Air Traffic Safety Risk Assessment

based on Rough Set and BP Neural Network

Lan Ma

, Weian Li and Zengxian Geng

Air Traffic Management College,

Civil Aviation University of China, Tianjin, China

Keywords: Air Traffic Safety, Rough Set, Attribute Reduction, BP Neural Network.

Abstract: The safety of air traffic control is an important link in the safety system of civil aviation industry. In order to

evaluate the safety risk of air traffic control in a more comprehensive and reliable way, proposing an air

traffic safety risk modeling and evaluation method based on rough set and BP neural network. After

analyzing the factors that may affect the safety in the actual work of ATC, 24 attribute variables which can

measure the safety risk of ATC are given. Aiming at the shortcomings of traditional neural network training

with high redundancy, slow convergence and easy to fall into local optimum, the attribute reduction method

is used to reduce the input attribute by rough set theory. Under the premise of not affecting the training

results and the accuracy of the data, removing the low correlation attributes with the results, the network

structure is simplified, the training times are reduced, and the training speed and accuracy of the neural

network are improved. Use the simplified condition attributes of the original data after rough attribute

reduction as input data, the conflict resolution object is as output data, using MATLAB to build the neural

network, and the trained network is tested and verified to be reliable. Compared with the model before the

reduction of the initial data, significantly improves the accuracy and efficiency. The model is verified by

examples The results show that the combination of rough set and BP neural network can accurately evaluate

the risk of air traffic control, change the risk assessment from qualitative to quantitative, and provide

guidance for the actual operation.

1 INTRODUCTION

The unsafe incidents in civil aviation operations can

be divided into five categories: aircraft operation,

aircraft maintenance, ground support, airport

operations and ATC safeguards. It can be seen that

some of the unsafe incidents are related to ATM

system. Air Traffic Management System is an

important part of the civil aviation system. It is also

a complex system of structural correlation. In recent

years, air traffic control accidents such as runway

incidents caused by security problems are common.

In order to ensure the safe operation of civil aviation

system, it is crucial to assess the air traffic safety

risk (Zellweger and Donohue 2015).

Ma Lan, female, Civil Aviation University of China,

Associate Professor, Ph.D. major research area:

Transportation planning, air traffic control information

processing.

1.1 Research Status

As people pay more attention to the safety of ATC,

more and more scholars inland and abroad are

engaged in the research of ATC safety, and they

have achieved some results. In the 1990s, academics

in the United States and European countries started

to study the theory of civil aviation safety risk

management. Shyur (2008) quantified the aviation

risk caused by human error with studying aviation

accidents and safety indicators. The benchmark risk

function was taken into aviation risk evaluation as

the quadratic function, then get a proportional risk

model to investigate non-linear aviation safety

factors and evaluate aviation risk; Al Basman and

Hu (2012) have studied the theory of stochastic

safety analysis that can be used in ATC systems and

proposed two ways that a multi-level Markov chain

and air traffic flow assessment to solve the safety

issues in the ATC environment;

Cruck and Lygeros

(2015)

have built a hybrid model which can be man-

Ma, L., Li, W. and Geng, Z.

Air Trafﬁc Safety Risk Assessment based on Rough Set and BP Neural Network.

DOI: 10.5220/0006889008630870

In Proceedings of the 13th International Conference on Software Technologies (ICSOFT 2018), pages 863-870

ISBN: 978-989-758-320-9

863

machine interaction, this model can be used in

complex situations to forecast, in order to percept air

traffic safety risk.

Domestic research on ATC safety risks started a

little later than abroad. However, in recent years

there have been some achievements. Many scholars

have used Bayesian analysis (Liao et al. 2015), gray-

level analysis(Guo et al. 2015), matter-element

analysis theory (Zhang et al. 2016) and other

methods to study the ATM safety problems. Based

on the analysis of triangular fuzzy mathematics and

ANP principle, Du et al. (2010) have established an

air traffic control security risk assessment model

with Fuzzy-ANP in view of the interaction of

security risk factors; China's civil aviation industry

now put forward the safety management system

(SMS) whose core is risk management to control the

security risk through the overall operation of the

various links (Lu 2017).

1.2 Introduction

ATC work experience and expert

advice

Air traffic control risk

assessment index system

ATC Risk Assessment Data

Rough set attribute reduction

Build neural network

Use test data to evaluate ATC

safety level

Figure 1: Research content and process.

This paper combines and draws on the

advantages of the existing researches, and on this

basis expands and builds the risk assessment index

system. Then it extends the qualitative research to

quantitative research that can be more universal, and

using rough set theory with attribute reduction

function makes the neural network model structure

(Wang 2013) more concise, and makes the

forecasting results fast and accurate. The research

content and process is shown in figure 1.

2 CONSTRUCTION OF RISK

ASSESSMENT INDEX SYSTEM

Based on a comprehensive analysis of the

characteristics and significance of air traffic control

system and its important position in the civil

aviation industry, air traffic control risk assessment

indicators (Luo et al. 2009) are generally divided

into four categories: human factors, equipment

factors, environmental factors and management

factors The factors are the first layer of the whole

index system, and then combine the examples of the

ATC operation and the opinions given by the ATC

experts. Each factor contains several sub-factors and

they are evaluated as an indicator in the subsequent

research. The final index system has the

characteristics of science, and fits the actual work of

the ATC. Air traffic control security risk assessment

indicator system in Figure 4.

In order to analyze the impact of each indicator

on the safety of ATC by means of qualitative and

quantitative analysis, investigate the senior

management of ATC system, controllers and ATC

experts in the form of questionnaires. And ask them

to combine their own actual work or research

conditions, or assessment of an unsafe event, and

mark each indicator based on their own knowledge

and experience, scoring criteria: 1 point - very good,

2 points - good, 3 points - ordinary, 4 points - poor,

5 points - very poor. After marking the index, given

the general

Evaluation of the security risk rating, there are

five levels: level 1 is lowest risk, level 2 is lower

risk, level 3 is medium risk, level 4 is higher risk,

and level 5 is highest risk. Summarizing above

work, the safety risk assessment system for ATC and

the results of the questionnaire form the data

foundation for the follow-up study in this paper.

3 INDEX REDUCTION BASED

ON ROUGH SET THEORY

Rough set theory (Wang et al. 2009) is used to

analyze and process data. It was first proposed by

Polish mathematician Z. Pawlak in 1982, which has

many advantages. It does not need too much raw

data to find the hidden rules of data, it is now widely

used in data mining, pattern recognition and other

fields.

Finally, complete content and organizational

editing before formatting. Please take note of the

SE-CLOUD 2018 - Special Session on Software Engineering for Service and Cloud Computing

864

following items when proofreading spelling and

grammar:

3.1 Rough Set Theory Applied to Air

Traffic Control

Define abbreviations and acronyms the first time

they are used in the text, even after they have been

defined in the abstract. In general, it is best to avoid

acronyms in the abstract unless they are critical.

Abbreviations such as IEEE, SI, MKS, CGS, SC,

DC, and RMS do not have to be defined. Do not use

abbreviations in the title or heads unless they are

unavoidable.

Obviously, the collated original data is too

complicated to subsequent computational studies.

Then we find that all the attributes are not equally

important. There are some redundant or inconclusive

attributes that can be eliminated, further obtaining a

more streamlined and intuitive decision table.

Given R is a set of reduced attributes to be

obtained, P is a set of condition attributes

corresponding to 24 ATM safety risk assessment

indicators, X is an instance set with removal

compatibility examples, EXPECT is a termination

condition for attribute dependencies, The following

algorithm:

Initialize, make

()

RcoreC=

()

CcoreC=−

1k =

• Remove all compatible instances of U, that is:

()

UPOSD=− (1)

• Calculation

()

card POS D

card U

(2)

• if

k EXPECT≥ , the algorithm terminates;

otherwise, if

() ()

POS D POS D= , returns

()

card POS D

card U

(3)

the algorithm terminates;

• For any

P∈

, calculation of

{}

()

card POS D

∪

= (4)

{}

()

{}

()

max_ /

m sizePOS D RPD

∪

=∪∪ (5)

• For all

P∈ , calculation

× with the

maximum value, and

given

{

}

RR P=∪ ,

{

}

PP p=− ;

• Return to the first step

3.2 Data Attribute Reduction

In the study, 310 scoring surveys were conducted on

the 24 safety risk assessment indicators of air traffic

control. The final survey data constituted the domain

U of the knowledge representation system, and then

formed a complete decision table. The safety risk

indicators {X1, X2, X3 , ..., X24} is the condition

attribute, the security risk level is the decision

attribute, the value of the attribute is 1, 2, 3, 4, 5,

which meets the discretization requirements of the

attribute reduction of the rough set. Show in table 1.

Table 1: Air traffic safety risk assessment decision table.

attribute C attribute D

X1 X2 X3 … X24 safety risk level

1 1 3 1 … 2 1

2 3 1 2 … 1 3

3 4 1 2 … 2 2

4 3 1 1 … 1 3

… … … … … … …

310 2 1 2 … 2 1

The reducing result is usually not the only one.

Finding all reductions or minimal reductions have

proved to be an NP-hard problem. The general

solution to this problem is heuristic search, which

builds a reduced set of attributes by computing

dependencies of attributes. And there are common

attribute reduction methods like Johnson greedy

algorithms and genetic algorithms.

Air Trafﬁc Safety Risk Assessment based on Rough Set and BP Neural Network

865

For processing the original data, using Rosetta

software to conduct attribute reduction, the software

is table logic data tool based on rough set theory

framework, it can be used to simplify the model. The

processed data excel tables imported into Rosetta

software for data complement and discretization.

And select the method to reduce the data, then you

can directly get attribute reduction results.

Table 2: Reduced air traffic safety risk assessment.

Factors number indicators

human

factors

Controller technical level is not up to

standard

Controller psychological quality is not

up to standard

Emergency response capacity is not

enough

Controller or department head safety

awareness and sense of responsibility

is not enough

equipment

factors

Communication equipment is not

working properly

Navigation device is not working

properly

Monitoring equipment is not working

properly

environ-

ment

factors

X13 Seat monitoring means imperfect

X14

Scene monitoring and guidance

system is not working properly

X16

Weather conditions and weather

disasters

manage-

ment

factors

X19 Poor management practices

X20

Department of high frequency of

conflict intensity

X23

Safety education and training system

is not perfect

After processing the data, the attributes after

reduction are shown in Table 2, it can be seen that

the attribute reduction obviously reduces the number

of attributes from the original 24 to 13. To a certain

extent, the reduction result also reflects the impact of

various factors on the safety of ATC, showing that

the most obvious impact of ATC safety is human

factor.

4 AIR TRAFFIC CONTROL

SECURITY RISK ASSESSMENT

BY NEURAL NETWORK

METHOD

After the text edit has been completed, the paper is

ready for the template. Duplicate the template file by

using the Save As command, and use the naming

convention prescribed by your conference for the

name of your paper. In this newly created file,

highlight all of the contents and import your

prepared text file. You are now ready to style your

paper; use the scroll down window in the formatting

toolbar (

Silva.et al. 2017).

4.1 Construction and Application of BP

Neural Network based on Rough

Set

Analyzing the reduction result of the safety risk

indicators assessment data of ATC. It can be

concluded that there is a certain causal relationship

between the attributes and the final safety

assessment, and the inherent law of the relation can

be tapped by neural network (Liu et al. 2010) to get

the quantitative and qualitative assessment of ATC

risk assessment.

Aiming at the shortcomings of traditional neural

network training with high redundancy, slow

convergence and easy to fall into local optimum, the

attribute reduction method is used to reduce the

input attribute by rough set theory. Under the

premise of not affecting the training results and the

accuracy of the data, removing the low correlation

attributes with the results, the network structure is

simplified, the training times are reduced, and the

training speed and accuracy of the neural network

are improved.

For the general pattern recognition problem,

adopting a three-layer neural network structure

(Bruin et al. 2017)

. The following is the step of

modeling the neural network of air traffic safety risk

assessment(Ma and Chang 2017):

• Initialize weights, assign random values in the

interval (0,1) to each connection weight and

threshold;

• Determine the number of neurons in each layer,

the number of neurons in the input layer is 13

attributes, the attribute values input after

discretization;

SE-CLOUD 2018 - Special Session on Software Engineering for Service and Cloud Computing

866

• Determine the number of hidden layer neurons.

In the three-layer network, choosing the number

of hidden layer neurons is a very complex issue.

It often requires designers' experience and

multiple tests to determine. If the number of

hidden layer neurons is n2, the number of input

layer neurons is n1, the number of output layer

neurons is m, choose the best n2 can refer to the

following formula:

nnma=++

Where a is the constant between

[

]

1,10

;

221

lognn=

However, the quantity is not fixed and needs to

be constantly adjusted by the actual training test. In

Table 3, the training errors when selecting different

hidden layers are listed. Therefore, the number of

hidden layer neurons is set as 9 in this study. Ensure

the training accuracy and improve the computing

speed.

Table 3: Implicit layer test error.

hidden layer

neurons

number

5 6 7 8 9

training

errors

0.00350 0.00928 0.00300 0.00916 0.00122

• Choose transfer function. The transfer function

of hidden layer neurons adopts S-type tangent

function:

()

−

=−

Output layer neurons transfer function using S-

type logarithmic function:

()

−

• The output layer should reflect the final five

security risk assessment levels. In order to

ensure the accuracy of the training results, we

classify the levels as level 1 (1 0 0 0 0), level 2

(0 1 0 0 0), level 3 (0 0 1 0 0), level 4 (0 0 0 1 0),

level 5 (0 0 0 0 1). Five neurons are set in the

output layer to quantify the above five levels.

After modeling, the 310 groups of data are

divided into training group and test group. The data

are normalized by MATLAB (Ge and Sun 2007) and

creating the complete network object by using the

newff function in the toolbox. The training function

trainlm uses Levenberg-Marquardt algorithm to train

the network, after which you can set the training

parameters in Table 4. Seen from Figure 2, after

eight training, the network performance to meet the

requirements, which is related to the network

structure and learning rate.

Table 4: Training parameters.

training times training goal learning rate

1000 0.001 0.1

Figure 2: Training results.

4.2 Combination of Rough Sets and

Neural Network Methods

Two variables were set for experiment, namely

whether using the improved BP neural network and

whether using the rough set theory was for data

preprocessing. Then four experimental methods

were formed and the four experimental methods

were compared. The results are shown in Table5.

From the research results, we can know that

using the rough set to reduce the attributes of the

initial data can remove the redundant attributes and

reduce the number of training sample attributes.

Then the neural network structure is more concise

and the operation time is reduced. The improved BP

Air Trafﬁc Safety Risk Assessment based on Rough Set and BP Neural Network

867

neural network is obviously superior to the BP

neural network before in terms of time and accuracy.

Combining rough set attribute reduction with neural

network modeling can simplify the model structure

and improve the operation efficiency and accuracy.

It is a scientific and effective method to deal with

such problems.

Table 5: Comparison of experimental methods.

Reduction Improved

Training

time

Number of

iterations

Training

error

No Yes 80 1670 0.068

No No 87 2000 0.14

Yes Yes 24 668 0.018

Yes No 41 1000 0.049

4.3 BP Neural Network Test

The neural network training of the data can be seen

from Figure 3. After 668 trainings, the network

performance has reached the requirements. It is

related to the network structure and the learning rate.

Figure 3: Neural network training performance.

Next, we test the trained network and select 5

sets of data as the test input data. The test code is

()

,YsimnetPtest= . The test data is shown in Table

6. According to the European norm theory, the error

of the test result is very small. It can be determined

that the network meets the requirements of air traffic

safety assessment after training.

5 CONCLUSION

Based on the indicator system of security risk

assessment of air traffic control, use the rough set

theory to reduce the influencing factors. It can

reduce the input attributes from 24 to 13, remove the

redundant attributes and simplify the complexity of

the network structure. Improve network training

rate, and get training results more quickly and

accurately. Design a complete BP neural network,

build a learning model, and excavate the potential

relationship within the data. It can be seen from the

test samples that the constructed network can meet

the requirements of air traffic control security risk

assessment, and provide the forecast risk assessment

level objectively and accurately. It is in favour of the

practical work of the ATC.

REFERENCES

Al Basman, M.& Hu, J. 2012, Approaches for stochastic

safety analysis arising in ATM application. Decision

and Control. IEEE, 2012:514-519.

Bruin, T D, Verbert, K.& Babuška, R. 2017， Railway

Track Circuit Fault Diagnosis Using Recurrent Neural

Networks. IEEE Transactions on Neural Networks &

Learning Systems, 2017, 28(3):523.

Cruck, E.& Lygeros, J. 2015, Hybrid modeling for the

evaluation of risk perception by air-traffic controllers.

Control Conference. IEEE, 2015:1278-1283.

Du H， Li H.& Yuan L. 2010, Risk assessment model for

air traffic control based on fuzzy-ANP method. China

Safety Science Journal, 2010, 20(12):79-85. (in

Chinese)

Ge Z.& Sun Z. 2007, Neural Network Theory and

Realization of MATLAB R2007 . Publishing House of

Electronics Industry, 2007

Guo J, Li Y.& Yang C. 2016, Classfication for general

aviation air traffic management operation security

based on gray analytic hierarchy process method.

Safety and Environmental Engineering, 2016,

23(2):162-166. (in Chinese)

Liao Y, Guo J.& Yang C. 2015, Probability assessment of

ATM safety based on bayes analysis. China Science

Paper, 2015(19):2271-2275. (in Chinese)

SE-CLOUD 2018 - Special Session on Software Engineering for Service and Cloud Computing

868

Air traffic safety risk assessment

index system

human factors

equipment factors

environment

factors

management

factors

Controller technical level is not up to standard

Controller psycho logical quality is not up to stand ard

Emergen cy response capacity is not enough

Controller misuses key instructions

Controller or department head safety awareness and sens e of responsibility is not

enough

cooperation of stations or departments is poor

Communication equipment is not working properly

Navigation device is n ot working properly

Monitoring equipment is not working p roperly

ATC system backup and emergency response capacity is not enough

Technical standards oversight

The number and layout of the con trol seats are not reasonable

Seat monitoring means imperfect

Scene monitoring and guidance system is not working properly

Military activities affected

Weath er cond itions and weather d isasters

Impact of airspace restrictions

Management standards oversight

Poo r management practices

Department conflict has hig h frequency and intens ity

Personnel do not agree with organizational g oals

Organizational structure is not reaso nable

Safety education and training sys tem is no t p erfect

Control procedures are not reasonable

Figure 4: Air traffic safety risk assessment index system.

Table 6: Test data.

U X1 X2 X3 ··· X20 X23 output value

expected

value

assessment

level

U1 1 1 3 ··· 3 1 0.11621 0.01528 0.82108 0.08179 0.00001 0 0 1 0 0 3

U2 3 1 3 ··· 1 1 0.01702 0.89169 0.11587 0.00964 0.00051 0 1 0 0 0 2

U3 3 3 1 ··· 3 1 0.04513 0.00079 0.99243 0.00050 0.00242 0 0 1 0 0 3

U4 1 2 2 ··· 2 3 0.17850 0.80827 0.00123 0.01184 0.00706 0 1 0 0 0 2

U5 2 1 2 ··· 2 1 0.88620 0.00984 0.16700 0.00888 0.00096 1 0 0 0 0 1

Air Trafﬁc Safety Risk Assessment based on Rough Set and BP Neural Network

869

Liu W, Wang P.& Sun D. 2010， Forecasting model of

road traffic accident based on improved BP neural

network. Computer Systems and Applications, 2010,

19(10):177-181.

Lu W. 2017, Talking about the concept of safety

management system and popularization in civil

aviation. China High Tech Enterprises, 2017(8):234-

235. (in Chinese)

Luo F, Jia G.& Chen G. 2009, Design and optimization of

index system for the ATC safety risk management.

China Safety Science Journal, 2009, 19(8):115-120.

(in Chinese)

Ma L.& Chang Y. 2017, Highway construction investment

risk evaluation using BP neural network model.

International Conference on Industrial Technology

and Management. IEEE, 2017:154-158.

Shyur, H J. 2008, A quantitative model for aviation safety

risk assessment. Computers & Industrial Engineering,

2008, 54(1):34-44.

Silva, T V D, Monteiro, R V A.& Moura, F A M. et al.

2017, Performance Analysis of Neural Network

Training Algorithms and Support Vector Machine for

Power Generation Forecast of Photovoltaic Panel.

IEEE Latin America Transactions, 2017, 15(6):1091-

1100.

Wang X. 2013, Several rough models and the combination

of rough set and neural network research. University of

Electronic Science and Technology of China, 2013. (in

Chinese)

Wang G, Hu J.& Zhang Q, et al. 2011, Granular

computing based data mining in the views of rough set

and fuzzy set. IEEE International Conference on

Granular Computing. IEEE, 2011:67-67.

Zhang X, Han S.& Yang C. 2016, An fuzzy synthetic

evaluation model for safety risk in air traffic control

based on matter element analysis . Journal of

transport information and Safety. 2016, 34(4):50-56.

(in Chinese)

Zellweger, A G.& Donohue, G L. 2015, Accident Risk

Assessment for Advanced Air Traffic Management.

Minerva Chirurgica, 2015, 14(4):214-5.

SE-CLOUD 2018 - Special Session on Software Engineering for Service and Cloud Computing

870