safety, and even lead to the hidden 
dangers of major accidents, so as to reduce the efficiency 
and safety of railway operation. At the same time, it also 
brings new challenges to railway signal equipment 
maintenance personnel to check and maintain signal 
equipment timely and accurately. 
High speed railway signal equipment is an 
important infrastructure to ensure high-speed train 
operation. The maintenance quality of signal equipment 
directly affects the traffic safety and transportation 
efficiency of high-speed railway. Signal equipment fault 
is diagnosed and handled according to the experience and 
knowledge of on-site maintenance personnel, which is 
easy to cause maintenance judgment error and 
maintenance time delay, and in serious cases, it will lead 
to equipment fault driving accident. The fault data of 
high-speed railway signal equipment records the fault 
phenomenon when the fault occurs in the form of text. 
The fault phenomenon is analyzed based on text data 
mining technology. Combined with the diagnosis results 
of experts on the fault phenomenon, the fault diagnosis 
model of signal equipment is studied to assist 
maintenance personnel to quickly locate the fault 
location and cause according to the fault phenomenon. It 
will be of great significance to further improve the safety 
guarantee level of high-speed railway. The basic activity 
diagram of train fault detection method is shown in 
Figure 1. 

344     Informatica 46 (2022) 343-354                                                                                                                              H. Liang et al. 
 
Figure 1: Activity diagram of Railway fault detection 
method 
 
This limitation of imbalanced faults of different 
signal equipment is addressed in this article. In order to 
study the signal equipment fault diagnosis method based 
on unbalanced samples based on text mining technology, 
two problems need to be solved: one is the processing of 
unbalanced samples, and the other is the construction of 
fault diagnosis and classification model.  
This article contributed in mainly using two 
methods to solve the sample imbalance problem: one is 
to synthesize the sample data by using data enhancement, 
under sampling or oversampling, and data generation 
methods such as SMOTE (Synthetic Minority 
Oversampling Technology) and ADASYN (Adaptive 
Synthetic Sampling). The other is to adjust the 
parameters of different categories for the classification 
learning algorithm. The sample synthesis algorithm can 
appropriately synthesize a few categories of samples 
according to the distribution of the overall samples, and 
can ensure that the sample data is not repeated. There are 
several articles which uses SVM-SMOTE method to 
automatically synthesize the few category samples of 
signal equipment fault, so as to solve the problem of 
signal equipment fault sample imbalance. This article 
utilizes the deep learning algorithm of artificial 
intelligence for investigating the interlocking faults in the 
railway transportation. This paper uses ADASYN data 
synthesis method to synthesize few category samples, 
uses TF-IDF to extract features and transform vectors, 
and proposes a deep learning integration method based 
on combined weight. The outcomes obtained for the 
proposed method reveals that BiGRU has better overall 
classification performance when evaluated on the index 
of primary and secondary fault classification accuracy. 
The rest of this article is structured as: review of 
literature is provided in section 2 followed by research 
methodology involved in analysis of fault diagnosis of 
railway unlocking system in section 3. Section 4 provides 
the experimental results and discussion along with 
concluding remarks in section 5.  
 
2 Related work 
In this section various state-of-the-art work in the 
field of railway signal interlocking fault based on 
artificial intelligence and other technologies is presented.  
With the advent of the intelligent era, artificial 
intelligence has become the mainstream technology in 
the world, and artificial intelligence technology has laid a 
solid research foundation [1]. Paek and Kim explores the 
future direction of education by examining the current 
impact of artificial intelligence and predicting the future 
impact [2]. Interlocking is a railway system, which can 
automatically control safety management route change 
and avoid train collision and derailment. Dobias and 
Kubatova analyzes the latest technologies used in several 
commercial interlocking equipment, and proposed the 
design and implementation of an interlocking system 
architecture based on FPGA technology [3]. In order to 
solve the problem of channel estimation based on 
demodulated reference signal (DMRs) in railway tunnel 
scene, Skiribou et al. proposed a deterministic model to 
accurately generate time-varying channel response [4]. 
Kiedrowski and Saganowski introduced a scheme of 
applying PLC technology to railway light signs. This 
paper introduces the structure of the network and a group 
of equipment to realize this specific type of wired sensor 
network, which is used to monitor the railway led sign 
network and maintenance parameters [5]. Yang et al.  
analyzed the requirements of clock synchronization of 
signal ground equipment in combination with the 
application status of clock synchronization of ground 
equipment in high-speed railway signal system. By 
analyzing the advantages and disadvantages of the 
world's mainstream satellite navigation system and the 
requirements of China's railway signal system, Beidou 
time service technology is selected as the clock 
synchronization technology of the ground equipment of 
high-speed railway signal system, and the overall scheme 
based on Beidou time service technology is constructed 
[6]. In order to evaluate the network access performance 
of railway signal equipment machine communication 
(MTC) in the next generation intelligent transportation 
system, Lin et al. divided the railway signal equipment 
machine communication traffic prediction model into 
station indoor model, station outdoor model and station 
outdoor model, and calculated the traffic and signaling 
overhead of the three models respectively. Based on 
Poisson distribution and Markov renewal process, an 
improved Markov modulated poisson process (immpp) 
for source traffic model is designed [7]. Wang et al.  
combined with the new technical characteristics of high-
speed railway, analyzed the current situation of lightning 
protection technology and lightning faults of foreign 
railway signal equipment. At the same time, the 
functions of intelligent technologies such as lightning 
activity location and lightning fault diagnosis are 

Application and Study of Artificial Intelligence in Railway…                                                Informatica 46 (2022) 343-354     345 
introduced, and the development direction of railway 
lightning protection in the future is prospected according 
to the characteristics of this technology [8]. In order to 
realize the real-time acquisition, monitoring and 
management of the technical status of railway signal 
equipment and meet the multi-dimensional business 
needs of railway signal system information sharing, data 
mining, analysis and display, Sahal et al. put forward the 
national technical big data platform of railway signal 
equipment on the basis of analyzing the current situation 
of railway signal system and the significance of signal 
big data platform construction [9]. Based on the common 
signal system equipment of rail transit stations at home 
and abroad, Cao et al. analyzed the common faults and 
their settings of the system, studied the common faults 
analysis, design and construction of the signal system, 
and developed the railway signal fault setting training 
system based on the core concept of fault safety design 
[10]. In order to solve the problem of railway 
transportation safety, Dong et al. carried out detection 
experiments on simulated images and real videos of 
railway signal lights based on machine vision. The image 
features of railway signal lights in different color spaces 
and their influence on railway signal light recognition are 
discussed [11]. 
Railway signal equipment safety is an important 
part of ensuring railway traffic safety, thus, it is very 
necessary to study a system that can diagnose the fault of 
railway signal equipment according to the actual 
situation. The literature suggests that there are many 
studies on using data synthesis method to solve the 
sample imbalance based on the deep learning of artificial 
intelligence approach [12-15]. This paper diagnoses the 
fault of high-speed railway signal equipment, improves 
the performance of equipment fault diagnosis, so as to 
improve the safety of railway.  
 
3 Research methods 
This section includes the description of small 
category sample generation based on ADASYN. The 
fault text features of high-speed railway signal are also 
represented in this section and fault diagnosis model is 
presented.  
High speed railway signal fault diagnosis forms a 
turnout fault diagnosis model with deliverable evaluation 
indexes through the training and optimization of the fault 
diagnosis model based on deep learning integration [16]. 
The turnout fault phenomenon of high-speed railway is 
input into the fault diagnosis model, and the model 
automatically outputs the type and cause of the fault, so 
as to realize the intelligent diagnosis of turnout 
equipment fault [17-19]. The architecture of this research 
work is depicted in Figure 2.  
 
Figure 2: Architecture of research work 
 
The basic structure of this research work includes 
pre-processing of data acquired from various sources. 
Further, the feature set is extracted followed by the 
classification of primary and secondary faults [20, 21]. 
At the final stage, accuracy values are determined for the 
proposed architecture. The development of artificial 
intelligence and Internet of Things is considered for 
several industrial applications and contributing towards 
social life [22-25]. 
3.1 Small category sample generation 
based on ADASYN 
ADASYN adaptive synthesis oversampling method 
is to adaptively synthesize a small number of samples 
according to the distribution of a small number of 
samples, and synthesize fewer samples where it is easy to 
classify and more samples where it is difficult to classify. 
The key of the synthesis algorithm is to find a probability 
distribution𝑟 𝑖 . Put 𝑟 𝑖 is the criterion for determining how 
many samples should be synthesized for each small 
category sample. 
The proportion of the number of secondary 
categories included in each primary fault category of 
high-speed railway signal turnout fault is 12:17:8: 
11:7:1:7. Therefore, ADASYN is used to synthesize 
fewer secondary fault category samples, and the 
imbalance of primary fault categories can be solved at 
the same time. The process of using ADASYN to 
adaptively generate turnout secondary few category 
samples is as follows: 
Step 1: Calculate the unbalance degree of few 
categories, 𝑑 = 𝑚 𝑠 /𝑚 𝑙 ，𝑚 𝑠 and 𝑚 𝑙 represent the 
number of samples with few categories and multiple 
categories respectively, 𝑑 ∈ ( 0, 1]. 
Step 2: Calculate the total number of small category 
samples to be synthesized, 𝐺 = ( 𝑚 𝑙 − 𝑚 𝑠 )× β ，β ∈
( 0, 1], indicating the expected imbalance degree of the 
whole sample after adding the synthetic sample, β= 1 
means that the sample category is completely balanced 
after adding the synthetic sample. 
Step 3: For each sample of a few categories𝑥 𝑖 . Find 
their K-nearest neighbors in n-dimensional space and 

346     Informatica 46 (2022) 343-354                                                                                                                              H. Liang et al. 
calculate the ratio𝑟 𝑖 = ∆
𝑖 /𝐾 ( 𝑖 = 1,2, … , 𝑚 ) , 𝑚 is the 
total number of samples, ∆
𝑖 is the number of multiclass 
samples in the k-nearest neighbor of𝑥 𝑖 , so 𝑟 𝑖 ∈ ( 0, 1] 。 
Step 4: According to 𝑟 𝑖 ̂ = 𝑟 𝑖 / ∑ 𝑟 𝑖 𝑚 𝑠 𝑖 =1
, regularize 𝑟 𝑖 . 
So 𝑟 𝑖 is the probability distribution, and ∑ 𝑟 𝑖 ̂ = 1. 
Step 5: Calculate the number of samples 𝑔 𝑖 = 𝑟 𝑖 ̂ ×
𝐺 to be synthesized for each small category sample 𝑥 𝑖 , 𝐺 
is the total number of synthetic samples. 
Step 6: According to the above steps, calculate the 
number of samples 𝑔 𝑖 synthesized by each small category 
sample𝑥 𝑖 . 
3.2 Fault text feature representation of 
high-speed railway signal equipment 
TF-IDF is a text feature representation method 
based on weighting idea. Its core idea is that if a word 
appears frequently in one document and low in other 
documents, it indicates that the word has high 
recognition in the document and assigns its high weight. 
The feature extraction of signal equipment fault text first 
needs to realize Chinese word segmentation [26-29]. 
Because the high-speed railway signal equipment fault 
text data contains professional words such as switch 
machine, red light band and sealer, this paper constructs 
railway signal professional thesaurus and loads the 
thesaurus into Jieba word segmentation tool to realize the 
accurate word segmentation of fault text. 
Text frequency (TF) in TF-IDF refers to the 
frequency of a given word in the document. For a given 
word 𝑡 𝑖 . In a document 𝑑 𝑗 , the degree of importance can 
be expressed as: 
 
𝑇𝐹
𝑖 ,𝑗 =
𝑛 𝑖 ,𝑗 ∑𝑛 𝑘 ,𝑗 𝑘 
(1) 
Where: 𝑛 𝑖 ,𝑗 is the number of occurrences of the i-th 
word in document 𝑑 𝑗 . ∑𝑛 𝑘 ,𝑗 𝑘 is the total number of 
occurrences of each word in document 𝑑 𝑗 . 
The inverse file frequency IDF is a measure of the 
general importance of a word. Its calculation formula is 
as follows. The larger the IDF, the better the ability to 
distinguish categories. 
 
𝐼𝐷𝐹 𝑖 = log
2
|𝐷 |
1 + |𝑗 : 𝑡 𝑖 ∈ 𝑑 𝑗 |
 
(2) 
 
Where: D is the total number of sample files, 
|𝑗 : 𝑡 𝑖 ∈ 𝑑 𝑗 | contains the number of documents in the 
word. If the word is not in the sample, it will cause the 
denominator to be zero. Therefore, adding 1 to the 
denominator is to avoid the situation that the 
denominator is 0. 
𝑊 𝑖 ,𝑗 = 𝑇𝐹
𝑖 ,𝑗 × 𝐼𝐷𝐹
𝑖 . Weight 𝜔 𝑖 ,𝑗 of words is 
obtained by multiplying the word frequency in the 
document by the low file frequency of the word in the 
whole document set. 
According to the TF-IDF feature weight calculation 
method, the characteristics of turnout fault samples based 
on text are calculated. The characteristics of a turnout 
fault sample are expressed as 𝑑 𝑖 = [𝜔 𝑖 1
 𝜔 𝑖 2
… 𝜔 𝑖 𝑚 ], m is 
the length of the sample, and the primary fault category 
and secondary fault category are expressed as matrix 𝒚 1
 
and 𝑦 2
 by one hot coding vectorization, 𝑦 𝑖 =
[0 1 0 … 𝑐 − 1], 𝑐 is the total number of categories, and 
the fault level I category feature is expressed as 𝐷 𝐿 1
=
[𝑑 𝑖  𝑦 1
], (i=1,2,…, n), n represents the total number of 
samples. The label of fault level I is also input into the 
feature vector by Fault secondary feature as a feature, 
𝐷 𝐿 2
= [[𝑑 𝑖  𝑦 1
] 𝑦 2
]. 
3.3 Deep learning integrated fault 
diagnosis model 
Integrated learning is to combine multiple weak 
supervised learning models to get a better and more 
comprehensive supervised learning model. The high-
speed railway turnout fault diagnosis model adopts 
BiGRU and BiLSTM neural networks as the weak 
supervised learning model, inputs the feature vectors 
extracted from the features into the embedded layer of 
BiGRU and BiLSTM neural networks respectively, and 
the two neural networks output the classification and 
prediction probability of the feature vectors in the 
Softmax layer through learning. The prediction results of 
the two neural networks are integrated and calculated by 
the combined weighted integration method, and finally 
the classification results of the input data by the deep 
learning integration model are output [30]. 
GRU and LSTM are variants of RNN neural 
network. Gating units are designed in neurons to 
effectively calculate and control the input and output of 
information, as shown in Figure 3. The design of this 
gating unit solves the problem of text sequence length 
dependence. Since the output of sigmoid function is 0 ~ 
1, 1 can mean that the information is retained, and 0 
means that the information is discarded, GRU and LSTM 
process the input information through sigmoid function, 
and tanh function processes the output information. 

Application and Study of Artificial Intelligence in Railway…                                                Informatica 46 (2022) 343-354     347 
 
Figure 3: Structural units of RNN and its variant neurons 
 
 
LSTM neural unit is composed of three gates, 
namely forgetting gate, input gate and output gate, as 
shown in Figure 3. LSTM first determines which 
information needs to be discarded through the forgetting 
gate, and calculates ℎ
𝑡 −1
 𝑥 𝑖 and output a vector between 0 
and 1, the vector represents what information neuron 
𝐶 𝑡 −1
 retains or discards. Then, the input gate is used to 
determine which information needs to be added in the 
neuron, and the candidate neuron 𝐶 𝑡 ̃
 is obtained by tanh's 
calculation using ℎ
𝑡 −1
 and 𝑥 𝑖 , which can be updated into 
the neuron. Finally, the output information is controlled 
by the output gate, and the LSTM neuron output is 
finally obtained by multiplying the 0 ~ 1 vector obtained 
by the output layer 𝑜 𝑡 and the neuron through the tanh 
layer. 
 
𝑓 𝑡 = 𝜎 ( 𝑊 𝑓 ∙ [ℎ
𝑡 −1
  𝑥 𝑖 ] + 𝑏 𝑓 ) (3) 
𝑖 𝑡 = 𝜎 ( 𝑊 𝑖 ∙ [ℎ
𝑡 −1
  𝑥 𝑖 ] + 𝑏 𝑖 ) (4) 
𝐶 𝑡 ̃
= 𝑡𝑎𝑛 ℎ( 𝑊 𝑐 ∙ [ℎ
𝑡 −1
  𝑥 𝑖 ] + 𝑏 𝑐 ) (5) 
𝐶 𝑡 = 𝑓 𝑡 ∗ 𝐶 𝑡 −1
+ 𝑖 𝑡 ∗ 𝐶 𝑡 ̃
 (6) 
𝑜 𝑡 = 𝜎 ( 𝑊 𝑜 ∙ [ℎ
𝑡 −1
  𝑥 𝑖 ] + 𝑏 𝑜 ) (7) 
ℎ
𝑡 = 𝑜 𝑡 ∗ tanh ( 𝐶 𝑡 ) (8) 
 
 
where: * is Hadamard product operator, which means 
multiplication of elements at the same position of the 
matrix. 
GRU is a variant of LSTM, as shown in Figure 3. It 
combines the forgetting gate and input gate into an 
update gate 𝑧 𝑡 . 𝑧 𝑡 controls how much information needs 
to be forgotten from the previous hidden layer ℎ
𝑡 −1
, how 
much information needs to be added to the current 
hidden layer ℎ
𝑡 ̃
, and then obtains ℎ
𝑡 . Reset gate 𝑟 𝑡 
controls how much previous information needs to be 
retained. When 𝑟 𝑡 is 0, ℎ
𝑡 ̃
 only contains the information 
of the current word. 
 
𝑧 𝑡 = 𝜎 ( 𝑊 𝑧 ∙ [ℎ
𝑡 −1
  𝑥 𝑡 ]) (9) 
𝑟 𝑡 = 𝜎 ( 𝑊 𝑟 ∙ [ℎ
𝑡 −1
  𝑥 𝑡 ]) (10) 
𝑟 𝑡 = tanh ( 𝑊 ∙ [𝑟 𝑡 ∗ ℎ
𝑡 −1 
  𝑥 𝑡 ]) (11) 
ℎ
𝑡 = ( 1 − 𝑧 𝑡 )∗ ℎ
𝑡 −1
+ 𝑧 𝑡 ∗ ℎ
𝑡 ̃
 
(12) 
 
The combination weighted integration method of 
LSTM and GRU combines the overall classification 
performance of a single neural network with the 
classification performance of each category by assigning 
weights. The combination weighted integration method 
includes overall weight and category weight. The higher 

348     Informatica 46 (2022) 343-354                                                                                                                              H. Liang et al. 
the overall classification performance of a single neural 
network, the higher the overall weight will be allocated. 
According to formula (13) and formula (14), the lower 
the error proportion of neural network in category 
classification, the better classification performance it has 
in this category, the higher the category weight will be 
allocated. Then add the overall weight of the neural 
network and the category weight according to equation 
(15) to recalculate the predicted value of the neural 
network in each category. This combined weighted 
integration method can avoid the influence of few values 
and extreme values in the integration method. 
 
𝜖 𝑖𝑗
=
𝑒𝑟𝑟𝑜𝑟 𝑁𝑢𝑚 𝑖𝑗
 𝑡𝑒𝑥𝑡 𝑁𝑢𝑚 𝑖𝑗
 
(13) 
𝛼 𝑖𝑗
= {
0             𝜖 𝑖𝑗
≥ 0.5
ln(
1−𝜖 𝑖𝑗
𝜖 𝑖𝑗
)     𝜖 𝑖𝑗
<0.5
 
(14) 
𝑃 𝑖 = ∑ ( 𝜔 𝑗 + 𝛼 𝑖𝑗
)
𝑛 𝑗 =1
∙ 𝑃 𝑖𝑗
 
(15) 
Where: 𝜖 𝑖𝑗
 is the classification error ratio of neural 
network ｊ in category i. 𝑡𝑒𝑥𝑡 𝑁𝑢𝑚 𝑖𝑗
 is the total number 
of samples of category i; e𝑟𝑟 𝑜𝑟 𝑁𝑢𝑚 𝑖𝑗
 is the number of 
classification error samples of neural network ｊ in 
category i. 𝛼 𝑖𝑗
 is the category weight of neural network 
ｊ in category i., 𝜔 𝑗 is the overall weight of neural 
network j, and ∑ 𝜔 𝑗 𝑛 𝑗 =1
. 
In order to improve the generalization ability of 
deep learning integration model, K-fold cross validation 
training model is adopted. K-fold cross validation is to 
randomly divide the whole training sample into K parts, 
one of which is used as the validation set and the other 
K-1 is used as the training set, and cycle K times until all 
data are selected once. 
 
4 Results and Analysis 
This section illustrates the result and analysis of 
overall weight distribution, weight calculation and the 
classification of deep learning integration model. 
4.1 Overall weight distribution of BiGRU 
and BiLSTM 
BiGRU and BiLSTM have the same network 
parameters, in which the embedded layer dimension is 
100, the hidden layer dimension is 512, K-fold cross 
validation K = 4, the number of iterations is 50, and the 
batch size is 256. After TF-IDF feature extraction and 
vector representation, the training set and verification set 
synthesized by ADASYN are input into BiGRU and 
BiLSTM networks for training. The change of loss 
function value in the training process of the two neural 
networks is shown in Figure 4. It can be seen from 
Figure 4 that with the increase of iteration times, the loss 
value of BiGRU is lower than that of BiLSTM, 
indicating that its overall classification performance is 
better. In the two neural networks, the loss function value 
of the primary classification is lower than that of the 
secondary classification, indicating that the evaluation 
index of the primary classification of the neural network 
is higher than that of the secondary classification. Both 
neural networks are between 40 ~ 50 iteration rounds, 
and the loss function value tends to be stable, indicating 
that the number of iteration rounds of 50 can make the 
neural network training reach the best state. 
 
 
(a): BiGRU primary classification training process 
 
 
(b): BiGRU primary classification training process 

Application and Study of Artificial Intelligence in Railway…                                                Informatica 46 (2022) 343-354     349 
 
(c): BiGRU secondary classification training process 
 
 
 
(d): BiGRU primary classification training process 
Figure 4 (a, b, c, d): Variation of loss value in K-cross 
training of BiGRU and BiLSTM neural networks 
 
After K = 4 training, 30% real samples are used to 
evaluate BiGRU and BiLSTM training models. The 
evaluation results are shown in Table 1 and is graphically 
presented in Figure 5.  
 
Method Level 
Accuracy 
rate 
Recall 
rate 
F1 value 
ADASYN
＋BiGRU 
Primary fault 
classification 
0.8742 0.8814 0.8779 
Secondary 
fault 
classification 
0.7828 0.7421 0.7619 
ADASYN
＋BiLSTM 
Primary fault 
classification 
0.8613 0.8765 0.8688 
Secondary 
fault 
classification 
0.7601 0.7581 0.7591 
BiGRU 
Primary fault 
classification 
0.7317 0.7098 0.7206 
Secondary 
fault 
classification 
0.7081 0.6712 0.6891 
BiLSTM 
Primary fault 
classification 
0.6912 0.7129 0.7019 
Secondary 
fault 
classification 
0.6371 0.6214 0.6292 
Table 1: Test results of K-fold cross validation + BiGRU 
and BiLSTM neural network 
 
  
 

350     Informatica 46 (2022) 343-354                                                                                                                              H. Liang et al. 
 
Figure 5: Graphical results of K-fold cross validation + BiGRU and BiLSTM neural network 
 
It can be seen from Table 1 that after using 
ADASYN less category synthesis method, the evaluation 
indexes of BiGRU network are higher than BiLSTM 
network under the same parameters, so BiGRU network 
should be assigned a higher overall weight. The original 
samples are trained with the same network structure and 
parameters. The test results are shown in Table 1. It can 
be seen that after ADASYN synthesizes a small number 
of samples, the classification indexes of the two neural 
networks are significantly improved, the first-class rating 
indexes of BiGRU network with good performance are 
increased by nearly 15%, and all evaluation indexes of 
BiGRU network are higher than those of BiLSTM 
network. It is further concluded that the performance of 
BiGRU is better than BiLSTM, and a higher overall 
weight can be assigned to BiGRU network. 
4.2 Weight calculation of BiGRU and 
BiLSTM  
In order to more comprehensively obtain the 
performance of neural network in each category 
classification, a few category samples synthesized by 
ADASYN and all real samples are used. A total of 6327 
samples are input into the trained ADASYN + BiGRU 
and ADASYN + BiLSTM neural networks. The category 
weight calculation results of the two neural networks in 
the primary classification are shown in Table 2.  
It can be seen from Table 2 that although BiGRU 
has higher overall evaluation index and higher overall 
weight than BiLSTM, the performance of the two neural  
 
 
networks are different in each category. BiLSTM has a 
larger category weight in the categories of security 
inspector, public works equipment and unknown reason, 
indicating that BiLSTM network has decision-making 
power in these three categories. Due to the large number 
of secondary classification categories of signal turnout 
equipment faults, considering the length, this paper only 
lists the weight calculation results of primary 
classification categories. 
4.3 Deep learning integration model and 
classification  
The various weights of the neural network are 
obtained through the above tests, and BiGRU should 
have higher overall weight than BiLSTM. Different 
overall weights are given to BiGRU and BiLSTM. The 
two deep learning neural networks are integrated through 
combined weighting, and the common classification 
prediction results are obtained through recalculation of 
the outputs of the two networks.  
Under different overall weight distribution, see 
Figure 6 for the evaluation indexes of level 1 fault 
classification and level 2 fault classification of the deep 
learning integration model (where G represents BiGRU 
and L represents BiLSTM). It can be seen from Figure 6 
that when the overall weight of BiGRU is 0.54 and the 
overall weight of BiLSTM is 0.46, the evaluation index 
of the deep learning integration model is the highest. The 
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
ADASYN ＋BiGRU
ADASYN ＋BiLSTM
BiGRU
BiLSTM
ADASYN ＋BiGRU
ADASYN ＋BiLSTM
BiGRU
BiLSTM
Primary fault classification Secondary fault classification
Accuracy rate Recall rate F1 value

Application and Study of Artificial Intelligence in Railway…                                                Informatica 46 (2022) 343-354     351 
final classification results of the deep learning integration 
model are shown in Table 3 and Figure 7.  
 
 
 
Classification Classification method 
Number of classification 
errors / total number of 
categories 
Recall rate Category weight 
Switch machine 
ADASYN ＋BiGRU 266/2053 0.1295 1.9048 
ADASYN ＋BiLSTM 288/2053 0.1403 1.8129 
External locking 
and installation 
device 
ADASYN ＋BiGRU 163/1251 0.1303 1.8983 
ADASYN ＋BiLSTM 192/1251 0.1534 1.7076 
Paste checker 
ADASYN ＋BiGRU 81/567 0.1428 1.7918 
ADASYN ＋BiLSTM 70/567 0.1235 1.9601 
Turnout control 
circuit equipment 
ADASYN ＋BiGRU 167/1280 0.1305 1.8968 
ADASYN ＋BiLSTM 189/1280 0.1477 1.7531 
Permanent way 
equipment 
ADASYN ＋BiGRU 62/440 0.1409 1.8077 
ADASYN ＋BiLSTM 55/440 0.1250 1.9459 
Supporting 
equipment 
ADASYN ＋BiGRU 86/614 0.1401 1.8147 
ADASYN ＋BiLSTM 80/614 0.1303 1.8984 
Unknown reason 
ADASYN ＋BiGRU 21/124 0.1694 1.5902 
ADASYN ＋BiLSTM 14/124 0.1129 2.0614 
Table 2: Calculation results of class I classification weight of signal turnout equipment fault 
 

352     Informatica 46 (2022) 343-354                                                                                                                              H. Liang et al. 
 
(a): First-level fault classification 
 
(b): BiLSTM secondary classification training process 
Figure 6: Evaluation index values of deep learning integration model under different overall weight distribution  
Method Level 
Accuracy 
rate 
Recall 
rate 
F1 value 
Deep 
learning 
integration 
model 
Primary fault 
classification 
0.9106 0.9389 0.9245 
Secondary 
fault 
classification 
0.8564 0.8612 0.8588 
Table 3: Classification test results of deep learning 
integration model 
 
 
Figure 7: Graphical representation of classification test 
results of deep learning integration model 
 
It can be seen from Table 3 and Figure 7 that 
compared with ADASYN + BiGRU neural network, the 
comprehensive evaluation index of primary fault 
classification is improved by about 5%, and the 
comprehensive evaluation index of secondary fault 
classification is improved by about 9%. Compared with 
ADASYN + BiLSTM neural network, the 
comprehensive evaluation index of primary fault 
classification is improved by about 6%, and the 
comprehensive evaluation index of secondary fault 
classification is improved by about 10%. 
 
5 Conclusions 
This paper studies the fault diagnosis model of 
signal turnout fault text data, uses ADASYN data 
synthesis method to synthesize few category samples. 
This article also uses TF-IDF to extract features and 
transform vectors, and puts forward a deep learning 
integration method based on combination weight. The 
sample synthesis algorithm can appropriately 
synthesize a few categories of samples according to 
the distribution of the overall samples. There are 
several articles which uses SVM-SMOTE method to 
automatically synthesize the few category samples of 
signal equipment fault, and solve the problem of 
signal equipment fault sample imbalance. Through 
experimental analysis, it is proved that deep learning 
integration is a method that can effectively improve 
the classification performance of turnout fault 
diagnosis model. At the same time, this method can 
also provide a new idea for railway text classification 
and fault diagnosis. This article utilizes the deep 
learning algorithm of artificial intelligence for 
investigating the interlocking faults in the railway 
transportation. This paper uses ADASYN data 
synthesis method to synthesize few category samples, 
uses TF-IDF to extract features and transform vectors, 
and proposes a deep learning integration method based 
on combined weight. The outcomes obtained for the 
proposed method reveals that BiGRU has better 
overall classification performance when evaluated on 
the index of primary and secondary fault classification 
accuracy. 
 
80%
82%
84%
86%
88%
90%
92%
94%
96%
Accuracy rate Recall rate F1 value
Deep learning integration model Primary fault
classification
Deep learning integration model Secondary fault
classification

Application and Study of Artificial Intelligence in Railway…                                                Informatica 46 (2022) 343-354     353 
References  
[1] Kong, J. (2020, December). Application and 
research of artificial intelligence in digital library. 
In International conference on Big Data Analytics 
for Cyber-Physical-Systems (pp. 318-325). 
Springer, Singapore. 
https://doi.org/10.1007/978-981-33-4572-0_47 
[2] Paek, S., & Kim, N. (2021). Analysis of worldwide 
research trends on the impact of artificial 
intelligence in education. Sustainability, 13(14), 
7941. 
https://doi.org/10.3390/su13147941 
[3] Dobias, R., & Kubatova, H. (2004, August). FPGA 
based design of the railway's interlocking 
equipments. In Euromicro Symposium on Digital 
System Design, 2004. DSD 2004. (pp. 467-473). 
IEEE.  
10.1109/DSD.2004.1333312 
[4] Skiribou, C., Elbahhar, F., & Elassali, R. (2021). 
DMRS-based channel estimation for railway 
communications in tunnel environments. Vehicular 
Communications, 29, 100340.  
https://doi.org/10.1016/j.vehcom.2021.100340 
[5] Kiedrowski, P., & Saganowski, Ł. (2021). Method 
of Assessing the Efficiency of Electrical Power 
Circuit Separation with the Power Line 
Communication for Railway Signs 
Monitoring. Transport and 
Telecommunication, 22(4), 407-416.  
10.2478/ttj-2021-0031 
[6] Yang, J., Bai, X., Zhang, Z., Yang, M., Pan, P., Liu, 
T., & Tao, T. (2021, May). Research on the 
application of BDS/GIS/RS technology in the high 
speed railway infrastructure maintenance. In IOP 
Conference Series: Earth and Environmental 
Science (V ol. 783, No. 1, p. 012168). IOP 
Publishing.  
10.1088/1755-1315/783/1/012168 
[7] Lin, J., Hu, X., Dang, J., & Wu, Z. (2019). Traffic 
model of machine-type communication for railway 
signal equipment based on MMPP. IET 
Microwaves, Antennas & Propagation, 13(8), 1072-
1079. 
https://doi.org/10.1049/iet-map.2018.6004 
[8] Wang, X., Guo, J., Jiang, L., Fu, J., & Li, B. (2016, 
August). Intelligent fault diagnosis and prediction 
technologies for condition based maintenance of 
track circuit. In 2016 IEEE International 
Conference on Intelligent Rail Transportation 
(ICIRT) (pp. 276-283). IEEE.  
10.1109/ICIRT.2016.7588745 
[9] Sahal, R., Breslin, J. G., & Ali, M. I. (2020). Big 
data and stream processing platforms for Industry 
4.0 requirements mapping for a predictive 
maintenance use case. Journal of manufacturing 
systems, 54, 138-151.  
https://doi.org/10.1016/j.jmsy.2019.11.004 
[10] Cao, Y ., Li, P., & Zhang, Y . (2018). Parallel 
processing algorithm for railway signal fault 
diagnosis data based on cloud computing. Future 
Generation Computer Systems, 88, 279-283.  
https://doi.org/10.1016/j.future.2018.05.038 
[11] Dong, C. Z., Ye, X. W., & Jin, T. (2018). 
Identification of structural dynamic characteristics 
based on machine vision 
technology. Measurement, 126, 405-416.  
https://doi.org/10.1016/j.measurement.2017.09.043 
[12] Jia, Z., & Sharma, A. (2021). Review on engine 
vibration fault analysis based on data 
mining. Journal of Vibroengineering, 23(6), 1433-
1445.  
https://doi.org/10.21595/jve.2021.21928 
[13] Yin, M., Li, K., & Cheng, X. (2020). A review on 
artificial intelligence in high-speed 
rail. Transportation Safety and Environment, 2(4), 
247-259.  
https://doi.org/10.1093/tse/tdaa022 
[14] Ren, X., Li, C., Ma, X., Chen, F., Wang, H., 
Sharma, A., & Masud, M. (2021). Design of multi-
information fusion based intelligent electrical fire 
detection system for green 
buildings. Sustainability, 13(6), 3405. 
https://doi.org/10.3390/su13063405 
[15] Sharma, D., Kaur, R., Sandhir, M., & Sharma, H. 
(2021). Finite element method for stress and strain 
analysis of FGM hollow cylinder under effect of 
temperature profiles and inhomogeneity 
parameter. Nonlinear Engineering, 10(1), 477-487.  
https://doi.org/10.1515/nleng-2021-0039 
[16] Afandizadeh, S., & Rad, H. B. (2021). Developing a 
model to determine the number of vehicles lane 
changing on freeways by Brownian motion 
method. Nonlinear Engineering, 10(1), 450-460. 
https://doi.org/10.1515/nleng-2021-0036 
[17] Shabaz, M., Sharma, A., Al Ajrawi, S., & Estrela, V . 
V . (2022). Multimedia-based emerging technologies 
and data analytics for Neuroscience as a Service 
(NaaS). Neuroscience Informatics, 2(3), 100067.  
https://doi.org/10.1016/j.neuri.2022.100067 
[18] Meher, M., & Rostamy, D. (2021). Hybrid of 
differential quadrature and sub-gradients methods 
for solving the system of Eikonal 
equations. Nonlinear Engineering, 10(1), 436-449.  
https://doi.org/10.1515/nleng-2021-0035 
[19] Mi, Z., Wang, T., Sun, Z., & Kumar, R. (2021). 
Vibration signal diagnosis and analysis of rotating 
machine by utilizing cloud computing. Nonlinear 
Engineering, 10(1), 404-413.  
https://doi.org/10.1515/nleng-2021-0032 
[20] Wang, H., Sharma, A., & Shabaz, M. (2022). 
Research on digital media animation control 
technology based on recurrent neural network using 
speech technology. International Journal of System 
Assurance Engineering and Management, 13(1), 
564-575. 
https://doi.org/10.1007/s13198-021-01540-x 
[21] Yousaf, B., Qaisrani, M. A., Khan, M. I., Sahar, M. 
S. U., & Tahir, W. (2021). Numerical and 

354     Informatica 46 (2022) 343-354                                                                                                                              H. Liang et al. 
experimental analysis of the cavitation and study of 
flow characteristics in ball valve. Nonlinear 
Engineering, 10(1), 535-545.  
https://doi.org/10.1515/nleng-2021-0044 
[22] Singh, P. K., & Sharma, A. (2022). An intelligent 
WSN-UA V-based IoT framework for precision 
agriculture application. Computers and Electrical 
Engineering, 100, 107912. 
https://doi.org/10.1016/j.compeleceng.2022.107912 
[23] Zeng, H., Dhiman, G., Sharma, A., Sharma, A., & 
Tselykh, A. (2021). An IoT and Blockchain‐based 
approach for the smart water management system in 
agriculture. Expert Systems, e12892. 
https://doi.org/10.1111/exsy.12892 
[24] Sharma, A., & Singh, P. K. (2021). UA V‐based 
framework for effective data analysis of forest fire 
detection using 5G networks: An effective approach 
towards smart cities solutions. International 
Journal of Communication Systems, e4826. 
https://doi.org/10.1002/dac.4826 
[25] Sharma, A., Singh, P. K., & Kumar, Y . (2020). An 
integrated fire detection system using IoT and 
image processing technique for smart 
cities. Sustainable Cities and Society, 61, 102332. 
https://doi.org/10.1016/j.scs.2020.102332 
[26] Zang, Y ., Shangguan, W., Cai, B., Wang, H., & 
Pecht, M. G. (2019). Methods for fault diagnosis of 
high-speed railways: A review. Proceedings of the 
institution of mechanical engineers, part O: journal 
of risk and reliability, 233(5), 908-922.  
https://doi.org/10.1177/1748006X18823932 
[27] Ting, L., Khan, M., Sharma, A., & Ansari, M. D. 
(2022). A secure framework for IoT-based smart 
climate agriculture system: Toward blockchain and 
edge computing. Journal of Intelligent 
Systems, 31(1), 221-236.  
https://doi.org/10.1515/jisys-2022-0012  
[28] Minea, M., Dumitrescu, C. M., & Dima, M. (2021). 
Robotic Railway Multi-Sensing and Profiling Unit 
Based on Artificial Intelligence and Data 
Fusion. Sensors, 21(20), 6876.  
https://doi.org/10.3390/s21206876 
[29] Luo, J., Wu, M., Gopukumar, D., & Zhao, Y . 
(2016). Big data application in biomedical research 
and health care: a literature review. Biomedical 
informatics insights, 8, BII-S31559.  
https://doi.org/10.4137/BII.S31559  
[30] Chen, F., Deng, P., Wan, J., Zhang, D., Vasilakos, A. 
V ., & Rong, X. (2015). Data mining for the internet 
of things: literature review and 
challenges. International Journal of Distributed 
Sensor Networks, 11(8), 431047. 
https://doi.org/10.1155/2015/431047